Proteins are not fixed forms but living processes — and the tools we use to understand them must reflect that truth. Researchers at Japan's Institute for Molecular Science have extended AlphaFold3, the Nobel-recognized AI, beyond its habit of seeing only a single frozen posture, teaching it instead to witness the full choreography of shapes a protein performs. By introducing a repulsive force that steers the model away from its own prior answers, the team has opened a path toward drug design that honors the dynamic nature of life itself.
Scientists enhance AlphaFold3 to predict protein shape-shifting critical for drug design
Proteins don't sit still. They shift shapes to do their work.
Mark
So AlphaFold3 won a Nobel Prize for predicting protein structures, but it's still missing something important?
Mimi
It's accurate at what it does—predicting the lowest-energy conformation. But proteins don't sit still. They shift shapes to do their work, and AlphaFold3 doesn't naturally show you those other shapes.
Mark
Why not? If the protein actually adopts those shapes, shouldn't the model predict them?
Mimi
The model is trained to find the most stable state, the one with the lowest energy. That's usually the resting state. When a drug molecule binds, the protein changes shape, but the model doesn't know to look for that unless you tell it to.
Luke
And the researchers are telling it to look by adding a repulsive force?
Mimi
Exactly. They run the prediction multiple times and add a penalty that pushes the model away from structures it's already found. That forces it to explore other conformational states.
Mark
Does it work?
Mimi
They tested it on ATP synthase, a protein that opens and closes its binding site. Standard AlphaFold3 only predicted the open state. AF3-ReD predicted open, closed, and intermediate states.
Luke
On one protein. How many did they test it on?
Mimi
The paper says "a variety of proteins," but the detailed example is ATP synthase. The method is new, so broader validation will come.
Mark
What does this mean for drug design?
Mimi
If you're designing a drug to fit into a protein's binding site, you need to know what shapes that site can adopt. Now you can predict multiple conformations instead of guessing from one frozen picture.
Luke
And the broader application?
Mimi
The same repulsive bias technique could be applied to designing new proteins and drug candidates from scratch, not just predicting existing ones. That could open up more possibilities.
Le Pouls
- AlphaFold3 won a Nobel Prize for predicting protein structures, yet it consistently defaults to a single conformation — a critical blind spot when proteins must shift shape to bind drugs or perform their biological roles.
- This limitation leaves drug designers working from incomplete maps, unable to see the multiple postures a protein adopts in the living body.
- A team at Japan's Institute for Molecular Science developed AF3-ReD, injecting a repulsive bias into AlphaFold3's diffusion process that pushes each new prediction away from structures already found, forcing the AI to explore the full conformational landscape.
- Tested on ATP synthase's F1β subunit, AF3-ReD successfully predicted open, closed, and intermediate states — where standard AlphaFold3 saw only one.
- The technique is now poised to accelerate molecular dynamics simulations and extend into the design of novel proteins and drug candidates, broadening the reach of generative AI across the life sciences.
Proteins are not fixed forms but living processes — and the tools we use to understand them must reflect that truth. Researchers at Japan's Institute for Molecular Science have extended AlphaFold3, the Nobel-recognized AI, beyond its habit of seeing only a single frozen posture, teaching it instead to witness the full choreography of shapes a protein performs. By introducing a repulsive force that steers the model away from its own prior answers, the team has opened a path toward drug design that honors the dynamic nature of life itself.
Proteins are dynamic machines, not static sculptures — they shift and reshape in response to molecular signals, and those shape-changes are central to how they function. Yet AlphaFold3, the AI that won the 2024 Nobel Prize in Chemistry for its structural predictions, carries a significant blind spot: it reliably finds only the lowest-energy conformation, the single frozen posture, and misses the range of shapes a protein naturally adopts. For drug design, this matters enormously — a protein often must change shape to bind a drug molecule, and a model that shows only one posture leaves researchers working with an incomplete picture.
Recognizing this gap, Jun Ohnuki and Kei-ichi Okazaki at Japan's Institute for Molecular Science developed AF3-ReD. The method is elegant in its logic. AlphaFold3 uses a diffusion generative model — the same class of AI behind image generation tools — that gradually moves atoms from random noise toward lower-energy positions, like water finding its way downhill. It settles on one conformation because that conformation sits at the energy minimum. AF3-ReD introduces a repulsive bias: each time a prediction is made, a bias energy term raises the cost of landing too close to a structure already found, nudging the model to explore elsewhere.
The results were striking. When tested on the F1β subunit of ATP synthase — a protein that opens its ATP-binding site by default and closes it upon binding — standard AlphaFold3 predicted only the open state, even when told ATP was present. AF3-ReD sampled the open conformation, the closed conformation, and intermediate states in between, and the approach proved effective across multiple proteins.
The implications reach beyond structure prediction. Starting molecular dynamics simulations from this richer set of conformations should make it faster to understand how proteins transition between states — the actual mechanism of their function. And because diffusion generative models now underpin not only structure prediction but also the design of new proteins and drug candidates, applying this repulsive bias to those domains could yield more diverse and effective therapeutic molecules. Published in JACS Au, the work is a concrete step toward computational protein science that reflects the living, shifting nature of its subject.
Proteins are not static sculptures. They are dynamic machines that shift and reshape themselves in response to molecular signals, and those shape-changes are essential to how they work. Yet the artificial intelligence systems we've built to predict protein structures—most notably AlphaFold3, which won the 2024 Nobel Prize in Chemistry for its breakthrough accuracy—have a significant blind spot: they typically predict only a single conformation, the lowest-energy state, and miss the multiple shapes a protein actually adopts when it's doing its job.
This limitation has real consequences for drug design. A protein that binds a drug molecule often needs to shift into a different shape to do so, and if your computational model can only show you one frozen posture, you're working with incomplete information. Researchers at Japan's Institute for Molecular Science and the Graduate University for Advanced Studies, SOKENDAI, recognized this gap and set out to fix it. Jun Ohnuki and Kei-ichi Okazaki's team developed a new method, called AF3-ReD, that coaxes AlphaFold3 into sampling the full repertoire of shapes a protein can adopt.
The innovation is elegant in its simplicity. AlphaFold3 works by using a diffusion generative model—the same class of AI that powers image generation tools—to predict structure. The model starts with atoms scattered randomly by noise, then gradually removes that noise, moving atoms toward positions of lower energy, like water flowing downhill. The reason AlphaFold3 settles on a single conformation is that this conformation sits at the lowest energy point. The researchers introduced a repulsive force into the process: each time AlphaFold3 makes a prediction, a bias energy term raises the energy whenever a new prediction gets too close to a structure the model has already found. This repulsion pushes the model away from its default answer and forces it to explore other conformational states.
The results are striking. When the researchers tested AF3-ReD on the F1β subunit of ATP synthase—a protein that normally keeps its ATP-binding site open but closes that site when ATP actually binds—standard AlphaFold3 predicted only the open conformation, even when told the protein had ATP attached. AF3-ReD, by contrast, sampled a far wider range: the open conformation, the closed conformation, and intermediate shapes in between. The method works across a variety of proteins, not just this one example.
The practical implications extend beyond structure prediction alone. Running molecular dynamics simulations from the multiple conformations AF3-ReD predicts should make it faster and more efficient to understand how a protein transitions from one shape to another over time—the actual mechanism of its function. And because diffusion generative models are now used not only in AlphaFold but also in designing entirely new proteins and drug candidates, applying this repulsive bias technique to those design applications could enable researchers to create more diverse and effective therapeutic molecules. The work, published in JACS Au, represents a concrete step toward making computational protein science more useful to the people trying to develop the next generation of medicines.
Citations marquantes
Proteins function by switching between multiple conformational states, but AlphaFold predicts only a single conformation for many proteins, thereby limiting its applicability to drug design.— Research group rationale (Ohnuki and Okazaki, IMS/SOKENDAI)
Envie de l'histoire complète ?
Lire l'original sur EurekAlert! ↗
Technology