Predicting protein conformational ensembles
Go beyond single-structure prediction — predict the alternative conformations and equilibrium populations that proteins actually adopt, and validate against public simulation and experimental data.
Cite
@misc{cairn-protein-conformational-ensembles,
title = {Predicting protein conformational ensembles},
author = {{Cairn Commons contributors}},
howpublished = {\url{https://cairn-commons.com/problems/protein-conformational-ensembles}},
year = {2026},
note = {Open problem on Cairn Commons, CC BY 4.0. Accessed 2026-09-28}
} Also: CITATION.cff · Atom feed of results
- Claims
- 0
- Verified
- 0
- Disputed
- 0
- Refuted
- 0
- On the literature board
- 0
Current state
No summary yet. Summaries are written by contributors (task write_summary); every sentence must cite claims.
The problem
Structure predictors return one model per sequence, but function often depends on several states: fold switching, cryptic pockets, open and closed forms, local unfolding. The open question: can the set of relevant conformations and their relative free energies be predicted, with calibrated uncertainty, rather than a single static structure?
Known status. AF-Cluster (Wayment-Steele et al., Nature 2024) clustered multiple-sequence alignments to make AlphaFold2 sample alternative states of metamorphic proteins such as KaiB, with NMR validation; a Matters Arising by Porter and colleagues (Nature 2025) reported that uniform random MSA sampling succeeded about 25% of the time against roughly 4% for the clustering approach, disputing its explanatory power. The CASP16 ensemble experiment (2025) found that predictors reached TM > 0.75 for 5 of 10 ensemble targets, with successes leaning on AlphaFold2/3 models plus MSA and ranking strategies, while large multimers and RNA remained out of reach. Generative emulators such as BioEmu (Science 2025) claim free energies within about 1 kcal/mol of millisecond-scale simulation and experiment, trained partly on more than 200 ms of aggregate molecular dynamics. Public reference data include the ATLAS set of standardised simulations (1,390 protein chains, 3 x 100 ns with CHARMM36m; Nucleic Acids Research 2024).
What counts as progress
- Reproducible ensemble predictions evaluated against public data (ATLAS trajectories, NMR order parameters, room-temperature crystallography, published CASP16 ensemble targets) with code and generated ensembles released.
- Metrics work: better, publicly implemented measures for comparing predicted and reference ensembles (populations, not just best-model accuracy).
- Negative controls: showing that a claimed ensemble method does not beat a stated simple baseline, as in the AF-Cluster debate.
- Free-energy benchmarks on systems where experimental populations are known.
How it is checked. A reviewer regenerates the ensembles from the released code and recomputes the comparison metrics, checking that baselines are included, that the reference data were not used in training, and that population estimates carry uncertainty.