Beyond AlphaFold 3’s limits: Repulsion helps predict diverse protein structures

A new technique can capture diverse structural changes in proteins that Google DeepMind’s artificial intelligence (AI) system AlphaFold 3 could not predict.

According to Asolife Sciences on the 1st (local time), researchers at the Institute for Molecular Science (IMS) of the National Institutes of Natural Sciences in Japan and the Graduate University for Advanced Studies (SOKENDAI) successfully sampled multiple structural states by introducing “repulsion” between predicted structures in AlphaFold 3.

Proteins fold into three-dimensional structures based on their amino acid sequences and perform their functions by changing structure in response to signals such as ligand binding. Predicting a folded structure from an amino acid sequence alone was a longstanding challenge, but Google DeepMind’s AlphaFold solved it with high accuracy.

For this achievement, John Jumper and Demis Hassabis jointly received the 2024 Nobel Prize in Chemistry. However, AlphaFold predicts only a single structure for many proteins, limiting its applications in the life sciences, including drug design.

The researchers explained that AlphaFold 3 predicts only certain structures because those structures have lower energy than others. AlphaFold 3 uses a diffusion generative model, which moves atoms in the direction of lower energy to find stable folded structures.

The researchers therefore repeatedly ran AlphaFold 3’s structure predictions and introduced a bias term that raises the energy as a new prediction’s atomic coordinates approach those of a previously predicted structure. This bias acts as a repulsive force, steering the model away from previous predictions and increasing the sampling of other structural states.

The resulting technique was named “AF3-ReD.” The researchers used the F1 beta subunit of adenosine triphosphate (ATP) synthase as an example. This protein normally has an open ATP-binding site, which closes when ATP binds.

Even when ATP was bound, AlphaFold 3 predicted only the open structure, while AF3-ReD broadly sampled the open and closed structures, as well as intermediate structures. The researchers said this made it possible to predict structural changes that were difficult to capture with conventional AlphaFold.

The findings were published in the international journal JACS Au. The researchers expect AF3-ReD to enable fast and accurate prediction of diverse protein structures and said that running molecular dynamics simulations on predicted structures could make it possible to study efficiently how proteins move from one structure to another.

They also noted that diffusion generative models have recently been used to design novel proteins and drug candidates, and projected that applying the repulsive bias from this study to design work could enable the design of a wider range of proteins and drugs.