A breakthrough in protein modelling has the potential to revolutionise healthcare delivery and drug development by enabling personalised treatment for complex diseases. At the heart of this advancement lies AbMap, a computational model that uses artificial intelligence (AI) to predict antibody structures and binding strengths with remarkable precision. By addressing the challenge of modelling hypervariable antibody regions, AbMap offers a new pathway for accelerating drug discovery and unravelling immune system mysteries.

Antibodies, crucial to the immune system, consist of long chains of amino acids that fold into complex structures. These structures, particularly the hypervariable regions, are essential for binding antigens and mounting an immune response. However, the extreme variability of these regions makes them difficult to predict using traditional AI models.

The hypervariable regions, located at the tips of the Y-shaped antibody, are unconstrained by evolutionary patterns. This lack of constraints creates a unique challenge for AI models, which typically rely on evolutionary signals to predict protein structures. As Singh, a researcher in the study, explains, “Part of the reason why language models can predict protein structure well is that evolution constrains these sequences in ways the model can decipher.” Without these constraints, predicting hypervariable regions becomes akin to decoding a language without grammar.

To overcome the challenge of predicting hypervariable regions in antibodies, researchers developed AbMap, an advanced model built on existing protein language models but enhanced with two specialised modules. The first module learnt from about 3,000 antibody structures in the Protein Data Bank. This allowed it to find patterns in highly variable sequences and guess what the structures would probably be. Complementing this, the second module analysed 3,700 antibody sequences to establish correlations between sequence variations and their binding affinities for antigens. Together, these modules allow AbMap to accurately predict antibody structure and binding strength. This makes it possible to quickly test millions of different antibody variants and pick the most promising ones for further testing in the lab.

Based on the amino acid sequences of antibodies, this combination lets AbMap guess both their structure and how well they bind. By leveraging this dual capability, researchers can now evaluate millions of antibody variants rapidly and identify the most promising candidates for experimental validation.

The application of AbMap in neutralising the spike protein of SARS-CoV-2, the virus responsible for COVID-19, demonstrated its potential. Starting with a small set of predicted antibodies, the model generated millions of variants by altering hypervariable regions. They then identified the most effective structures for neutralising the virus.

Together with Sanofi, we did experiments that showed that 82% of the chosen antibodies had better binding affinities than the first candidates. This efficiency underscores AbMap’s ability to reduce the time and cost of early-stage drug development. Singh underscores the significance of this approach by asserting, "Drug companies are hesitant to concentrate their efforts on a single candidate." They prefer a set of strong candidates to ensure they have options if one fails during preclinical trials.”

AbMap has transformative potential for healthcare systems by addressing critical challenges across drug development, personalised medicine, vaccine design, and immune system research. It can predict early on which antibody candidates will work, which cuts down on the time and money needed for preclinical and clinical trials. This makes it easier to get new treatments to market. By analysing structural differences in antibodies across individuals, AbMap enables the design of personalised therapies that offer tailored solutions for autoimmune diseases, infectious diseases, and cancers. Its ability to model how antibodies and antigens interact also helps the development of vaccines, guiding the creation of vaccines that make immune responses stronger and more targeted. Beyond these applications, AbMap also deepens our understanding of immune system variability, uncovering why certain individuals respond differently to infections and paving the way for preventive strategies against diseases like HIV and severe infections. This convergence of capabilities positions AbMap as a pivotal tool for advancing precision medicine and transforming health systems.

Beyond drug discoveries, AbMap offers profound insights into personalised medicine by addressing critical questions about individual immune responses. For example, why do some people develop severe COVID-19 while others remain asymptomatic? Or why are some individuals naturally resistant to HIV?

Traditional approaches, such as single-cell RNA sequencing, focus on antibody sequence comparisons. However, these methods often overlook functional similarities between structurally similar antibodies with different sequences. AbMap bridges this gap by prioritising structural predictions, revealing significantly more overlap in antibody repertoires across individuals than previously thought.

“This is where a language model fits beautifully,” Singh notes. “It combines the scalability of sequence analysis with the accuracy of structure-based predictions.”

As researchers refine AbMap, its applications are expected to expand, including exploring immune responses to diverse pathogens. This technology, which blends computational biology and immunology, represents a scalable and precise approach to understanding the immune system.

Tools like AbMap have the potential to revolutionise the treatment and prevention of diseases for healthcare systems worldwide. By offering faster, more cost-effective solutions, it brings us closer to an era where treatments are not only effective but also tailored to individual needs—ushering in the promise of precision medicine.