New machine-learning algorithms developed at the University of Michigan offer a promising solution for improving the effectiveness of antibody treatments against diseases like Parkinson’s, Alzheimer’s, and colorectal cancer.
These algorithms can identify problematic regions in antibodies that make them prone to binding with non-target molecules, potentially reducing their therapeutic efficacy.
The research, led by Peter Tessier, the Albert M. Mattocks Professor of Pharmaceutical Sciences at U-M, was published in Nature Biomedical Engineering. The study’s key innovation is its ability to pinpoint specific positions within antibodies causing binding issues and suggest modifications without introducing new problems.
Antibodies play a vital role in fighting diseases by binding to specific antigens on disease-causing agents, such as the spike protein on the COVID-19 virus. However, antibodies designed to bind strongly and quickly to their intended targets can sometimes bind with non-antigen molecules or even other antibodies of the same type. This unintended binding can compromise the antibody’s effectiveness and result in the formation of thick solutions that are challenging to administer.
Tessier explained that the ideal antibody should fulfill three criteria: it should bind tightly to its target, repel other antibodies, and ignore unrelated molecules in the body. Antibodies failing to meet these criteria are unlikely to be successful drugs.
The research team assessed the activity of 80 clinical-stage antibodies and found that 75% of them interacted with the wrong molecules or with each other. Modifying the amino acids that constitute an antibody can alter its 3D structure, potentially preventing these undesirable interactions. However, making arbitrary changes to antibodies can create more problems than solutions, given the numerous amino acid positions in an antibody.
To address this challenge, the team developed machine-learning models trained on experimental data from clinical-stage antibodies. These models can accurately identify modifications that enhance an antibody’s behavior, meeting all three criteria, with an accuracy ranging from 78% to 88%. This significantly streamlines the antibody modification process and reduces the need for extensive lab testing.
Emily Makowski, the study’s first author and a recent Ph.D. graduate in pharmaceutical sciences, highlighted the efficiency of their models, noting that exploring modifications for a single antibody takes only about two workdays, compared to months of experimental testing.
The potential of machine learning to optimize therapeutic antibodies is gaining recognition within the biotech industry. Companies are increasingly turning to these techniques to enhance the next generation of therapeutic antibodies. Tessier mentioned that some companies have already sought their expertise in identifying specific areas within their antibodies that require modification.
In summary, the University of Michigan’s machine-learning algorithms offer a promising approach to improve the effectiveness of antibody treatments for various diseases by identifying and rectifying issues that hinder their performance, ultimately accelerating drug development and improving patient outcomes.
