Researchers at the Institute of Science and Technology Austria have developed a way to guide AlphaFold3 with real experimental measurements, a step that could make protein prediction more faithful to the restless behavior of molecules inside living systems. The work, described in an official announcement and published in Nature Biotechnology, aims to help the AI model produce shifting structural ensembles rather than a single preferred shape.
The advance matters because proteins rarely sit still. They bend, flex, open, close and pass through fleeting shapes that can control how cells work. AlphaFold has transformed biology by predicting protein structures from amino acid sequences with remarkable accuracy. The ISTA-led team is now pushing that predictive power toward a more dynamic view of molecular life.
AlphaFold’s one-shape problem
AlphaFold has become one of the most influential tools in modern structural biology. It can take a protein sequence and predict a three-dimensional shape that often matches experimental structures with striking precision. That achievement helped make AI a central part of biological research.
The challenge comes from the way much of structural biology has been recorded. Many databases are dominated by stable structures captured through crystallography. These entries are immensely valuable, yet they tend to present proteins as sharply defined objects. A biological protein can behave more like a moving machine.
According to the ISTA team, AlphaFold3 often collapses a heterogeneous structure into one dominant conformation. In plain terms, a protein with many possible poses may appear as one favored pose in the model. That single picture can miss local changes caused by chemical surroundings or experimental conditions.
This matters for the flexible regions of proteins. In crystal structures, loops or mobile segments may appear faint or missing because they refuse to settle into a single fixed position. Those zones can be crucial for binding partners, carrying signals, or switching a protein between active states.
The new approach treats protein ensembles as a central target. Instead of asking only for one most likely shape, the model can be guided toward a family of shapes that agrees with measurements from the lab. That gives researchers a richer picture of how proteins may behave in real biological settings.
How experiments steer the AI
The ISTA-led method uses experimental data to guide AlphaFold3 during structure generation. The team showed that the AI can be steered with measurements from NMR spectroscopy, X-ray crystallography and cryo-EM. These methods each offer a different window into molecular structure.
NMR can report on proteins in solution. X-ray crystallography can reveal high-resolution details from crystals. Cryo-EM can capture large molecular machines and complexes in frozen samples. When these measurements are incorporated into the model, AlphaFold3 can generate conformations that remain consistent with the data.
The approach also allows data about dynamics to enter the prediction process. Some measurements describe how ordered or flexible specific regions of a protein are. That information can help the model decide where a structure should stay firm and where it should explore multiple positions.
ISTA professor Alex Bronstein framed the goal in simple terms. “Proteins are highly dynamic molecules,” he said. Modeling that dynamism, he added, could reveal the functional importance of protein motion.
The study was led by Bronstein with Ailie Marx of Tel-Hai University of Kiryat Shmona and MIGAL, ISTA professor Paul Schanda and Sanketh Vedula of Princeton University and the Broad Institute. Advaith Maddipatla, a doctoral student in Bronstein’s group, is the first author of the Nature Biotechnology study.
Why protein motion matters
Protein function often depends on motion. Enzymes shift as they bind molecules. Receptors change shape as they transmit signals across cell membranes. Immune proteins can move through several arrangements while recognizing targets.
A fixed structure can still be deeply useful. It gives scientists a map of atoms and surfaces. It can reveal pockets where drugs might bind. It can show how mutations disturb a fold. The new work expands that map into something closer to a time-aware molecular portrait.
Researchers have long known that some protein regions behave like hinges, flaps, or loose loops. These motions can happen over many timescales, from rapid local flickers to slower rearrangements. The ISTA approach aims to make those changes easier to model from sequence and experimental measurements.
Paul Schanda described the ambition as a way to let the model explore functionally meaningful poses. “We want our model to be able to visit all these conformations,” he said.
That ability could be especially useful for proteins that pass quickly through important intermediate states. A single structure can capture a start or end point. An ensemble can show the routes a molecule may take as it performs its job.
A new language for molecular fuzziness
Structural biology has developed a powerful visual vocabulary. Scientists describe helices, beta sheets, loops, ribbons and cartoon representations. These terms helped generations of researchers interpret molecular architecture.
The ISTA researchers argue that protein movement calls for a richer graphical language. Static diagrams can make a molecule look frozen. Flexible zones may be represented as missing density or dashed connections, which can make them seem secondary to the folded core.
Bronstein wants those elusive regions to become part of the main story. “We want to tackle all the questions that crystallographers couldn’t answer in the past,” he said.
The team’s broader goal is to recover structural information that has been difficult to express in existing databases. The Protein Data Bank has become a foundation of biology, medicine and biotechnology. Its records have trained prediction systems and guided countless experiments. A future generation of databases may need to represent motion more directly.
That is where molecular fuzziness becomes valuable. In many experiments, a blurred region can indicate a range of real conformations. The ISTA method treats that signal as information. If an AI model can learn from it, protein prediction could move closer to the physical behavior of molecules.
What this could change in drug discovery
Drug discovery often depends on shape. A medicine may work by fitting into a pocket, blocking a moving part, or stabilizing one state of a protein over another. When a protein shifts between conformations, each state can expose different opportunities for intervention.
Experiment-guided AlphaFold3 could help researchers explore those options with more realism. A model that produces measurement-consistent ensembles can suggest which shapes are plausible under specific conditions. That could help scientists understand why some compounds bind well and others fail.
The same idea could support inverse protein design, also known as inverse folding. In that field, researchers design sequences expected to fold into desired structures. If future tools can design ensembles across time, engineered proteins could be built for movement as well as shape.
Advaith Maddipatla described the work as a path toward AI models that respond to experimental reality. “Experiment-guided AlphaFold paves the way for future predictive models that are ‘experimentally aware’,” he said.
The current work remains a proof of concept for a broader future. The researchers have also pursued related studies on faster inference and on uncovering previously unmodeled conformations in β2-microglobulin. Together, the efforts point toward structural prediction systems that treat proteins as moving ensembles, with experiments helping keep the AI grounded.
