Researchers at the EPFL (École Polytechnique Fédérale de Lausanne) have developed an AI-based generative framework capable of creating full, all-atom models of proteins, including their dynamic movements. The system, titled Latent Diffusion for Full Protein Generation (LD-FPG), aims to solve a long-standing limitation in computational biology where existing tools like AlphaFold typically produce only static protein snapshots rather than simulating the complex conformational shifts that govern protein function.

Proteins act as molecular machines that undergo subtle rearrangements, particularly in side chains, which dictate how they interact with drug candidates and other molecules. While static modeling remains useful, capturing these 'movies' of protein motion is essential for advancing drug discovery. The LD-FPG framework shifts the modeling approach from predicting direct atomic coordinates to learning a low-dimensional map of shape changes, significantly reducing the complexity of the task.

To achieve this, the team utilizes a graph neural network (GNN—a machine learning model that processes data structured as nodes and edges) to compress structural data into a simplified latent space. Once trained, the AI generates new protein structural ensembles that are then converted back into high-resolution, all-atom representations. The researchers demonstrated the efficacy of LD-FPG by generating dynamic representations of the dopamine D2 receptor in both active and inactive states. This receptor is a critical target for the pharmaceutical industry, and the team has released the dataset with open access to support further study.

Published in the Proceedings of NeurIPS 2025, the research suggests that modeling dynamic behavior rather than static shape could improve the accuracy of virtual drug screening processes. Researchers Patrick Barth and Pierre Vandergheynst note that while the framework offers a new paradigm for structural biology, the ultimate quality of such systems depends heavily on clean, well-evaluated input data rather than simply scaling the volume of training sets.

Researchers at the EPFL (École Polytechnique Fédérale de Lausanne) have developed an AI-based generative framework capable of creating full, all-atom models of proteins, including their dynamic movements. The system, titled Latent Diffusion for Full Protein Generation (LD-FPG), aims to solve a long-standing limitation in computational biology where existing tools like AlphaFold typically produce only static protein snapshots rather than simulating the complex conformational shifts that govern protein function.

Proteins act as molecular machines that undergo subtle rearrangements, particularly in side chains, which dictate how they interact with drug candidates and other molecules. While static modeling remains useful, capturing these 'movies' of protein motion is essential for advancing drug discovery. The LD-FPG framework shifts the modeling approach from predicting direct atomic coordinates to learning a low-dimensional map of shape changes, significantly reducing the complexity of the task.

To achieve this, the team utilizes a graph neural network (GNN—a machine learning model that processes data structured as nodes and edges) to compress structural data into a simplified latent space. Once trained, the AI generates new protein structural ensembles that are then converted back into high-resolution, all-atom representations. The researchers demonstrated the efficacy of LD-FPG by generating dynamic representations of the dopamine D2 receptor in both active and inactive states. This receptor is a critical target for the pharmaceutical industry, and the team has released the dataset with open access to support further study.

Published in the Proceedings of NeurIPS 2025, the research suggests that modeling dynamic behavior rather than static shape could improve the accuracy of virtual drug screening processes. Researchers Patrick Barth and Pierre Vandergheynst note that while the framework offers a new paradigm for structural biology, the ultimate quality of such systems depends heavily on clean, well-evaluated input data rather than simply scaling the volume of training sets.