The newly developed experiment-guided AlphaFold technology (green) can predict protein structures across a wider conformational space than the existing AlphaFold3 (red). Provided by Maddipatla et al./Nature Biotechnology
Google DeepMind’s artificial intelligence (AI) tool AlphaFold, which predicts three-dimensional protein structures, has now evolved to the point where it can forecast the dynamic motions of proteins. It can predict multiple structures that proteins may adopt in real biological environments.
A research team led by Professor Alex Bronstein at the Institute of Science and Technology Austria (ISTA), together with international collaborators including Ayelet Marx, principal investigator at the MIGAL Galilee Research Institute in Israel, and Sanketh Vedula, postdoctoral researcher at Princeton University in the United States, has developed an “experiment-guided AlphaFold” technology that supplements AlphaFold3’s protein structure predictions using experimental data. The results were published on the 29th (local time) in the international journal Nature Biotechnology.
AlphaFold, which was recognized with the 2024 Nobel Prize in Chemistry, is an AI tool that predicts three-dimensional protein structures with high accuracy using only amino acid sequences. It simplifies the various three-dimensional structures that a protein can adopt into a single dominant conformation, and therefore does not fully capture the fact that some structures may change depending on diverse biological environments or experimental conditions.
This limitation stems from the fact that AlphaFold has mainly been trained on static protein structure data. In the Protein Data Bank (PDB), a protein structure database, most entries are derived from X-ray crystallography. X-ray crystallography is a method that uses X-ray diffraction to analyze the atomic arrangement of proteins. Although it is a core technique in protein structural studies, it observes proteins in their crystalline state, leading to their interpretation as fixed shapes.
In reality, proteins constantly fluctuate and change shape inside cells. Protein motion and conformational changes play a crucial role in enabling proteins to perform their functions.
The team proposed a method that combines experimental data with the AlphaFold prediction process to model ensembles of protein structures. Reflecting the fact that proteins do not exist in only one form but can switch among multiple conformations over time and under different conditions, they developed the experiment-guided AlphaFold technology.
Experiment-guided AlphaFold uses information that was previously dismissed as blurry or noisy in conventional structural analyses as valuable clues. The researchers interpreted certain regions in X-ray crystallography data that are not clearly resolved and have been labeled as “flexible regions” as potential indicators of a protein’s actual motions and diverse conformations.
Even in cases where AlphaFold3 (red) completely mispredicts key structural features, experiment-guided AlphaFold (green) generates models that match the experimentally measured values (white). Provided by Maddipatla et al./Nature Biotechnology
Experiment-guided AlphaFold is also designed to interpret more finely the evolutionary information encoded in amino acid sequences that make up proteins. Amino acids that have co-evolved are likely to be located close to each other within the protein or be functionally linked. The team explained that combining experimental data with evolutionary information allows the model to better reflect the sequence and structural context unique to each protein.
The researchers expect that experiment-guided AlphaFold will help advance protein complex prediction and inverse protein design. Inverse protein design is a technique that starts by specifying a desired three-dimensional structure and then works backward to identify the amino acid sequence that can form that structure; it plays an important role in bioengineering and drug discovery. If proteins can be designed not as single fixed structures but as ensembles that change over time, it would enable more realistic protein design.
“Proteins are highly dynamic molecules,” said Bronstein. “If we can predict this dynamism with experiment-guided AlphaFold, we may be able to comprehensively organize all structural information stored in the PDB within a few years.” He added, “We hope this will lead to next-generation prediction models that capture the diverse states of protein structures.”
Advaith Maddipatla, a PhD student at ISTA, said, “Experiment-guided AlphaFold has demonstrated that it can uncover protein conformations missed by conventional approaches,” adding, “Our ultimate goal is to develop this model into a standard tool for biological research.”
doi.org/10.1038/s41587-026-03166-5
Unlike AlphaFold3 (red), which simplifies protein structures into a single conformation, experiment-guided AlphaFold (green) uses nuclear magnetic resonance data to predict a broader ensemble of structures that proteins can adopt. The experimental model (white) is human ubiquitin protein. Provided by Sanketh Vedula
Copyright ⓒ DongA Science. All rights reserved.