AI gets a protein-shape upgrade

AlphaFold3 revolutionized structural biology by predicting how proteins fold — earning its creators a 2024 Nobel Prize in Chemistry. But there's been a catch: proteins don't just hold one shape. They constantly shift between different conformational states to carry out their functions, and AlphaFold3 has largely been limited to predicting just one of those states.

Researchers at Japan's Institute for Molecular Science (IMS) have now addressed that limitation. Their new method, AF3-ReD, introduces a repulsive bias into AlphaFold3's diffusion model — essentially a force that pushes each new prediction away from previously generated structures. The result? The AI explores a much broader landscape of protein shapes, including intermediate states that were previously out of reach. In tests, AF3-ReD successfully predicted both open and closed conformations of the protein F1β, as well as intermediate states — something standard AlphaFold3 failed to do.

Key Takeaways

• AF3-ReD samples diverse protein conformations rapidly and accurately, without rebuilding AlphaFold3 from scratch
• The method can be paired with molecular dynamics simulations to study how proteins transition between states over time
• The repulsive bias approach could also be applied to AI-driven design of novel proteins and drug candidates, enabling more diverse and targeted drug development

Why it matters: Proteins are the primary targets of most drugs. Understanding how they shift shape — not just their default form — is critical for designing therapies that work precisely when and where they're needed. AF3-ReD could meaningfully accelerate that process.