Conformational changes in proteins are vital to their function yet remain challenging for state-of-the-art artificial intelligence, such as AlphaFold3, to predict. Researchers at the Institute for Molecular Science (IMS), and the Graduate University for Advanced Studies, SOKENDAI introduced a repulsive force between predicted structures, allowing AlphaFold3 to sample the multiple conformational states that its default settings rarely capture.
Proteins are chain-like molecules made of amino acids, and they fold into a three-dimensional structure determined by the sequence of those amino acids. In response to cues such as the binding of a ligand (a molecule that attaches to the protein), they then switch between different shapes of that structure, known as conformational states, to carry out their functions, such as synthesizing or transporting substances. Predicting the folded structure from the amino acid sequence alone had been a long-standing challenge in the protein sciences, until researchers at Google DeepMind developed AlphaFold, an AI that achieves highly accurate structure prediction. For this achievement, the researchers, John Jumper and Demis Hassabis, shared the 2024 Nobel Prize in Chemistry. However, while proteins function by switching between multiple conformational states, AlphaFold is known to predict only a single conformation for many proteins, thereby limiting its applicability to the life sciences, including drug design. Therefore, the research group of Jun Ohnuki and Kei-ichi Okazaki at the Institute for Molecular Science (IMS), National Institutes of Natural Sciences, and the Graduate University for Advanced Studies, SOKENDAI, set out to develop a novel AlphaFold-based method for sampling conformational changes.
The group has now developed a new sampling scheme that introduces a repulsive force between predicted structures in AlphaFold and has succeeded in predicting protein conformational changes that had been difficult for AlphaFold with its default settings. The results will be published online in JACS Au .
The latest version, AlphaFold3 (AF3), uses a diffusion generative model, a powerful class of AI also used for image generation, for structure prediction. The diffusion generative model first creates an initial state in which the atoms of the protein are scattered at random by noise, and then removes that noise, moving the atoms toward positions of higher probability. In the language of physics, positions of higher probability correspond to positions of lower energy. In other words, the diffusion generative model moves atoms down the gradient of the energy and thereby finds a low-energy folded structure (Figure 1A). The reason AF3 predicts only one particular conformation is that this conformation lies at a lower energy than the others.
The researchers therefore repeated the AF3 structure prediction multiple times and introduced a bias energy term that raises the energy whenever a new prediction approaches the atomic coordinates of a previously predicted structure (Figure 1B). With this bias built into the AF3 diffusion model, a repulsive force acts during structure prediction so that the model avoids approaching earlier predictions, which enhances the sampling of other conformational states.
The resulting scheme, named AF3-ReD, proved able to predict conformational changes in a variety of proteins. Figure 2 shows, as an example, the structure prediction of the F 1 β subunit of adenosine triphosphate (ATP) synthase. F 1 β normally adopts a conformation with its ATP-binding site open, and changes to a closed conformation when ATP binds. AF3, however, predicts the open conformation even for ATP-bound F 1 β. AF3-ReD, by contrast, sampled a far wider range, reaching the open and closed conformations as well as intermediate conformations between them. Introducing a repulsive force between predicted structures thus makes it possible to predict conformational changes that had been difficult for AlphaFold with its default settings.
AF3-ReD makes it possible to predict diverse protein conformations rapidly and accurately. Running molecular dynamics simulations from the predicted structures should now also make it efficient to investigate how a protein moves from one conformation to another over time. Moreover, diffusion generative models have in recent years been used not only in AlphaFold but also in the design of novel proteins and drug candidates. Applying the repulsive bias introduced in this study to such design work is expected to enable more diverse protein and drug design.
JACS Au
Computational simulation/modeling
Not applicable
Enhanced sampling of protein conformations in AlphaFold3 with repulsive bias in the diffusion generative model
26-Aug-2026