henhaobio
Sign in

ScienceMolecular BiologyPeer-reviewed journal article

Scalable emulation of protein equilibrium ensembles with generative deep learning

Sarah LewisiD, Tim HempeliD, José Jiménez-LunaiD, Michael GasteggeriD, Yu XieiD, Andrew Y. K. Foong, Víctor García Satorras, Osama AbdiniD, Bastiaan S. Veeling, Iryna ZaporozhetsiD, Yaoyi CheniD, Soojung YangiD, Adam FosteriD, Arne SchneuingiD, Jigyasa NigamiD, Federico Barbero, Vincent StimperiD, Andrew M. CampbelliD, Jason YimiD, Marten LieneniD, Yu ShiiD, Shuxin ZhengiD, Hannes SchulziD, Usman MuniriD, Roberto Sordillo, , Frank NoéiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Following the sequence and structure revolutions, predicting the dynamical mechanisms of proteins that implement biological function remains an outstanding scientific challenge. Several experimental techniques and molecular dynamics (MD) simulations can, in principle, determine conformational states, binding configurations and their probabilities, but suffer from low throughput. Here we develop a Biomolecular Emulator (BioEmu), a generative deep learning system that can generate thousands of statistically independent samples from the protein structure ensemble per hour on a single graphical processing unit. By leveraging novel training methods and vast data of protein structures, over 200 milliseconds of MD simulation, and experimental protein stabilities, BioEmus protein ensembles represent equilibrium in a range of challenging and practically relevant metrics. Qualitatively, BioEmu samples many functionally relevant conformational changes, ranging from formation of cryptic pockets, over unfolding of specific protein regions, to large-scale domain rearrangements. Quantitatively, BioEmu samples protein conformations with relative free energy errors around 1 kcal/mol, as validated against millisecond-timescale MD simulation and experimentally-measured protein stabilities. By simultaneously emulating structural ensembles and thermodynamic properties, BioEmu reveals mechanistic insights, such as the causes for fold destabilization of mutants, and can efficiently provide experimentally-testable hypotheses.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 2 (25 Feb 2025) · CC BY-NC-ND 4.0

BioEmu, a diffusion model trained on structures, molecular dynamics trajectories, and stability assays, samples protein equilibrium ensembles and folding free energies within ~1 kcal/mol in GPU-hours.

Experimental biophysics and all-atom molecular dynamics (MD) simulations can resolve protein conformational states and their thermodynamic probabilities, but both are too slow and costly to scale across proteomes. By integrating clustered structural databases, over 200 milliseconds of MD trajectories, and high-throughput experimental folding free energies, BioEmu emulates equilibrium structural ensembles four to five orders of magnitude faster than MD.

Key results

  1. ShownBioEmu qualitatively samples functionally relevant domain motions, local unfolding transitions, and cryptic binding pockets in proteins with low sequence similarity to the training set.

    Across curated benchmarks with <= 40% sequence similarity to training data, BioEmu sampled 85% of reference domain-motion structures within <= 3 A RMSD, captured 72% (locally folded) and 74% (locally unfolded) states across 20 proteins, and sampled 85% of holo cryptic-pocket structures (49% of apo structures) within <= 1.5 A RMSD. · 3 Sampling conformational changes related to…; Figure legends

  2. ShownFine-tuned BioEmu models quantitatively reproduce two-dimensional free energy landscapes of fast-folding proteins and CATH domains at force-field-level accuracy.

    Against ground-truth all-atom MD trajectories, leave-one-out DESRES-finetuned models achieved a mean average error of 0.74 kcal/mol across 12 fast-folding proteins (0.30 kcal/mol for BBA up to 1.63 kcal/mol for lambda-repressor), and the main BioEmu model achieved a macrostate free energy MAE of 0.91 kcal/mol across 17 test CATH domains with > 100 us of MD data. · 4 Emulating MD equilibrium distributions; Figure legends

  3. ShownBioEmu generates equilibrium conformational ensembles four to five orders of magnitude faster than all-atom molecular dynamics simulations.

    Drawing 10,000 independent conformations required < 1 GPU-minute for Chignolin to ~1 GPU-hour for lambda-repressor on a single NVIDIA Titan V GPU, compared with 2,000 GPU-hours (Chignolin) to > 100,000 GPU-hours (NTL9) for DESRES MD simulations. · 2 Model; 4 Emulating MD equilibrium distributions; Figure legends

  4. ShownProperty-prediction fine-tuning (PPFT) enables BioEmu to predict experimental folding free energies and mutant stability changes while providing structural explanations.

    Trained with 8-step PPFT on MEGAscale data, BioEmu achieved a folding free energy (Delta G) MAE < 0.8 kcal/mol and Spearman r > 0.65, sampled folded states (native contacts > 0.65) for 26 stable ProThermDB proteins (Delta G < -8 kcal/mol), and linked destabilizing mutants to structure loss in HHH_rd1_0335 (I7P Delta Delta G 1.3 vs 2.1 kcal/mol), 2JWS (I24D, 1.5 vs 2.9 kcal/mol), and 2ZW1 (1.9 vs 1.2 kcal/mol). · 5 Predicting protein stabilities; S.3.6 Training on folding free energies via…; Figure legends

Methods, data and reproducibility

BioEmu pairs frozen AlphaFold2 Evoformer representations with an invariant point attention (IPA) denoising diffusion score model over coarse-grained backbone frames (C-alpha coordinates and SO(3) orientations) sampled in 100 second-order steps. Pretrained on AlphaFold Database sequence clusters and fine-tuned on >200 ms of reweighted Amber all-atom MD trajectories and ~776,000 MEGAscale Delta G values via PPFT.

  • Clustered AlphaFold Database (AFDB, July 2024 snapshot) with structural diversity for pretraining · multiple species · n = >93 million initial AFDB clusters filtered to structurally diverse clusters
  • All-atom molecular dynamics simulation datasets (octapeptides, CATH1, CATH2, MEGAsim, DDR1, SETD8, SARS-CoV-2, MHC2, Barnase-Barstar) · multiple species · n = >200 ms MD (1,100 octapeptides, ~1,100 CATH domains, 271 WTs, 21,458 mutants) · https://osf.io/4r8x2/
  • DESRES fast-folding protein all-atom MD trajectories generated on the Anton supercomputer · multiple species · n = 12 proteins (8.2 ms cumulative simulation time)
  • MEGAscale high-throughput experimental protein folding stability dataset (cDNA display proteolysis) · multiple species · n = ~776,000 entries (over 750,000 experimental measurements)
  • Multi-conformation and stability benchmarks (OOD60, domain motion, cryptic pocket, local unfolding, ProThermDB, CALVADOS IDPs) · multiple species · n = 19 OOD60, 22 domain motion, 34 cryptic pocket, 21 unfolding, 26 ProTherm, 65 IDP
Methods & tools
BioEmu, AlphaFold2, OpenMM, GROMACS, MMseqs2, Foldseek, ColabFold, AFCluster, AlphaFlow, pdbecif, BioPython, scikit-learn
Software & environment
foldseek (version 9.427df8a), pdbecif (version 1.5), mmseqs2 (version 15.6f452), Uniclust30 (August 2018)
Assay conditions & compute
Inference: <1 GPU-minute to ~1 GPU-hour per 10,000 samples on a single NVIDIA Titan V GPU; in-house MD simulations run on Azure NC4as T4 v3 compute instances
Reproducibility gaps
DESRES fast-folding protein trajectories were used under license; public repository URLs for the in-house generated MD datasets (octapeptides, CATH1, CATH2, MEGAsim) are not printed in the text.
Limitations (3)
  • BioEmu models only single soluble protein chains with a coarse-grained backbone representation at a fixed temperature of 300 K, omitting explicit side-chains, multi-chain complexes, membranes, small-molecule ligands, and conditioning on temperature or pH. · stated by the authors
  • The generative model samples conformational distributions empirically without providing an explicit statistical-mechanical potential energy function u(x) for exact Boltzmann reweighting. · stated by the authors
  • In the cryptic pocket benchmark, the model displays a strong preference for holo states (85% coverage within 1.5 A RMSD) and only recovers 49% of apo reference conformations. · stated by the authors

Read this if you study protein dynamics, cryptic pockets, or thermostability and want a generative diffusion model that samples equilibrium backbone ensembles in GPU-hours.

Citations
350
FWCI
104.62
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISClifton '26Bhakat '26Tang '24Janson '25Janson '22Fan '25
76% semantic matchBiophysics
Bryony RJ Clifton, Adam Graham Grieve, Robin Adam Corey·bioRxiv·Aug 9, 2026