RNA folding kinetics and energy landscapes
Why an RNA molecule does not simply adopt its lowest-energy structure, how energy landscapes are summarised by barrier trees, how folding is computed on a landscape that changes during transcription, and how fast a structure is reached.
An RNA molecule does not fall into its lowest-energy structure. It moves through a space of alternative structures, one base pair at a time, and it can get stuck on the way. In the cell it starts to fold while it is still being made, so the first part of the chain has already chosen a structure before the rest exists. For a riboswitch, for a designed switch and for many regulatory RNAs this is not a detail. The structure that matters is the one reached in time, not the one that is most stable.
This page collects the work on this site on that question, from the formal description of an energy landscape in 2002 to learned folding time distributions in 2025.
The landscape and its barrier tree
An energy landscape is the set of all secondary structures of a sequence with their free energies, connected by elementary moves. Its local minima are the structures a molecule can rest in, the saddle points between them set the barriers, and a barrier tree summarises this in one diagram: leaves are the minima, internal nodes the lowest saddle connecting two basins, branch lengths the barrier heights.
Barrier trees of degenerate landscapes (Zeitschrift für Physikalische Chemie 2002) makes this precise for the degenerate case, in which many structures share the same energy, and implements it in the barriers program. The input is deliberately general, i.e. a finite set of states, a neighbourhood relation and an energy for each state, which is why the same tool applies to lattice proteins and spin systems as well as to RNA.
The point of the tree is what comes after it. Once the landscape is partitioned into basins connected by barriers, the basins can be treated as macrostates and the folding process as a Markov process between them.
Folding as motion between basins
Efficient computation of RNA folding dynamics (Journal of Physics A 2004) does exactly that. Instead of simulating many stochastic trajectories through the full structure space, it integrates the flow of population between barrier-separated basins. Bistable molecules and switches become computable on timescales that direct simulation does not reach, and the results agree with the stochastic simulations where both can be run.
Getting the landscape in the first place is its own problem, since enumerating all structures is impossible beyond small sequences. Exploring the lower part of discrete polymer model energy landscapes (Europhysics Letters 2006) floods the landscape upward from the optimal and near-optimal structures, so that only the part relevant for the basin structure is expanded. The examples are lattice proteins, the algorithm is not specific to them, and Memory-efficient exploration of RNA energy landscapes (Bioinformatics 2014) brings the approach back to RNA with a local variant that keeps the macrostate model exact while fitting into memory. That paper also compares the exact transition model with two barrier-based approximations and shows where the approximations mislead.
A related question is how to look at all this. Visualization of lattice-based protein folding simulations (IV 2006) treats trajectories and transitions as something to be inspected rather than summarised, which matters once the state space is too large to read as a table.
Folding while the chain grows
During transcription the landscape itself changes: every added nucleotide opens structures that did not exist a moment earlier. BarMap and RNA folding on dynamic energy landscapes (RNA 2010) handles this by computing a barrier tree for each transcript length and mapping the macrostates of one onto those of the next, so that the population can be carried forward rather than recomputed. The same construction covers temperature changes and refolding after cleavage.
For long molecules even that is expensive. Folding kinetics of large RNAs (Journal of Molecular Biology 2008) takes the heuristic route with Kinwalker, which assembles a folding trajectory from locally optimal substructures and interleaves folding with transcription, up to roughly 1500 nucleotides. It reproduces experimentally studied cases such as the delayed cloverleaf formation in bacteriophage RNA and the Hok system.
Riboswitches: where timing is the mechanism
A riboswitch decides between two structures depending on whether a small molecule is bound, and for many of them the decision is made during transcription. Co-transcriptional folding and metastable states in riboswitch function (Journal of the American Chemical Society 2017) resolves every transcription intermediate of the 2′dG riboswitch from Mesoplasma florum by NMR at single-nucleotide resolution. Ligand responsiveness depends on the transcript length: certain intermediates expose a binding-competent arrangement, slightly longer ones already offer competing folds. The switch works within a narrow kinetic window.
Co-transcriptional riboswitch modeling with ViennaRNA (Methods 2018) models the same system computationally by adding ligand binding to the mapped-landscape framework, with binding treated as all-or-none for a defined competent structure and a concentration-dependent rate. The model reproduces the mechanism: an early competent state captures the ligand and thereby steers the molecule away from the conformation that would otherwise take over.
The design side follows directly. In silico design of ligand-triggered RNA switches (Methods 2018) makes the design objective explicit and includes folding kinetics in the evaluation, because a candidate that satisfies the structural constraints can still fail if its switching path is slow or passes through the wrong intermediate.
How long does it take?
A compact way to state folding kinetics is the distribution of first passage times, i.e. how long a molecule needs to reach a target structure for the first time. Obtaining that distribution requires many simulation runs. KinPFN (ICLR 2025) approximates the whole distribution from a handful of example folding times with a prior-data fitted network trained on synthetic multi-modal priors. It does not replace the simulator. It sits on top of it, reports the cumulative distribution in one forward pass, and saves most of the simulation effort, which makes it possible to compare many candidates rather than a few. The workshop paper states the same idea in shorter form.
Where this connects
Folding kinetics is the reason why a predicted equilibrium structure is not the end of the story. How such a prediction is computed, and how far it can be trusted, is the subject of the tutorials on RNA structure prediction with ViennaRNA; probing data, which constrain the structures a molecule actually populates, are described in Predicting RNA structures from sequence and probing data (Methods 2016). The terms used here are defined in the glossary, e.g. metastable state, co-transcriptional folding and BarMap. The posts are also collected on the topic pages RNA folding kinetics, energy landscapes and co-transcriptional RNA folding.