Antibody Discovery uses artificial intelligence for multi-dimensional optimization

In many areas, traditional techniques are being turbocharged by fast, rational, computer-driven techniques—that is, artificial intelligence (AI). And there is ample evidence that monoclonal antibody (mAb) discovery is one of those areas. It has long benefited from applications of wet chemistry and analytics. And now it is starting to benefit from AI as well. For example, AI technologies are already searching for new antibody sequences that can bind to specific targets. In addition, AI technologies promise to solve many development-stage problems, such as manufacturability.

Without AI, antibody discovery can be tedious. “The problem with the way antibodies are discovered—and the way they are optimized for high affinity, high stability, low viscosity, long half-life in vivo—is that there is too much trial and error,” says Dr. Peter M. Tessier, a professor of chemical biology at the University of Michigan.

AI reduces the need for experimentation by identifying high-quality monoclonal antibody candidates much faster, at the discovery stage. With its potential to optimize multiple properties simultaneously, AI can help narrow the number of mutations required to improve antibody performance and make it ready for clinical testing.

Tessier says that the current benefits of AI-based monoclonal antibody discovery have less to do with de novo design of new antibodies and more to do with “rapid, efficient identification of antibodies with impressive combinations of properties in a much shorter time frame.”

Finding new chemistry

Many experts believe that AI-driven monoclonal antibody design will enable developers to leverage the structures and properties of different antibody sequences, proposing new design space for chemistry and perhaps addressing unmet medical needs. One such expert is Harshitha Gamahewage, Business Operations Manager at BigHat Biosciences.

“Traditional antibody engineering is done manually—that is, designed by a scientist—or through biased or random sequence mutagenesis—for example, through an affinity maturation library,” Gamahewage says. Developers can only hope to improve or fix a relatively small number of antibody properties or deficiencies, such as affinity, purity, or viscosity.

In this zero-sum game, gains in one property are usually achieved at the expense of others, and improvements are limited to sequence space within a few mutations of the starting sequence. Furthermore, standard approaches have difficulty in conceiving new chemical spaces or formats other than standard mAbs.

“AI approaches can work with any number of existing sequences, model multiple properties simultaneously, generalize to a more diverse sequence space, and generate or propose new sequences that meet certain design conditions,” Gamahewage continues. “These algorithms can take in more data and model more simultaneous complexity than a typical scientist can hold in their brain at one time.”

By learning underlying sequences and/or structural determinants of antibody properties from thousands or even millions of sequences, AI models can more efficiently propose potentially more diverse solutions to difficult engineering problems or, in the case of a new discovery, design an antibody from scratch.

What is needed

“Machine learning protein design is exciting and evolving, with new capabilities emerging every month,” notes BigHat’s Gamahewage, “but ultimately, the proof of concept of these techniques will be their ability to bring antibodies into the clinic to treat patients. As we develop increasingly sophisticated approaches, the most exciting applications are yet to come.”

This may seem extraordinary to old-school chemists and biologists, but the only requirements for designing antibody sequences and structures today are a laptop, open-source software, and software libraries. The advent of AI provides an additional dimension, primarily the ability to invent and provide instructions for previously unknown antibody structures and functions, that is, to explore uncharted chemical space.

Although AI-assisted antibody discovery involves a relatively small investment in hardware and software, an exceptionally high level of skill is required to program and maintain the software, design experiments, and decide which molecules to promote.

“At a minimum, you’d need skills in Python or a deep learning framework like PyTorch, as well as Biopython and ANARCI skills to deal with antibody sequences,” notes Gamahewage. “In addition, to overcome the data shortage in de novo antibody design, AI implementers need to leverage general antibody or structure predictors like AlphaFold and IgFold, protein language models like ESM or AntiBERTy, a protein design diffusion model like RFdiffusion, or even rational design tools like Rosetta and the structure browser PyMol. All of this can be conveniently run on a single GPU on a decent laptop.”

However, because models cannot accurately predict how AI-designed antibodies will perform in real life, a research lab is needed to validate any computational designs and test for stability, propensity to aggregate, target affinity, purity, and other properties.

Combining competences

So rather than eliminating or reducing the reliance on wet chemical assays and their associated analysis, AI-based monoclonal antibodies are increasing the demands on these tools in terms of speed, robustness, and reliability. Earlier this year, ImmunoPrecise Antibodies (IpA), a Canadian AI-biopharm research and technology company, not only acquired Carterra’s surface plasmon resonance (SPR) instrument platform, but also formally announced the acquisition as part of its commitment to AI. SPR, a rapid biophysical assay for quantifying molecular affinity, provides affinity data in seconds and is very economical in test materials.

At the time, IpA CEO Dr. Jennifer Bath noted that SPR would complement her company’s “capabilities in high-throughput antibody discovery, manufacturing, and screening” and allow the company to “come closer to creating the fastest, most cost-effective drug discovery pipeline.”

Not only are there AI-focused antibody discovery companies acquiring analytical and laboratory skills, but there are also traditional antibody discovery organizations acquiring AI skills. GENE More formal collaboration is expected to develop between AI companies and enterprises with established expertise in discovery and development.

For example, in April 2024, BigHat Biosciences partnered with Johnson & Johnson subsidiary Janssen Biotech to combine the larger company’s drug discovery, clinical development and data science expertise with BigHat’s Milliner platform, a suite of machine learning technologies integrated with a rapid wet lab to aid in the design and selection of high-quality antibodies for multiple therapeutic targets in neuroscience.

Milliner integrates synthetic biology-based rapid wet lab with machine learning technologies into a complete antibody discovery and engineering platform, with the goal of engineering antibodies with more complex functions and improved biophysical properties. This approach, BigHat claims, reduces the difficulty of designing therapeutic proteins to treat a range of chronic and life-threatening diseases while significantly accelerating the discovery and validation of candidates.

While there is work underway to strengthen collaboration between traditional biopharmaceutical companies and AI, joint ventures will not be limited to traditional biopharmaceutical and AI companies, as various services companies will also seek to join forces to provide AI-powered design and discovery services.

For example, in March 2024, BioGeometry, a digital biology company specializing in AI-driven protein design, and Sanyou Biopharmaceuticals, a biology R&D services company, joined forces to leverage their expertise to create a next-generation antibody drug discovery platform. Under the agreement, Sanyou will integrate its wet chemistry expertise into GeoBiologics, BioGeometry’s generative AI antibody design platform.

In late 2023, Absci, an AI antibody discovery company, partnered with AstraZeneca to use AI to find new cancer treatments. The agreement leverages Absci’s AI-based protein design platform with AstraZeneca’s oncology expertise to accelerate discovery and development.

Absci’s Integrated Drug Creation platform combines generative AI and scalable wet lab technologies based on millions of protein-protein interactions. AI is used at the front end (discovery) with more or less conventional analytics and chemical testing and has the potential to drive discovered proteins to the clinic faster. A six-week workflow is anticipated.

The announcement follows Absci’s publication on de novo antibody design and validation using the company’s “zero-shot” generative AI model. With this model, developers can design antibodies that bind to specific targets without the need for training data from antibodies that are known to bind to those targets. In other words, the target is the only requirement. This allows the generation of antibodies that are structurally and chemically distinct from those in existing antibody databases, including versions of all three heavy-chain complementarity-determining regions that are most critical for target binding. Absci says its discovery platform finds up to 30 times more hits than those found using standard antibody library methods.

Improvement of all features

In antibody discovery, AI has focused on identifying molecules with high target affinity and new chemical space. But AI also has the potential to improve almost every attribute of an antibody. One such attribute is developability—the likelihood that a molecule will proceed smoothly through the chemistry, manufacturing, and control (CMC) process at a reasonable cost and within a reasonable time frame.

In a paper published in early 2023, researchers from Shanghai-based WuXi Biologicals proposed AI-based screening to “reduce the risk of an antibody candidate with poor development ability progressing to the CMC stage” and that this should be done “as early as possible… in a rapid and efficient manner, using small amounts of test materials.” The researchers argued that this approach could be applied not only to monoclonal antibodies but also to advanced therapies such as bispecific antibodies, multispecific antibodies, antibody-drug conjugates, and other monoclonal antibody derivatives.