Breaking through AlphaFold's limits to predict how proteins change shape

Researchers at the Institute for Molecular Science (IMS) and SOKENDAI have developed a new method to predict conformational changes in proteins using Google's AlphaFold3. Conformational changes are crucial for protein function, but AlphaFold3 typically predicts only one conformation for a given protein. The team introduced a repulsive force between predicted structures in AlphaFold3, allowing the AI to explore multiple conformational states that its default settings often miss.

By adding a bias energy term to the AlphaFold3 diffusion model, the researchers created a repulsive force that prevents the model from approaching previously predicted atomic coordinates. This new approach, named AF3-ReD, successfully predicted conformational changes in various proteins, including the F1β subunit of adenosine triphosphate (ATP) synthase.

AF3-ReD can predict open, closed, and intermediate conformations of ATP-bound F1β, which were previously difficult for AlphaFold3 to capture. This enhanced sampling of protein conformations has the potential to improve conformational modeling and downstream molecular dynamics simulations. Additionally, the repulsive bias introduced in this study could accelerate protein design and drug discovery efforts.