AF3-ReD steers AlphaFold3 toward multiple protein conformations
AF3-ReD applies a repulsive bias against structures found in earlier AlphaFold3 prediction rounds, steering diffusion sampling into new regions. The method by Jun Ohnuki and Kei-ichi Okazaki can search for open, closed and intermediate protein conformations without retraining the model. Its structural diversity does not replace experimental validation or establish the true thermodynamic distribution of protein states.
Artificial Intelligence··Evening
Repulsive bias pushes repeated answers away
AF3-ReD applies a repulsive bias when a new AlphaFold3 prediction approaches atomic coordinates generated in earlier rounds. Once a structure has been found, that region becomes less attractive in the next round, steering diffusion sampling toward other conformational regions. The method by Jun Ohnuki and Kei-ichi Okazaki does not retrain AlphaFold3; it changes the search during generation. The aim is to find several plausible shapes a protein may adopt instead of repeatedly returning to one dominant structure.[1], [2]
Three shapes emerged for the F1 beta subunit
The researchers tested the method on the F1 beta subunit of ATP synthase. Default AlphaFold3 returned the open conformation even for an ATP-bound example. AF3-ReD reached open, closed and intermediate conformations for the same system. The result shows that repulsive bias can expose different structural clusters for one target. The finding belongs to the cases examined; the sources do not claim that the method will produce the same diversity across all proteins.[1], [2]
Diversity does not establish thermodynamic populations
The conformations generated by AF3-ReD do not calculate how long a protein occupies each state or the rate of transitions between them. The diffusion model's search landscape is also a learned generative space, not the protein's physical potential energy. The method does not replace experimental validation or molecular-dynamics simulation. Its clearest contribution is a broader set of candidate structures that may provide different starting points for those later studies.[1], [2]
