
from board games to protein design
(Nanowerk News) Scientists have successfully applied reinforcement learning to challenges in molecular biology.
A team led by UW Medicine researchers developed powerful new protein design software adapted from strategies proven proficient in board games such as chess and Go. In one experiment, a protein made with the new approach was found to be effective at producing useful antibodies in mice.
Reinforcement learning is a type of machine learning in which computer programs learn to make decisions by trying different actions and receiving feedback. Such an algorithm can learn to play chess, for example, by testing millions of different moves that result in wins or losses on the board. This program is designed to learn from this experience and get better at making decisions over time.
To create reinforcement learning programs for protein design, scientists give computers millions of simple starting molecules. The software then performs ten thousand attempts to randomly boost each one towards a predetermined goal. The computer elongates the protein or bends it in a certain way until it learns how to turn it into the desired shape.
The findings are reported in Science (“Top-down protein architecture design with reinforcement learning”), indicating that this breakthrough will soon lead to a more efficacious vaccine. More broadly, the approach could lead to a new era in protein design.
“Our results show that reinforcement learning can do much more than mother board games. When trained to solve long-standing puzzles in protein science, the software excels at creating useful molecules,” said senior author David Baker, professor of biochemistry at the UW School of Medicine in Seattle and recipient of the 2021 Breakthrough Prize in Life Sciences.
“If this method is applied to the right research problem,” he says, “it can accelerate progress in various scientific fields.”
This research is a milestone in leveraging artificial intelligence to conduct protein science research. The potential applications are wide ranging, from developing more effective cancer treatments to making new textiles that are biodegradable.
Isaac D. Lutz, Shunzhi Wang, and Christoffer Norn, all members of the Baker Lab, led the research.
“Our approach is unique in that we use reinforcement learning to solve the problem of making protein shapes fit together like pieces of a puzzle,” explained co-lead author Lutz, a doctoral student at the UW Medicine Institute for Protein Design. “This was not possible using previous approaches and has the potential to change the types of molecules we can build.”
As part of this research, scientists produce hundreds of AI-designed proteins in the laboratory. Using electron microscopy and other instruments, they confirmed that many computer-generated forms of protein were indeed realized in the laboratory.
“This approach proved to be not only accurate but also highly customizable. For example, we asked the software to create a spherical structure with no holes, pinholes, or large holes. Its potential for fabricating all kinds of architectures has not been fully explored yet,” said lead co-author Shunzhi Wang, a postdoctoral scholar at the UW Medicine Institute for Protein Design.
The team concentrated on designing new nanoscale structures consisting of many protein molecules. This requires designing the protein components themselves and the chemical interfaces that allow the nanostructures to self-assemble.
Electron microscopy confirmed that many AI-designed nanostructures can be formed in the laboratory. As a measure of how accurate the design software is, scientists observed many unique nanostructures in which each atom was found to be where it was intended. In other words, the deviation between intended and realized nanostructures is on average less than one atom wide. This is called an atomically accurate design.
The authors foresee a future in which this approach will enable them and others to make therapeutic proteins, vaccines, and other molecules that cannot be made using previous methods.
Researchers from the UW Medicine Institute for Stem Cell and Regenerative Medicine used a primary cell model of vascular cells to show that the designed protein scaffolds outperform previous versions of the technology. For example, because receptors that help cells receive and interpret signals are clustered more densely on denser scaffolds, they are more effective at increasing vascular stability.
Hannele Ruohola-Baker, a UW School of Medicine professor of biochemistry and one of the study’s authors, spoke of the implications of the investigation for regenerative medicine: “The more accurate the technology, the more it opens up potential applications, including vascular treatment. for diabetes, brain injury, stroke and other cases where blood vessels are at risk. We can also envision more precise delivery of the factors we use to differentiate stem cells into different cell types, giving us new ways to regulate cell development and aging processes.”