Proteins are like tiny machines that dance and switch on and off
New AI system captures proteins' full range of motion and structural changes, unlike AlphaFold which produces static snapshots. Framework uses graph neural networks to compress protein data, then reconstructs high-resolution dynamic models for drug targeting.
- EPFL researchers developed LD-FPG, an AI framework that generates complete, all-atom protein models including dynamic movements
- The system uses graph neural networks to compress protein structure data into a simplified map, then reconstructs high-resolution models
- Framework successfully modeled dopamine D2 receptor in both active and inactive states, with dataset released openly for research
EPFL researchers developed LD-FPG, an AI framework that generates complete, dynamic all-atom protein models including side-chain movements, advancing drug discovery beyond static structure prediction.
The lock-and-key metaphor has long guided drug discovery: find the right molecule, fit it into a protein's binding pocket, and trigger the cellular cascade that heals or harms. But proteins are not locks. They are machines in constant motion, their atoms shifting and rearranging in ways that determine whether a drug will actually work. For decades, the best computational tools available could only capture static snapshots—a single frozen moment in a protein's life. That limitation has just shifted.
Researchers at EPFL's School of Life Sciences have developed an artificial intelligence framework called Latent Diffusion for Full Protein Generation, or LD-FPG, that does something previous systems could not: it generates complete, all-atom models of proteins in motion, including the subtle rearrangements of side chains that govern how proteins interact with other molecules. The work, published in the Proceedings of NeurIPS 2025, represents a fundamental change in how computational biologists can approach drug design.
The problem LD-FPG solves is both conceptual and practical. Systems like Google DeepMind's AlphaFold revolutionized structural biology by predicting the spatial position of every atom in a protein—but doing so requires enormous computing power and deep expertise in both biology and computer science. The EPFL team, led by Patrick Barth of the Laboratory of Protein and Cell Engineering and Pierre Vandergheynst of the Signal Processing Laboratory, took a different path. Instead of trying to predict the exact coordinates of every atom, they built a model that learns a simplified, low-dimensional map of how a protein's shape changes over time. "Proteins are like tiny machines that dance and switch on and off to work," says Aditya Sengar, a researcher in Barth's lab, "but generating this 'movie' in full detail has been an unsolved challenge. Our LD-FPG framework is the first to do this."
The technical approach hinges on graph neural networks—a way of representing proteins as mathematical graphs where atoms are nodes and chemical bonds are edges. This compressed representation allows the AI to learn the underlying patterns of protein motion without the computational burden of tracking every atom individually. Once trained, the model generates new latent data for entirely novel protein structures. Those simplified representations are then converted back into high-resolution, fully detailed proteins complete with side chains and dynamic movements.
The team tested the framework on the dopamine D2 receptor, one of the most-studied G-protein coupled receptors in the world. GPCRs are a major focus of the global drug development industry because they sit on cell membranes and respond to external signals. The researchers generated high-fidelity, dynamic models of the dopamine D2 receptor in both its active and inactive states—capturing the full range of motion for a complex drug target that has resisted detailed dynamic modeling. They have released this dataset with open access to accelerate further research.
The implications for drug discovery are substantial. Current virtual screening processes involve considerable trial and error; computational models help narrow the field, but they remain crude approximations of biological reality. A tool that can model not just a protein's static shape but its actual movements could transform how researchers identify promising drug candidates. "LD-FPG opens the door to designing new medicines that target a protein's dynamic behavior, not just its shape," Barth says. "Our work represents a new paradigm for computational biology, and a meaningful step forward at the interface of AI and structural biology."
The team is already working to refine the framework for greater accuracy and to scale it to larger, more complex proteins. But Vandergheynst, the signal processing expert, sounds a note of caution about the broader AI landscape. Many assume that feeding massive datasets to AI models will automatically solve scientific problems or replace human researchers. In reality, much of that data is noisy or poorly evaluated. "We need human scientists to produce the clean data and rigorous benchmarks AI requires," he says. The future of computational drug discovery, it seems, will depend not just on better algorithms but on better science.
Citações Notáveis
Proteins are like tiny machines that dance and switch on and off to work, but generating this 'movie' in full detail has been an unsolved challenge. Our LD-FPG framework is the first to do this.— Aditya Sengar, LPCE researcher
LD-FPG opens the door to designing new medicines that target a protein's dynamic behavior, not just its shape. Our work represents a new paradigm for computational biology.— Patrick Barth, Laboratory of Protein and Cell Engineering
A Conversa do Hearth Outra perspectiva sobre a história
Why does it matter that the model captures motion instead of just shape? Can't you just design a drug that fits the static structure?
A protein isn't a static target. It's constantly moving, breathing, shifting its shape. A drug that fits perfectly into one conformation might not fit at all when the protein moves into another state. If you only design against the snapshot, you miss half the biology.
So AlphaFold was incomplete?
Not incomplete—revolutionary for what it did. But it was like having a photograph of a dancer. You can see the pose, but you don't understand the choreography. LD-FPG gives you the full performance.
How does compressing the data actually help? Doesn't that lose information?
It's counterintuitive, but the compression is strategic. Instead of tracking millions of atomic coordinates, the model learns the underlying patterns of how proteins deform. It's like learning the rules of movement rather than memorizing every frame. Once you know the rules, you can generate new frames that are physically realistic.
What happens next? Is this ready for drug companies to use?
Not yet. The team is still refining it for accuracy and scaling it to larger proteins. But they've proven the concept works on a real drug target—the dopamine receptor. That's the proof that matters. The practical tools will follow.
Why did they release the dopamine receptor data openly?
Because they understand that science moves faster when everyone can build on the same foundation. One lab's dataset becomes another lab's starting point. That's how paradigm shifts actually happen.