If you have an ACS member number, please enter it here so we can link this account to your membership. (optional)

ACS values your privacy. By submitting your information, you are gaining access to C&EN and subscribing to our weekly newsletter. We use the information you provide to make your reading experience better, and we will never sell your data to third party members.


Drug Development

An AI tool saves time in improving protein drugs

Researchers skip high-throughput screen while optimizing antibody affinity

by Laurel Oldach
July 9, 2024

Photo collage showing an antibody, a 96-well plate, a multiple sequence alignment and a circuit board.
Credit: Madeline Monroe/C&EN/Shutterstock/Wikimedia Commons
With a tool to predict useful mutations, researchers had to test only tens rather than millions of antibody variants.

Though scientists developed several monoclonal antibody drugs to treat COVID-19 during the pandemic, viral evolution has made most of them almost useless. “As the variants came along, each of the antibodies started losing their authorization from the FDA” because they could not neutralize new viral variants, says biochemist Peter S. Kim.

Researchers in Kim’s and Brian Hie’s laboratories at Stanford University used an artificial intelligence tool to update two outmoded antiviral antibodies—without the months of lab work that would usually accompany such a project (Science 2024, DOI: 10.1126/science.adk8946). The tool may help cut a time-consuming optimization step from future drug development efforts.

The study addresses a central problem in protein engineering: the number of possible changes to a protein is vast, but most of those changes are harmful to protein function. Therapeutic antibodies are hundreds of amino acids long. Nineteen amino acid substitutions can be made at each position, and substitutions at any number and combination of positions might be useful. The potential combinations are dizzying. To find the rare tweaks that improve an antibody-drug candidate, researchers currently conduct affinity maturation studies, screening thousands or even millions of variants in a painstaking search that can take years.

In this study, researchers used a computational tool to narrow down the search space. They began with a previously published protein language model—a neural network that can predict amino acid sequence based on protein structure. Some scientists train AI tools on functional data from high-throughput binding screens, but this software receives no more information about protein function than what is conveyed by the structure.

The team put the technique to the test using two neutralizing antibodies that were developed to treat COVID-19 but are ineffective against strains of the virus circulating today. The researchers gave the model a structure of each antibody bound to its viral target protein and constrained its suggested changes to those that don’t alter the antibody’s backbone structure. With this technique, “we can test on the order of 10s of molecules instead of millions,” Kim says. And about half the variants they tested improved the antibody. After just two rounds of evolution, the group found a variant that neutralized a pseudovirus 25 times better than the original antibody.

According to Rob Meijers of the Institute for Protein Innovation, a nonprofit that develops antibodies for research, the new approach could be very useful to pharmaceutical companies—especially those developing antibodies to viral targets. He would like to see more data on how it works for other targets and for non-antibody proteins.

This tool depends on a starting antibody and performs best when given the structure of an antibody-target complex. Meijers adds that many in the field hope that antibodies can someday be designed from scratch.



This article has been sent to the following recipient:

Chemistry matters. Join us to get the news you need.