New Method Lets AlphaFold3 See a Protein's Other Shapes

Researchers in Japan have found a way around one of AlphaFold's biggest blind spots: its tendency to predict only a single shape for proteins that actually switch between several. The team, from the Institute for Molecular Science (IMS) and the Graduate University for Advanced Studies, SOKENDAI, built a new method called AF3-ReD that pushes Google DeepMind's AlphaFold3 (AF3) to find shapes it otherwise misses.

Proteins are chains of amino acids that fold into three-dimensional shapes. Many switch between different folded shapes, called conformational states, to do their job — for example, changing shape when a ligand, a molecule that binds to the protein, attaches to it. AlphaFold, the AI system that made accurate structure prediction possible and earned creators John Jumper and Demis Hassabis a share of the 2024 Nobel Prize in Chemistry, normally predicts only one lowest-energy shape, even for proteins with several working states.

AlphaFold3 uses a diffusion model, the type of AI also used for image generation, which starts from a randomly scattered structure and removes noise step by step until it settles into the lowest-energy shape. Jun Ohnuki and Kei-ichi Okazaki's group added a repulsive bias to this process: each time AF3 makes a new prediction, an energy penalty pushes it away from shapes the model has already produced, steering it toward other stable conformations instead of the same one.

The team tested AF3-ReD on the F1β subunit of ATP synthase, an enzyme whose ATP-binding site normally opens and then closes once ATP, adenosine triphosphate, the molecule cells use to store energy, binds to it. Standard AF3 predicts only the open shape even when ATP is present. AF3-ReD sampled the open shape, the closed shape, and several conformations in between. The results are published in the journal JACS Au.

The researchers say running molecular dynamics simulations, computer models tracking how atoms move over time, from AF3-ReD's structures could show how a protein moves between shapes, and that the same bias could apply to diffusion-based AI tools already used to design new proteins and drug candidates.