Antibody treatments may be able to activate the immune system to fight diseases such as Parkinson’s disease, Alzheimer’s disease, and colorectal cancer, but when antibody treatments bind to themselves or other molecules that are not markers of the disease, It will be less effective.
Now, a new machine learning algorithm developed at the University of Michigan can highlight problem areas in antibodies that tend to bind to non-target molecules.
This model can be used to pinpoint the location of the antibodies that are causing the problem and change their location to correct the problem without creating new problems. â
Peter Tessier, UM’s Albert M. Mattox Professor of Pharmacy and corresponding author of the study in the journal Nature Biomedical Engineering
âThis model is useful because it can be used with existing antibodies, new antibodies in development, and even antibodies that have not yet been created.â
Antibodies fight diseases by binding to specific molecules called antigens on the disease-causing agent, such as the spike protein on the virus that causes COVID-19. Once bound, the antibodies either directly inactivate the harmful virus or cell or signal the body’s immune cells to do so.
Unfortunately, antibodies designed to bind very strongly and quickly to a specific antigen can also bind to non-antigenic molecules, eliminating the antibody before it can target the disease. Such antibodies tend to combine with other antibodies of the same type, in the process forming a thick solution that does not easily pass through the needle that delivers the antibody drug.
“The ideal antibody should do three things at the same time: bind tightly to what it’s supposed to, repel each other, and ignore everything else in the body,” Tessier said. .
Antibodies that don’t check all three boxes are unlikely to be successful as drugs, but many clinical-stage antibodies do not. In the new study, Tessier’s team measured the activity of 80 clinical-grade antibodies in the lab and found that 75% of the antibodies interacted with the wrong molecule, interaction, or both.
By changing the amino acids that make up antibodies, and thus the 3D structure of antibodies, it may be possible to prevent antibodies from malfunctioning. This is because the structure of the antibody determines what it can bind to. But some changes can cause more problems than fixes, and the average antibody has hundreds of different amino acid positions that can be changed.
âIt takes about two working days in our model to examine all the changes in a single antibody, which is significantly less than measuring each modified antibody experimentally. “At best, this can take several months,” says Emily Makowsky, a recent Ph.D. She is a pharmacy graduate and the study’s lead author.
The team’s model, trained on experimental data collected from clinical-stage antibodies, is able to identify ways to modify the antibody, allowing it to check all three boxes with 78% to 88% accuracy . This narrows down the number of antibody modifications that chemical and biomedical engineers need to manufacture and test in their labs.
âMachine learning is the key to accelerating drug development,â said Tiexin Wang, a doctoral student in chemical engineering and study co-author.
Biotech companies are already beginning to realize the potential of machine learning to optimize the next generation of therapeutic antibodies.
“Some companies have developed antibodies that they’re very excited about because they have desirable biological activities, but we know that there are problems when trying to use these antibodies as drugs.” said Tessier. “So we’re pointing out specific places where antibodies need to be modified and helping several companies do this.”
This research was funded by the Center for Biomolecular Interaction Technology, the National Institutes of Health, the National Science Foundation, and the Albert M. Mattox Chair, and was conducted in collaboration with the Biointerface Institute and EpiVax Inc.
The University of Michigan and Sanofi have filed a patent for an experimental method that provides data used to train the algorithm.
Mr. Tessier has received honoraria for invited talks related to this research from GlaxoSmithKline, Bristol-Myers Squibb, Janssen, and Genentech.
Tessier is also a professor of chemical engineering and biomedical engineering.
sauce:
Reference magazines:
EK Makowsky, other. (2023). Optimizing therapeutic antibodies for reducing self-association and nonspecific binding with interpretable machine learning. natural biomedical engineering. doi.org/10.1038/s41551-023-01074-6.