A disordered protein can fold without staying folded

An intrinsically disordered protein is not one structure waiting to be found. It is a population of interconverting conformations, most common, some rare, and none sufficient on its own to describe the molecule. That makes the central computational challenge very different from conventional structure prediction. The problem is not producing one plausible frame. It is sampling enough of the population to discover which frames the protein actually visits.

Our study with Prof. Kresten Lindorff-Larsen and colleagues at the University of Copenhagen, published in Nature Communications, examines that sampling problem directly. We tested On-the-fly Probability Enhanced Sampling in a multithermal ensemble, known as OPES multiT, across peptides and intrinsically disordered proteins from 15 to 71 residues. The method generated atomistic ensembles that were broadly consistent with established simulation approaches while exploring low-population conformations more efficiently.

The ensemble is the molecule

Classical structural biology has been extraordinarily successful at resolving proteins that settle into stable folds. A high-resolution structure can identify an active site, explain a catalytic mechanism, and guide molecular design. But the same representation becomes incomplete when the subject remains in motion.

For an intrinsically disordered protein, a beautiful static structure can imply permanence where none exists. Local helices may form and dissolve. Distant residues may meet briefly and separate. A compact state that matters for recognition may account for only a small fraction of the population. The biological information lies in the distribution of those events.

This is why ensemble modeling must answer two questions at once. Has the simulation explored conformational space broadly enough? And does the resulting population agree with independent measurements? Sampling without experiment can overinterpret numerical states. Experiment without an atomistic ensemble can report averages without revealing the structures underneath them.

A sampling problem, not a rendering problem

Unbiased molecular dynamics can describe motion at atomic resolution, but it can remain trapped in local free-energy minima over accessible simulation times. Replica-exchange methods address this by running many related simulations and exchanging configurations between them. These methods can work well, but their cost and parameter requirements increase for larger, heterogeneous systems.

OPES multiT takes a different route. It uses potential energy as a collective variable so that a single simulation replica can diffuse across a defined temperature range under a converged bias potential. The resulting frames can then be reweighted to recover an ensemble at a target temperature. Production simulations do not exchange configurations, and the method does not require a structure-based reaction coordinate that presupposes the state being sought.

Across the systems in the paper, OPES produced reweighted ensemble averages broadly consistent with REST2 and unbiased simulations. The exact acceleration depended on the protein and observable, so there is no single universal speed-up number. The consistent result was broader exploration and faster decorrelation for multiple structural properties, particularly access to low-population compact or helical states.

In a personal companion essay, I compared this distinction to photography. A perfect still records exquisite detail, while a computational camera combines many frames to capture a moving subject honestly. The analogy is useful, but the scientific test remains stricter: every frame must belong to a physically coherent population, and the population must remain accountable to experiment.

What ACTR revealed

The longest protein examined was ACTR, a 71-residue transcriptional coactivator. ACTR is disordered in isolation and folds when it binds the nuclear coactivator binding domain of CBP. In the bound complex, three ACTR helices contribute to the interaction. The question was whether the unbound ensemble ever visits related structured states on its own.

OPES explored a broader ACTR conformational landscape than REST2 or the previously published unbiased simulations used for comparison. After experimental reweighting, the landscape contained three principal states with progressively more secondary and tertiary structure. The most structured state accounted for approximately 2.6 percent of the ensemble. In that state, helices 1 and 2 were largely formed, helix 3 was partly formed, and long-range contacts increased as secondary structure accumulated.

The state was not one rigid mini-protein. Clustering showed several related conformations with different inter-helical orientations. One cluster resembled a collapsed helical bundle with a defined hydrophobic core. Related compact states were visited repeatedly and reversibly in independent OPES simulations, supporting their interpretation as metastable parts of the apo ensemble rather than one-way simulation artefacts.

That distinction matters. ACTR did not simply sample its final bound structure in the absence of CBP. Simulations initiated from the experimental complex showed that the precise bound arrangement was unstable without the partner. What persisted more readily was secondary structure, while the helices reorganised into alternative compact, non-native tertiary arrangements.

Experiment decides what counts

The simulated population was tested against extensive solution measurements. NMR chemical shifts report local secondary structure. Paramagnetic relaxation enhancement measurements constrain long-range contacts. Residual dipolar couplings and small-angle X-ray scattering provide independent checks on local orientation and global dimensions.

Bayesian maximum-entropy reweighting improved agreement across these observables without forcing one conformation to explain every signal. The OPES ensemble reweighted with chemical shifts and PRE data provided balanced agreement with the restrained measurements and with RDC and SAXS data held out for cross-validation. Importantly, the experiments did not uniquely prove one structural model. They showed that the additional highly helical states sampled by OPES were compatible with the complete set of measurements considered.

This is the standard an ensemble must meet. A rare state becomes scientifically interesting when it is physically sampled, remains part of a coherent population, and does not break agreement with measurements that were not used to construct it.

What the result means

The paper establishes OPES multiT as a practical approach for generating atomistic ensembles of disordered proteins and for exposing functionally relevant conformations that conventional sampling can miss. It also shows why enhanced sampling and experimental reweighting are most powerful together.

The result does not demonstrate exhaustive sampling. It does not establish that the rare ACTR states contain a validated druggable pocket. It does not describe a drug. It provides something earlier and necessary: a physically and experimentally supported hypothesis about transient tertiary structure in an unbound disordered protein.

For molecular discovery, that is a consequential step. A low-population conformation can influence binding, regulation, or folding even when it is invisible to an average structure. By recovering those states and testing the population against experiment, we can decide which fleeting shapes deserve the next experiment.

The deeper account of this work, including the contribution of Prof. Kresten Lindorff-Larsen and the University of Copenhagen, is available on our collaboration page.

Disclosure

This work was sponsored by Peptone Ltd. Michele Invernizzi, Sandro Bottaro, and Kamil Tamiola are employed by Peptone. Kresten Lindorff-Larsen holds stock options in Peptone, is a consultant for the company, and received sponsored research support. Julian O. Streit declared no competing interests. The complete funding and competing-interest statements are available in the published paper.