Many drug and antibody discovery pathways focus on intricately folded cell membrane proteins. When drug candidate molecules bind to these proteins, they trigger a chemical cascade that changes cell behavior, like a key in a lock. Therefore, understanding how proteins fold and move is essential to developing drugs that interact well with their targets.
Artificial intelligence (AI) is a very useful tool for generating new protein structures, but most systems, including Google DeepMind’s AlphaFold, focus on generating static “snapshots” of proteins. Subtle rearrangements of atoms within structures called side chains that affect the protein’s interactions with other molecules are not captured.
Scientists from the EPFL School of Life Sciences are currently working with data processing experts from the School of Engineering to solve this problem. Researchers led by Patrick Barth of the Laboratory for Protein and Cell Engineering (LPCE) and Pierre Vanderheinst of the Signal Processing Laboratory (LTS2) have developed an AI-based production framework called Latent Diffusion for Complete Protein Production (LD-FPG). This generates a complete all-atomic structural ensemble of the protein and its motion.
“Proteins are like little machines that dance and switch on and off to function, but producing this ‘movie’ in detail has been an outstanding challenge,” says LPCE researcher Aditya Sengar. “Our LD-FPG framework is the first to achieve this. Instead of trying to predict the exact coordinates of atoms in space, our model learns a low-dimensional map of protein shape changes. This conceptual change allows for the generation of all-atom dynamics.”
We note that this new framework can generate a full range of behaviors for complex drug targets such as G protein-coupled receptors (GPCRs), which are a focus of the global drug development industry.
“LD-FPG opens the door to the design of new drugs that target not only the shape of proteins but also their dynamic behavior. Our work represents a new paradigm in computational biology and a meaningful advance at the interface of AI and structural biology,” said Barth. The work has been published below. Proceedings of NeurIPS 2025.
Capturing the dance of proteins
Systems like AlphaFold use AI to predict the spatial location of every atom in a protein, which requires vast computational power and expertise in biology and computer science. LD-FPG simplifies this problem using something called a graph neural network (GNN). GNN treats each protein like a mathematical graph. Atoms represent “nodes” and bonds between them represent “edges.” This low-level representation is used to essentially compress the protein structural data into a simplified map, or potential map.
An AI model then studies this map and “learns” a representation of protein structure and movement. Once trained, the model generates latent data with an entirely new structure. Finally, these simplified data are translated into high-resolution proteins with side chains and dynamic movements.
In one experiment, the team generated high-fidelity dynamic representations of both the active and inactive states of the dopamine D2 receptor. This protein detects the neurotransmitter dopamine and controls important cellular responses, making it one of the most studied GPCRs. The researchers have made this dataset open access to facilitate further research.
“In addition to enhancing biological understanding, we believe our research will help improve the virtual screening process for proteins, which currently involves a lot of trial and error, thereby helping to accelerate drug discovery,” Senger says.
Going forward, the team aims to streamline the AI framework to further improve accuracy and realism, allowing it to model larger proteins. However, Vanderheinst emphasizes that high-quality data will continue to be the foundation of success. “Many people think that feeding large datasets to AI models will automatically solve scientific problems or replace researchers. But much of that data is noisy or poorly evaluated. Just as we need journalists to protect against disinformation, we need human scientists to generate the clean data and rigorous benchmarks that AI requires.”
sauce:
Lausanne Federal Institute of Technology
Reference magazines:
Aditya Sengar, Ali Hariri, Daniel Probst, Patrick Barth, Pierre Vanderheinst (2025). Generative modeling of all-atom protein conformations using latent diffusion on graph embeddings. NeurIPS 38 Minutes. https://papers.nips.cc/paper_files/paper/2025/hash/23be5bb3a432d3ccfe991562897ebf02-Abstract-Conference.html
