Boltzmann-weighted conformational ensembles of multi-state proteins from sequence
Hardest of the structure problems here: ground truth is indirect, force fields cannot yet serve as truth, and the best CASP16 methods resolved both states for only about half of targets. It sits above fold switching (58), which it generalises, by roughly the gap between a special case and a general theory. Application value 45 (cryptic pockets, allostery, stability) and leverage 52 because fold switching, kinetics, design and free-energy methods all draw on it, level with the neutron-star equation-of-state problem (52).
The problem
Predicting all the shapes a flexible protein adopts, and how often each one occurs.
- 1Protein flips between shapes
- 2Odds set by free energy
- 3Simulations are slow and imperfect
- 4AI emulators trained on simulations
- 5Test against real measurements
Target. At equilibrium a protein samples conformations with probability . For two states A and B, the population ratio is ; an error of 1 kcal/mol in shifts the ratio about five-fold. The task is to output a set of structures with correct populations for proteins of up to roughly 500 residues.
Built from
earlier problems → this oneThe problems this one is built from: progress or tools it needs (builds on), special cases that come first (generalizes), and weaker results on the way (implies). On the right, what builds on it.
- Start with
- One classical force field accurate for folded proteins, disordered proteins and weak protein–protein associationComputational Chemistry · D51.5
- Predicting fold-switching proteins and both of their folds from sequenceComputational Medicine · D58
- Step 2
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceThis problem
- Then
- Predicting drug-target residence times (koff) from structure within a factor of threeComputational Medicine · D61
- De novo enzymes with native-like efficiency (kcat/KM ≈ 10⁵ M⁻¹ s⁻¹) without directed evolutionComputational Medicine · D66
- Binding affinity within 1 kcal/mol for novel scaffolds across protein familiesComputational Medicine · D63
- Prospective protein–ligand binding free energies at 1 kcal/mol RMSE across diverse, difficult targets
Path to understanding
foundations → problemThe research entries this problem rests on, in the order to read them. Each links to its explainer.
- Start with
- Gravitational N-body simulationAlgorithm
- Step 2
- Molecular dynamics for biomoleculesAlgorithm · host
- Then
- Boltzmann-weighted conformational ensembles of multi-state proteins from sequenceThe problem
State of the art
- BioEmu (Science 2025). A generative model trained on over 200 ms of MD plus experimental stability data draws thousands of independent conformations per GPU-hour and captures cryptic-pocket opening, local unfolding and domain motions; the authors report free-energy errors of about 1 kcal/mol relative to reference MD and experiment, and correlations above 0.6 with folding free energies of small proteins and their mutants. Generalisation to large multi-domain proteins, ligand-dependent states and membrane proteins is limited, and the 'reference' is itself a force-field MD.
- AlphaFlow and MSA subsampling. AlphaFold fine-tuned with flow matching (Jing et al., 2024) generates ensembles that match MD marginals better than MSA subsampling on the ATLAS MD dataset; as for BioEmu, evaluation is against simulation.
- CASP16 ensembles (2024). For ten multi-state ensembles, some groups produced reasonably accurate models of both states (TM-score above 0.75) for five targets; apart from one ligand-induced target with an apo template, predictors generally missed the structural details that distinguish states, and accuracy was clearly lower than for single-state targets. The best approaches ran many AlphaFold2 models with enhanced MSAs and sampling and then selected by quality score.
- Co-folding models. The Boltz-2 preprint states that it does not significantly deviate from AlphaFlow or BioEmu on MD ensemble tasks and "still often fails to capture large conformational changes".
- Not shown. No method has been validated against held-out experimental ensemble data at the scale and accuracy specified above.
Realistic applications
Ensemble prediction would identify cryptic and allosteric pockets (druggable states absent from crystal structures), predict the effect of mutations on stability and function, support design of conformationally switchable proteins, and supply the conformers needed for docking, free-energy and kinetics calculations.
What counts as solved
A refereed blind benchmark (CASP-style or a prospective set of newly solved multi-state proteins) with pre-registered metrics, in which one openly available method meets the state-recovery, free-energy-error and observable-agreement criteria above for at least 50 targets, with the analysis performed by an independent group and not relying on MD of the target as ground truth.
How it gets verified
ReproductionA training run, experiment or computation rerun by someone else.
One openly available method must meet the state-recovery, free-energy-error and observable-agreement criteria for at least 50 targets on a refereed blind benchmark (CASP-style or newly solved multi-state proteins) with pre-registered metrics, analysed by an independent group.
Claims under verification
+ add a claimNo claimed results on the Verification board for this problem yet.
Where the work is
Progress needs new theory for validating ensembles, large-scale compute for reference simulation, and experimental ensemble data of known populations.
Bottlenecks
ranked by votes- #1EmpiricalNo ground-truth ensembles
Experimental data (NMR, HDX, smFRET, cryo-EM heterogeneity) are indirect and under-determined; emulators are validated mostly against MD.
Community grades
Sign in to add your own grade.
Connected problems
Progress
- Sep 2, 2025CASP16 ensembles assessedsource
Both states reproduced (TM-score > 0.75) for five of ten targets; RNA and oligomers hardest.
- Jul 10, 2025BioEmu in Sciencesource
Peer-reviewed version reports free-energy accuracy of about 1 kcal/mol against MD and experiment for small proteins.
- Dec 6, 2024BioEmu preprintsource
Generative equilibrium-ensemble emulator trained on MD and stability data; thousands of samples per GPU-hour.
- Feb 7, 2024AlphaFlowsource
Flow-matching fine-tuning of AlphaFold generates protein ensembles with improved agreement to MD marginals.