Researchers at the Institute for Molecular Science (IMS) and the Graduate University for Advanced Studies, SOKENDAI, have developed a new method called AF3-ReD, which introduces a repulsive force between predicted structures to help AlphaFold3 sample multiple protein conformational states. According to the research group of Jun Ohnuki and Kei-ichi Okazaki, this novel sampling scheme overcomes the default tendency of AlphaFold to predict only a single conformation, opening new avenues for drug design and life sciences.
The Challenge of Protein Conformational Changes in AlphaFold3
Composed of amino acids linked together in chains, proteins fold into three-dimensional shapes defined entirely by their specific amino acid sequences. During their functional activities—such as transporting or synthesizing various substances—these macromolecules shift between different physical forms known as conformational states, driven by triggers like the binding of a ligand (a molecule that attaches to the protein). While researchers at Google DeepMind developed AlphaFold, an AI that achieves highly accurate structure prediction, and the researchers, John Jumper and Demis Hassabis, shared the 2024 Nobel Prize in Chemistry for this achievement, AlphaFold is known to predict only a single conformation for many proteins, thereby restricting its overall utility across life sciences fields including drug discovery.
Did you know? Aimed at structure prediction, AlphaFold3 (AF3) utilizes a diffusion generative model—a sophisticated category of AI technology also frequently deployed in image generation—by first scattering protein atoms randomly via noise into an initial state, and subsequently stripping away that noise to guide the atoms toward high-probability coordinates, with findings to be published online in JACS Au.
How AF3-Red Uses Repulsive Force Sampling
To capture diverse protein shapes, Jun Ohnuki and Kei-ichi Okazaki’s research team executed the AF3 structure prediction process repeatedly while incorporating a bias energy term designed to increase the energy level whenever a fresh prediction nears the atomic coordinates of any previously generated structure. Built straight into the AF3 diffusion model, this bias creates a repulsive force during structure prediction that prevents the model from duplicating earlier outputs, consequently boosting the exploration of alternative conformational states.
Predicting Open and Closed States in F1Beta
The resulting scheme, named AF3-ReD, proved able to predict conformational changes in a variety of proteins. Normally, F1β features an open conformation at its ATP-binding site, which shifts into a closed form upon ATP binding. AF3, however, predicts the open conformation even for ATP-bound F1β. By contrast, AF3-ReD achieved a much broader sampling span, successfully capturing the open state, the closed state, and the intermediate conformations lying between them.
Implications for Drug Design and Molecular Dynamics
Conducting molecular dynamics simulations starting from these predicted structures should now streamline the process of examining how a protein transitions between different conformations over time. Furthermore, diffusion generative models have recently seen widespread employment not just within AlphaFold, but also for creating brand-new proteins and prospective drug compounds. Integrating the repulsive bias developed in this research into those design efforts is anticipated to facilitate a wider variety of protein and drug development possibilities.
Frequently Asked Questions
What is AF3-ReD?
AF3-ReD is a new sampling scheme developed by researchers at the Institute for Molecular Science (IMS) and the Graduate University for Advanced Studies, SOKENDAI, that introduces a repulsive force between predicted structures in AlphaFold3, allowing it to sample multiple conformational states.
Why did AlphaFold traditionally predict only one protein conformation?
According to the researchers, this conformation lies at a lower energy than the others because the diffusion generative model moves atoms down the gradient of the energy to find a low-energy folded structure.
Where are the findings published?
The results will be published online in JACS Au.
Explore More: AF3-ReD makes it possible to predict diverse protein conformations rapidly and accurately.