Scientists have uncovered surprising new insights into how enzymes work. These findings could help scientists design better enzymes, advance AI tools for biomedical research, and develop new therapies. Enzymes are specialized proteins that act as catalysts, speeding up chemical reactions in living organisms. Without them, our bodies would not be able to accomplish essential tasks such as digesting food, producing energy, and repairing DNA. Yet scientists still do not fully understand how small genetic changes can affect enzyme performance, which can disrupt critical biological processes in the body.  

To address this question, High-Risk, High-Reward Research (HRHR) NIH Director's Early Independence Award recipient, Margaux Pinney, Ph.D., and her research team at the University of California, San Francisco investigated this by studying the function of adenylate kinase (ADK), an essential enzyme that helps regulate energy in cells found across nearly all forms of life. They compared hundreds of different versions of ADK from organisms adapted to extreme environments, ranging from very cold to extremely hot conditions. Using a new high-speed testing method, the researchers measured how efficiently each version of the enzyme worked. They also evaluated whether artificial intelligence (AI) models could predict enzyme performance from genetic sequences alone.  

Dr. Pinney and her team developed a large-scale map of how the ADK enzyme performs. Surprisingly, they found that some versions of ADK worked up to 1,000 times faster or slower than other versions, even though their overall structures were nearly identical. The team described this variation as an “enzyme landscape,” like a map of hills and valleys where peaks represent enzymes that work very well, and valleys represent those that do not. Although the landscape is rugged, it remains connected so that genetic or structural changes to the enzyme can move them across the landscape, sometimes leading to improved performance. While enzymes evolve in different environments, the team found that environmental factors like temperature do not consistently predict how efficiently this enzyme functions.

In their research, Pinney’s lab found that AI models trained on protein sequences alone had limited ability to predict enzyme activity. However, when combined with the team’s experimental data, AI predictions became far more accurate. This highlights the importance of experimental data for building reliable AI models of enzyme function. AI models have tremendous promise for biomedical research, but it is important that they are used carefully and tested for accuracy.

The findings challenge assumptions about how enzymes evolve and show that multiple structural solutions can produce highly active enzymes. The findings also highlight the importance of pairing AI models with real-world data from large-scale experiments to improve its usefulness and provide insights into how genetic changes may impact protein function and health. These insights could help scientists design better enzymes, develop new medicines, and advance therapies.

Reference:  

Muir, D. F., Asper, G. P. R., Notin, P., Posner, J. A., Marks, D. S., Keiser, M. J., & Pinney, M. M. (2025). Evolutionary-scale enzymology enables exploration of a rugged catalytic landscape. Science, 388(6752), eadu1058. https://doi.org/10.1126/science.adu1058