A Repulsive Force Fix Lets AlphaFold3 See Multiple Protein Shapes, Not Just One

AlphaFold changed structural biology by solving a problem that had stumped biochemists for decades: predicting a protein's 3D shape from its amino acid sequence alone. That work earned John Jumper and Demis Hassabis a share of the 2024 Nobel Prize in Chemistry. But there's a catch that anyone doing serious drug design or enzymology already knows about. Proteins aren't static. They flex, open, close, and shift between multiple conformational states to do their jobs, and AlphaFold has a well-documented habit of predicting just one of those states, even when the biology depends on several.

Researchers at the Institute for Molecular Science (IMS) and the Graduate University for Advanced Studies, SOKENDAI, just published a fix in JACS Au that's clever precisely because it's simple. Jun Ohnuki and Kei-ichi Okazaki didn't retrain the model or bolt on new architecture. They added a repulsive force inside AlphaFold3's existing diffusion process, and it worked well enough to pull out conformations the default model had been missing entirely.

To understand why this matters, you need to know how AlphaFold3 actually generates structures. Unlike its predecessors, AF3 uses a diffusion generative model, the same general class of AI behind image generators like DALL-E and Stable Diffusion. It starts with atoms scattered randomly by noise, then iteratively removes that noise, nudging atoms toward positions of higher probability. In physics terms, higher probability means lower energy. The model is essentially sliding downhill along an energy landscape until it settles into a folded structure.

That's a great strategy for finding a stable structure. It's a bad strategy for finding all of them, because if one conformation sits in a deeper energy well than the others, the model will gravitate there every single time. Run AF3 a hundred times on the same sequence and you'll likely get the same answer a hundred times, even if the real protein spends meaningful time in two or three other shapes.

How the Repulsive Bias Works

Ohnuki and Okazaki's approach, which they've named AF3-ReD, tackles this head-on. The method runs AF3's structure prediction repeatedly, but each time, it adds a bias energy term that penalizes the model whenever a new prediction gets too close to the atomic coordinates of a structure it already generated.

Here's the practical effect:

• Every previous prediction becomes a soft "no-go zone" in conformational space
• The diffusion process still moves atoms downhill toward low energy, but now it has to route around already-explored territory
• Repeated runs get pushed toward different low-energy basins instead of collapsing back into the same one
• No retraining, no architecture changes, just an added term during the sampling process

The team tested this on the F1β subunit of ATP synthase, a well-studied protein that's a good stress test precisely because its conformational behavior is already known experimentally. F1β normally sits with its ATP-binding site open. When ATP binds, it snaps shut. Default AF3, true to form, predicted the open conformation even when fed the ATP-bound version of the protein, missing the closed state entirely.

AF3-ReD didn't just recover both states. It sampled the open conformation, the closed conformation, and a range of intermediate structures in between, essentially mapping out the transition pathway rather than just the two endpoints. That's a meaningfully richer picture than a single static structure, and it's the kind of detail that matters if you're trying to understand how a protein actually moves, not just what it looks like at rest.

Key Takeaways

The immediate value here is obvious for anyone working in structure-based drug design. A lot of small-molecule drugs work by binding to a specific conformational state of a target protein, sometimes deliberately targeting a less common or transient shape rather than the dominant one. If your prediction tool only ever shows you the most stable conformation, you're structurally blind to an entire class of potential binding sites and mechanisms.

There's also a nice downstream benefit for computational efficiency. Molecular dynamics (MD) simulations, the gold-standard method for watching how a protein actually moves between conformations over time, are notoriously expensive to run from scratch because you often need to simulate long stretches of time just to catch a rare transition. If AF3-ReD can hand you snapshots of the open state, the closed state, and several intermediates upfront, MD simulations can start from those points and focus computational effort on the transition dynamics rather than the initial search for relevant states. That's a meaningful head start.

Diffusion models have also become the backbone of a lot of generative protein and drug design work beyond structure prediction alone, including tools that design entirely new protein sequences or candidate molecules from scratch. The repulsive bias technique described here isn't inherently tied to AlphaFold3's specific implementation. It's a general trick for pushing a diffusion model to explore more of its own output space instead of settling into whatever mode happens to be easiest to find. The paper's authors flag exactly this: applying the same bias to protein and drug design pipelines could push those tools toward more diverse, less repetitive outputs too.

None of this changes the underlying accuracy of AlphaFold3's structure predictions. AF3-ReD is a sampling enhancement layered on top of an already-trained model, not a claim that the model's fundamental understanding of protein physics has improved. But for a field where "which conformation are we even looking at" has been a persistent blind spot, a lightweight fix that requires no retraining and plugs directly into existing AF3 workflows is the kind of practical improvement that's likely to get adopted fast, especially by teams already running AF3 predictions at scale for drug discovery pipelines.