Protein structure changes now predictable with new AlphaFold-based method

A research group led by Assistant Professor Jun Ohnuki and Associate Professor Kei-ichi Okazaki of the Institute for Molecular Science and the Graduate University for Advanced Studies (SOKENDAI) has developed a new method that introduces repulsive forces between structures predicted by AlphaFold. Using this approach, the team succeeded in predicting protein conformational changes that were difficult to capture with conventional AlphaFold. The findings were published online in JACS Au.

AlphaFold is widely recognized as a highly accurate model for predicting protein structures from amino acid sequences. Its developers, Dr. John Jumper and Dr. Demis Hassabis, were awarded the 2024 Nobel Prize in Chemistry. However, AlphaFold often predicts only a single dominant structure for many proteins and therefore cannot adequately capture the structural transitions between multiple conformations that occur during protein function.

In the latest version, AlphaFold 3 (AF3), which was used in this study, protein structure prediction is based on a diffusion generative model, an AI technology also used in image generation and other applications.

In diffusion generative models, an initial state is created in which the atoms that make up a protein are randomly dispersed by noise. As the noise is gradually removed, the atoms are guided toward positions with higher probabilities of occurrence. These high-probability positions correspond to lower-energy states. In other words, diffusion generative models move atoms along energy gradients to identify low-energy folded structures. The tendency of AF3 to predict only certain structures stems from the fact that those structures correspond to lower-energy states than alternative conformations.

To address this issue, the researchers repeatedly performed AF3 predictions and introduced a bias energy term that increases the energy whenever a new prediction approaches the atomic arrangement of a previously predicted structure. By incorporating this bias into the diffusion model, a repulsive force is effectively introduced during structure prediction, discouraging the model from revisiting previously identified conformations and promoting exploration of alternative structures.

The researchers demonstrated that this approach, called AF3-ReD, can predict conformational changes in a wide range of proteins. As one example, they analyzed the F₁β subunit of ATP synthase. F₁β normally adopts an open conformation at its ATP-binding site, but changes to a closed conformation upon ATP binding. However, AF3 predicts an open conformation even when ATP is bound to F₁β. In contrast, AF3-ReD was able to explore not only the open and closed conformations but also intermediate states between them.

These results show that introducing repulsive forces between predicted structures makes it possible to predict conformational changes that were difficult to capture using conventional AlphaFold. AF3-ReD enables the rapid and accurate prediction of diverse protein conformations.

In the future, these predicted structures could serve as starting points for molecular dynamics simulations, making it more efficient to investigate how proteins transition between conformations over time. Diffusion generative models are also increasingly being used to design novel proteins and drug candidate molecules.

Applying the repulsive-bias approach developed in this study to such design processes could increase the diversity of generated proteins and drug candidates, potentially enabling the discovery of novel candidates that would otherwise be difficult to obtain.

Journal Information
Publication: JACS Au
Title: Enhanced Sampling of Protein Conformations in AlphaFold3 with Repulsive Bias in the Diffusion Generative Model
DOI: 10.1021/jacsau.6c00596