Unlocking antibody assembly secrets with AI

During this unique study, scientists sought to understand if artificial intelligence could be used to predict how the interior of antibodies are assembled in the body. Antibodies, which are comprised of 'heavy' and 'light' protein chains, are produced by B cells within the immune system and protect against viruses and bacteria.

Franca Fraternali, Professor of Integrative Computational Biology at University College London, said:

"Until now, it was widely assumed that the pairing of heavy and light chains within antibodies occurred at random. Using Immunomatch, we show for the first time that this assembly is, in fact, highly specific. Understanding these pairing rules is crucial for predicting antibody stability and performance and opens the door to the rational design of more effective therapeutics."

To learn more, scientists created ImmunoMatch, based on an antibody-specific language model, which was applied to heavy and light chain antibody sequences collected from millions of single human B cells. The AI model was able to identify and predict pairings of chains, giving scientists an invaluable insight into how antibodies are combined.

The team also showed that ImmunoMatch can accurately analyze antibody sequences from immune cells actively responding to disease, including those from hematological cancers and B cells within solid tumours. These insights could accelerate the rational design of new therapeutic antibodies.

Professor Deborah Dunn-Walters, Professor of Immunology, at the University of Surrey, said:

"Using AI has helped us discover that the combinations of 'heavy' and 'light' chains are not as random as we previously thought.

"This information enables us to learn the nature-derived rules governing how proteins are combined to create functional antibodies.

"Antibodies are the single largest class of modern therapeutics. Around a quarter of all newly approved therapeutics are monoclonal antibodies, so understanding how antibodies are made is critical for their effective design."

This study was published in Nature Methods.