Scalable emulation of protein equilibrium ensembles with generative deep learning.
Saved in:
| Title: | Scalable emulation of protein equilibrium ensembles with generative deep learning. |
|---|---|
| Authors: | Lewis, Sarah, Hempel, Tim, Jiménez-Luna, José, Gastegger, Michael, Xie, Yu, Foong, Andrew Y. K., Satorras, Victor García, Abdin, Osama, Veeling, Bastiaan S., Zaporozhets, Iryna, Chen, Yaoyi, Yang, Soojung, Foster, Adam E., Schneuing, Arne, Nigam, Jigyasa, Barbero, Federico, Stimper, Vincent, Campbell, Andrew, Yim, Jason, Lienen, Marten |
| Source: | Science. 8/14/2025, Vol. 389 Issue 6761, p1-10. 10p. |
| Subjects: | Protein structure, Deep learning, Molecular dynamics, Conformational analysis, Free energy (Thermodynamics) |
| Abstract: | Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that emulates protein equilibrium ensembles by generating thousands of statistically independent structures per hour on a single graphics processing unit (GPU). BioEmu integrates more than 200 milliseconds of molecular dynamics (MD) simulations, static structures, and experimental protein stabilities using new training algorithms. It captures diverse functional motions—including cryptic pocket formation, local unfolding, and domain rearrangements—and predicts relative free energies with 1 kilocalorie per mole accuracy compared with millisecond-scale MD and experimental data. BioEmu provides mechanistic insights by jointly modeling structural ensembles and thermodynamic properties. This approach amortizes the cost of MD and experimental data generation, demonstrating a scalable path toward understanding and designing protein function. Editor's summary: Proteins can typically adopt a range of conformations in solution. Molecular dynamics simulations have been used to generate ensembles with realistic probabilities, but the required computational resources are large even for simple problems. Lewis et al. developed a generative deep learning protein dynamics emulator called BioEmu that is trained on extensive molecular dynamics results, protein structures, and experimental data. Using a tiny fraction of the power used in traditional simulations, BioEmu generates accurate conformational ensembles for medium-sized soluble proteins. The ability to quickly, cheaply, and accurate generate such data will enable researchers to generate and test many more hypotheses and shift computational resources to validation and investigation of more difficult problems. —Michael A. Funk INTRODUCTION: Proteins constitute the functional building blocks of life and are central to drug discovery and biotechnology. We now have technologies to determine protein sequence and predict protein structure at the genomic scale, but this is not the case for protein function. Protein function relies on dynamical mechanisms, particularly the transitions between long-lived protein structures (conformational states) and the association with and dissociation from other proteins and ligands (compositional states). The coupling between conformational and compositional state changes, and the probability of these states under a given set of conditions (temperature, solvation, concentration), determine "how proteins work" on a molecular scale. Although biophysical experiments and molecular dynamics (MD) simulations can reveal such structure-dynamics relationships with high accuracy, these methods suffer from low throughput. RATIONALE: As a step toward solving this throughput challenge, we developed a biomolecular emulator (BioEmu) that samples the approximate equilibrium distribution of structures of single protein chains. BioEmu is a generative deep-learning system that can generate thousands of statistically independent structure samples per hour on a single graphics processing unit (GPU). BioEmu leverages AlphaFold to encode the protein sequence into a rich sequence-structure representation, which inputs into a diffusion model that efficiently samples three-dimensional structures. BioEmu was trained in three stages: It was pretrained on a processed version of the AlphaFold database (AFDB) in such a way as to incentivize the model to associate each protein sequence with a diverse set of structures. Training was continued on a vast dataset of MD simulations of thousands of proteins and more than 200 ms of aggregate simulation time. And finally, BioEmu was fine-tuned on more than 500,000 experimental protein stabilities using a technology developed here, property-prediction fine-tuning (PPFT). RESULTS: We tested BioEmu on a variety of protein systems that are dissimilar from training proteins and benchmarked its performance on three tasks: (i) Predicting known conformational changes including large domain motions, local unfolding transitions, and the formation of cryptic binding pockets while achieving success rates of sampling the known references of between 55 and 90%. (ii) Emulating equilibrium distributions of both protein folding and native-state conformational transitions that can be generated by high-throughput MD simulation, demonstrating errors in free-energy differences below 1 kcal/mol and speedups of four to five orders of magnitude. (iii) Predicting experimentally measured stabilities of folded states of small proteins by directly generating equilibrium ensembles and explaining structure-stability relationships of mutants, achieving errors below 1 kcal/mol and correlation coefficients greater than 0.6 for both absolute folding free energies and folding free-energy changes of mutants. CONCLUSION: BioEmu has various practical use cases, including complementing present MD simulation workflows, interpreting protein experiments in terms of structural mechanisms, identifying binding pockets and allosteric mechanisms in drug discovery, and generating ensembles for dynamical protein design. Our demonstration that the large upfront costs of MD simulation and experimental data generation can be amortized and that the prediction error decreases with an increasing amount of diverse training data indicates a path forward for predicting biomolecular function at the genomic scale. Illustration of the BioEmu model and workflow.: BioEmu generates equilibrium protein structure ensembles by combining AlphaFold's sequence representation with a diffusion model trained on vast simulation and experimental data. These ensembles enable rapid computation of properties such as protein stability, achieving speeds that are orders of magnitude faster than MD simulation. [Emu illustration by F.N.] [ABSTRACT FROM AUTHOR] |
| Copyright of Science is the property of American Association for the Advancement of Science and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.) | |
| Database: | Psychology and Behavioral Sciences Collection |
|
Full text is not displayed to guests.
Login for full access.
|
|
| Abstract: | Following the sequence and structure revolutions, predicting functionally relevant protein structure changes at scale remains an outstanding challenge. We introduce BioEmu, a deep learning system that emulates protein equilibrium ensembles by generating thousands of statistically independent structures per hour on a single graphics processing unit (GPU). BioEmu integrates more than 200 milliseconds of molecular dynamics (MD) simulations, static structures, and experimental protein stabilities using new training algorithms. It captures diverse functional motions—including cryptic pocket formation, local unfolding, and domain rearrangements—and predicts relative free energies with 1 kilocalorie per mole accuracy compared with millisecond-scale MD and experimental data. BioEmu provides mechanistic insights by jointly modeling structural ensembles and thermodynamic properties. This approach amortizes the cost of MD and experimental data generation, demonstrating a scalable path toward understanding and designing protein function. Editor's summary: Proteins can typically adopt a range of conformations in solution. Molecular dynamics simulations have been used to generate ensembles with realistic probabilities, but the required computational resources are large even for simple problems. Lewis et al. developed a generative deep learning protein dynamics emulator called BioEmu that is trained on extensive molecular dynamics results, protein structures, and experimental data. Using a tiny fraction of the power used in traditional simulations, BioEmu generates accurate conformational ensembles for medium-sized soluble proteins. The ability to quickly, cheaply, and accurate generate such data will enable researchers to generate and test many more hypotheses and shift computational resources to validation and investigation of more difficult problems. —Michael A. Funk INTRODUCTION: Proteins constitute the functional building blocks of life and are central to drug discovery and biotechnology. We now have technologies to determine protein sequence and predict protein structure at the genomic scale, but this is not the case for protein function. Protein function relies on dynamical mechanisms, particularly the transitions between long-lived protein structures (conformational states) and the association with and dissociation from other proteins and ligands (compositional states). The coupling between conformational and compositional state changes, and the probability of these states under a given set of conditions (temperature, solvation, concentration), determine "how proteins work" on a molecular scale. Although biophysical experiments and molecular dynamics (MD) simulations can reveal such structure-dynamics relationships with high accuracy, these methods suffer from low throughput. RATIONALE: As a step toward solving this throughput challenge, we developed a biomolecular emulator (BioEmu) that samples the approximate equilibrium distribution of structures of single protein chains. BioEmu is a generative deep-learning system that can generate thousands of statistically independent structure samples per hour on a single graphics processing unit (GPU). BioEmu leverages AlphaFold to encode the protein sequence into a rich sequence-structure representation, which inputs into a diffusion model that efficiently samples three-dimensional structures. BioEmu was trained in three stages: It was pretrained on a processed version of the AlphaFold database (AFDB) in such a way as to incentivize the model to associate each protein sequence with a diverse set of structures. Training was continued on a vast dataset of MD simulations of thousands of proteins and more than 200 ms of aggregate simulation time. And finally, BioEmu was fine-tuned on more than 500,000 experimental protein stabilities using a technology developed here, property-prediction fine-tuning (PPFT). RESULTS: We tested BioEmu on a variety of protein systems that are dissimilar from training proteins and benchmarked its performance on three tasks: (i) Predicting known conformational changes including large domain motions, local unfolding transitions, and the formation of cryptic binding pockets while achieving success rates of sampling the known references of between 55 and 90%. (ii) Emulating equilibrium distributions of both protein folding and native-state conformational transitions that can be generated by high-throughput MD simulation, demonstrating errors in free-energy differences below 1 kcal/mol and speedups of four to five orders of magnitude. (iii) Predicting experimentally measured stabilities of folded states of small proteins by directly generating equilibrium ensembles and explaining structure-stability relationships of mutants, achieving errors below 1 kcal/mol and correlation coefficients greater than 0.6 for both absolute folding free energies and folding free-energy changes of mutants. CONCLUSION: BioEmu has various practical use cases, including complementing present MD simulation workflows, interpreting protein experiments in terms of structural mechanisms, identifying binding pockets and allosteric mechanisms in drug discovery, and generating ensembles for dynamical protein design. Our demonstration that the large upfront costs of MD simulation and experimental data generation can be amortized and that the prediction error decreases with an increasing amount of diverse training data indicates a path forward for predicting biomolecular function at the genomic scale. Illustration of the BioEmu model and workflow.: BioEmu generates equilibrium protein structure ensembles by combining AlphaFold's sequence representation with a diffusion model trained on vast simulation and experimental data. These ensembles enable rapid computation of properties such as protein stability, achieving speeds that are orders of magnitude faster than MD simulation. [Emu illustration by F.N.] [ABSTRACT FROM AUTHOR] |
|---|---|
| ISSN: | 00368075 |
| DOI: | 10.1126/science.adv9817 |
