Reading an MFE RNA Structure, and Where It Is Wrong
6 min read · Updated October 30, 2026
A folding program returns one structure and a number, and both are easy to over-read. The structure is one member of an ensemble; the number is a free energy at a single temperature under one parameter set.
The quickest way to calibrate how much to trust either is to fold a molecule whose structure is already known.
Dot-bracket, in one paragraph
A dot-bracket string is the same length as the sequence. A dot is an unpaired base; a matched pair of brackets is a base pair, with the opening bracket at the 5' partner. Brackets nest, so ((((....)))) is a four-base-pair stem closing a four-base loop — which is what GGGGAAAACCCC folds to, at −5.4 kcal/mol.
Nesting is also the format's limit: a pseudoknot, where two stems interleave rather than nest, cannot be written in plain dot-bracket and is not predicted by this class of algorithm at all.
A molecule whose answer is known
Yeast tRNA-Phe is 76 nucleotides and its cloverleaf has been known for fifty years: an acceptor stem pairing bases 1–7 with 66–72, a D-stem at 10–13 with 22–25, an anticodon stem at 27–31 with 39–43 holding a seven-base loop whose middle three bases are the anticodon, and a T-stem at 49–53 with 61–65.
Folded at 37 °C on the Turner 2004 parameters, it comes back at −32.7 kcal/mol with 24 base pairs. The acceptor stem is exactly right. The D-stem is exactly right. The T-stem is right to within a base pair.
The anticodon stem is not. The prediction extends it past where it stops, pairing 32 with 44, 33 with 43, 34 with 42 and 35 with 40 — a stem with base 41 bulged out of it, which closes the anticodon loop and base-pairs two of the three anticodon bases, the part of a tRNA that has to be single-stranded to do its job.
What that one example teaches
The fold got the global architecture right and a functionally critical local detail wrong, which is the characteristic failure of minimum-free-energy prediction rather than an unlucky case. Benchmarks put single-sequence accuracy around 40–70% of base pairs recovered, and it falls as sequences get longer.
It also shows what the energy is for. −32.7 kcal/mol is the free energy of that structure under that model; it is not a confidence, and a structure 0.5 kcal/mol worse may be the one the molecule adopts. A second structure within a fraction of a kcal/mol of the minimum is common.
So the useful questions are comparative: is this region more structured than that one, does this 5' UTR fold more tightly than the variant next to it, does a mutation open a stem. The unusable question is what is the structure.
37 °C is in the model, not in your protocol
The energies are measured at 37 °C and the fold is computed there. A structure predicted at 37 °C is not the structure at your annealing temperature, at 4 °C on the bench, or at 70 °C in a reverse-transcription step — and nothing about the output says which temperature you had in mind.
Nor is it a two-strand prediction. A fold is one molecule on its own; an oligo pairing with its template or with another oligo is a co-folding question with a different tool.
Practical rules
- Compare candidates rather than trusting one fold. Three 5' UTR variants folded the same way are a usable ranking; one fold is an illustration.
- Treat a long, very stable stem as real and a marginal one as noise: the accuracy is concentrated in the strong features.
- Read the engine and the parameter set. The same sequence folded under different parameter versions gives different energies, and they are not interchangeable between implementations.
- If a pseudoknot matters to your molecule, this class of prediction cannot see it, and the absence of one in the output means nothing.
- Fold the RNA you actually have. A DNA sequence folded as RNA is an approximation of a transcript, not of the DNA.
Frequently asked questions
What does the MFE value mean?
The free energy of the predicted structure under the model's nearest-neighbour parameters at 37 °C, in kcal/mol — more negative is more stable. It is a property of that one structure, not a confidence in it, and it is not comparable between molecules of different lengths, since a longer sequence has more pairs to contribute.
How accurate is an MFE RNA structure prediction?
Benchmarks put single-sequence accuracy at roughly 40–70% of base pairs recovered, falling as sequences get longer. Folded at 37 °C, yeast tRNA-Phe comes back with its acceptor, D- and T-stems correct and its anticodon stem over-extended so that the anticodon is paired — architecture right, a functionally critical detail wrong.
What is dot-bracket notation?
A string the same length as the sequence, where a dot is an unpaired base and a matched pair of brackets is a base pair. Brackets nest, which is why a pseudoknot cannot be written in it — and why this class of algorithm does not predict one.
Can I use a predicted structure at my annealing temperature?
No. The parameters are measured at 37 °C and the fold is computed there, so the structure says nothing about 55 °C or 4 °C. Temperature-dependent folding and two-strand hybridisation are separate calculations.
Related tools
Fold an RNA to its minimum-free-energy secondary structure — dot-bracket, MFE, base-pair list and an exportable arc diagram, with nothing to install.
Find every stretch of a transcript that can pair over the Shine-Dalgarno sequence or the start codon, with the length and melting temperature of each duplex — the mechanism behind every translational switch, and the first thing to check when a construct is silent.
Nearest-neighbor Tm, ΔG, hairpins and primer dimers for any oligo or primer pair.
Predict the translation initiation rate at every start codon, and design a 5' UTR to hit a target expression level.
See repeats, duplications and inversions that a single best alignment hides.