Computing an RNA secondary structure takes a second. Deciding what to do with the result is where a project's money goes. Seven ways a prediction gets used for something it did not say, what each one costs a project in months and experiments at Austrian public funding rates, and the check that would have caught it. Written for the person who signs the timesheet, with the central number computed on the first 400 nucleotides of a real viral genome.
Computing a secondary structure is the cheapest step in RNA biology. What is done with the result is among the most consequential. RNAfold on a few hundred nucleotides finishes in well under a second on a laptop, and the output looks finished too: one line of brackets, one energy. Nothing about it says that this is the most probable of many structures the molecule adopts, not the structure it has.
The money is spent afterwards, when the line of brackets becomes a mutant series, a synthesis order, or a panel in a grant application. This is a list of the ways I have seen that step go wrong, in courses, in the issue tracker and in manuscripts I have reviewed, with the cost of each worked out at current rates. None of them is a flaw in the method. Each is the method answering a question that was not the one being asked.
Three unit prices
The arithmetic below uses three numbers, stated once here so the multiplication is yours to check. The salary figure is an Austrian public funding rate, so the sums are what a project here would pay. Rates elsewhere in Europe differ, the order of magnitude does not.
- A doctoral month. The FWF standard personnel cost rate for a doctoral candidate at 30 hours a week is EUR 50,220 a year for applications from January 2026, employer contributions included. Six months of a doctoral student cost half of that, EUR 25,110.
- A probing experiment. SHAPE-MaP or DMS-MaPseq on one RNA costs a few thousand euros in reagents and sequencing and two to four weeks of somebody's time, before any analysis.
- A synthesis round. A handful of oligonucleotides is a few hundred euros. A series of modified mRNA constructs for an expression assay runs to four or five figures.
Delay has a price of its own that none of these capture. A manuscript revision is three to six months. A grant deadline missed is a year.
1. The minimum free energy structure, taken as the structure
What it looks like. A figure with one structure in it. A mutant designed to disrupt "the helix at positions 120 to 135". A sentence in a manuscript saying the 5' UTR forms a stem-loop, singular.
Why the model produces it. The MFE structure is the single most stable one, and for an RNA of a few hundred nucleotides that is a weak distinction. Here are the first 400 nucleotides of the SARS-CoV-2 genome (NC_045512.2), the 5' UTR and the start of ORF1ab, one of the best-studied RNA structures of the decade:
RNAfold -p --noPS < sars2_1-400.fa
AUUAAAGGUUUAUACCUUCCCAGGUAACAAACCAACCAACUUUCGAUCUCUUGUAGAUCUGUUCUCUAAACGAACUUUAAAAUCUGUGUGGCUGUCACUCGGCUGCAUGCUUAGUGCACUCACGCAGUAUAAUUAAUAACUAAUUACUGUCGUUGACAGGACACGAGUAACUCGUCUAUCUUCUGCAGGCUGCUUACGGUUUCGUCCGUGUUGCAGCCGAUCAUCAGCACAUCUAGGUUUCGUCCGGGUGUGACCGAAAGGUAAGAUGGAGAGCCUUGUCCCUGGUUUCAACGAGAAAACACACGUCCAACUCAGUUUGCCUGUUUUACAGGUUCGCGACGUGCUCGUACGUGGCUUUGGAGACUCCGUGGAGGAGGUCUUAUCAGAGGCACGUCAACAU
......(((((.(((((....)))))..))))).......(((((((((.....))))).((((.......))))........((((((((.((.((((.(((.....))).)))))).))))))))..((((((.....))))))...(((((((((((..(((((...(((.((((((((.....((((((.(((((......)))))..)))))).........(((((((.((......)))))))))(((....)))))))))))))).))))).))))...)))))))))))...((((((.......((..((((((...))))))..))))))))...(.(((((.(((((((((((((.....))).)))))..))))).))))))..... (-132.70)
......(((((.(((((....)))))..))))).........,.(((((.....))))).((((.......))))........((((((((.((.((((.(((.....))).)))))).))))))))..{(((((.....)))))}.{.(((((((((((..(((((...(((.((((((((.,,..((((((.(((((......)))))..))))))......,..(((((((.((......)))))))))(((....)))))))))))))).))))).))))...))))))),).....((((((.......((..((((((...))))))..))))))))...|}(((((.(((((((((((((.....)))}})))}.,))))).)))))...... [-138.14]
......(((((.(((((....)))))..)))))...........(((((.....))))).((((.......))))........((((((((.((.((((.(((.....))).)))))).))))))))..((((((.....))))))...(((((((((((..(((((...(((.((((((((.....((((((.(((((......)))))..)))))).........(((((((.((......)))))))))(((....)))))))))))))).))))).))))...))))))).......((((((.......((..((((((...))))))..)))))))).....(((((.(((((((((((((.....))).)))))..))))).)))))...... {-130.40 d=21.45}
frequency of mfe structure in ensemble 0.000147239; ensemble diversity 35.27RNAsubopt -e 1 < sars2_1-400.fa | tail -n +3 | wc -l # 104
The MFE structure has a free energy of −132.7 kcal/mol and a probability of 0.015 percent in the ensemble it was drawn from, i.e. about one molecule in seven thousand. There are 104 structures within 1 kcal/mol of it, 1,641 within 2 and more than 17,000 within 3. The ensemble diversity, i.e. the expected number of base pairs by which two structures drawn from the population differ, is 35.
This does not mean the prediction is bad, and the literature lets us check. Probing and conservation studies agree on five stem-loops in this region, SL1 to SL5, with SL5 spanning positions 150 to 294 and ending with a three-way junction (Rangan et al. 2020, Manfredonia et al. 2020, Huston et al. 2021). They sit one after the other on an otherwise unpaired backbone. All five are in the MFE structure above (the third arm of the SL5 junction, SL5c, is the one piece the model does not place), and the base pairs that form them have mean probabilities of 0.90 (SL1), 0.95 (SL2), 0.81 (SL3), 0.97 (SL4) and 0.93 (SL5) in the ensemble. The helices are well supported.
The MFE structure also contains one thing the literature does not have. Nucleotides 41 to 44 pair with 295 to 298, four base pairs that close SL2 to SL5 into a multiloop. Those four pairs have probabilities between 0.23 and 0.30. The centroid structure, the third line of the output above, leaves them out, and so does the maximum expected accuracy structure (RNAfold --MEA). The one element of the prediction that the probing data reject is the one element the dot plot marks as weak, and the line of brackets gives no hint of it. The question to ask is therefore about a helix, never about a structure.
What it costs. A mutant series designed against a helix with a base pair probability of 0.3 tests a helix that is absent in seven out of ten molecules. Design, cloning, the assay and the months before anyone notices that nothing is consistent come to about a quarter of a doctoral year, i.e. about EUR 12,500, plus reagents.
What prevents it. RNAfold -p, and the dot plot read before anything is ordered. The RNAfold tutorial in this series exists for this.
2. A fragment folded out of its context
What it looks like. A 3' UTR cut out of the transcript and folded alone. A target site for an siRNA or an antisense oligonucleotide judged on the fold of the 50 nucleotides around it.
Why the model produces it. Base pairs do not respect the boundaries you draw. A nucleotide near the edge of a fragment has lost half of its possible partners, so the prediction for that region describes a molecule that does not exist.
The same RNA shows it. The annotated 5' UTR ends at position 265, and folding exactly those 265 nucleotides, as a reasonable person would, gives a structure in which SL2 is missing (its nucleotides pair elsewhere, the best pair inside the element has a probability of 0.20) and SL5 is cut through the middle, because SL5 runs to position 294 and the boundary of the fragment is inside it. With 400 nucleotides both are back, SL2 with pair probabilities of 0.95. Nothing about the sequence changed. The boundary moved a conserved hairpin out of the prediction.
The accessibility question is more sensitive still, and here the mistake deserves to be shown as a mistake. The first of these two commands is the one a person makes who has the UTR in a file and wants to know how accessible an eight-nucleotide window is:
RNAplfold -W 240 -L 160 -u 8 < sars2_5utr.fa # the mistake: the UTR by itself RNAplfold -W 240 -L 160 -u 8 < sars2_1-600.fa # the UTR in its context
The sequence in the first file is 265 nucleotides and the window is 240, so there is no position in that file that lies more than one window from a cut. The RNAplfold tutorial measures how far the border effect reaches and ends with the rule to read no accessibility from the first or last window of whatever was handed to the program. By that rule the first command yields nothing that may be read at all. It still prints numbers, and they look like every other number the program prints:
| position | UTR alone | UTR within 600 nt |
|---|---|---|
| 215 | 5.8 × 10-8 | 4.7 × 10-4 |
| 230 | 0.070 | 0.011 |
| 265 | 0.035 | 0.00097 |
At the 3' end of the fragment the number is 36 times too high. Fifty nucleotides upstream it is four orders of magnitude too low. Near a boundary the result is not an approximation of the right number, it has no relation to it, and nothing in the output says so.
What it costs. A synthesis round aimed at a site chosen on the wrong number, and the weeks of testing that follow. For oligonucleotides the reagents are cheap and the time is not. For a series of modified mRNAs neither is.
What prevents it. Fold the whole molecule, or compute the local quantity with the fragment inside its context, which is what RNAplfold is for. The RNAplfold tutorial works through the choice of window and span.
3. Modified nucleotides folded as unmodified ones
What it looks like. A therapeutic mRNA in which every uridine has been replaced by N1-methylpseudouridine, folded with the default parameters, i.e. as if it were made of uridine.
Why the model produces it. The default energy parameters describe the four standard nucleotides. Modified bases pair and stack differently. Before ViennaRNA 2.8.0 the parameter files for them had to be found and loaded separately, so most people did not.
What it costs. For a therapeutics group this is the most expensive item on the list. Construct design, a synthesis series and an expression assay, all resting on a calculation for the unmodified sequence, while the molecule that leaves the synthesiser carries the modified base at every uridine position. A single synthesis series of modified mRNAs can reach five figures before the first measurement.
What prevents it. Since ViennaRNA 2.8.0 the parameter files for pseudouridine and N1-methylpseudouridine ship inside the library, so the fix is one option on the command line. The 2.8.0 release note has the details.
4. Probing data that constrained nothing
What it looks like. A SHAPE-MaP dataset, generated at real expense, shown as a bar chart beside the predicted structure and never put into the prediction. Or put in with the default method, no thought about normalisation, and read afterwards as a per-nucleotide verdict of paired or unpaired.
Why the model produces it. A reactivity is a continuous, noisy measure of nucleotide flexibility. It is not a pairing state. It enters the model as a pseudo energy, and the three methods for doing so give different structures from the same data, which is why the method and its parameters have to be chosen and then reported.
What it costs. The experiment itself, a few thousand euros and several weeks, is the smaller part. The larger part is a figure that looks like validation and is not. A reviewer usually finds this before the authors do.
What prevents it. --shape with a method chosen on purpose, and the constrained prediction reported next to the unconstrained one. The probing-data tutorial covers the three methods and what to report.
5. A structure claim with nothing behind it
What it looks like. One dot-bracket string in a figure legend. No base pair probabilities, no ensemble diversity, no alternative structure, no statement of what was checked.
Why the model produces it. RNAfold prints one structure by default, and one structure is what gets pasted into the figure.
What it costs. Reviewers ask now, more often than not, how well supported the structure is. Answering takes an afternoon if the ensemble was computed at the time and a revision cycle if it was not: three to six months for a manuscript. In a grant application there is no revision cycle, only the next deadline, a year away.
What prevents it. Show the dot plot or the base pair probabilities, report the ensemble diversity, and say in the methods which quantities were examined. This is a few minutes of work when the ensemble was computed in the first place.
6. What the model leaves out, assumed to be in it
What it looks like. A secondary structure reported for an RNA whose fold is decided by something the calculation does not contain. tRNA is the usual example. For yeast tRNA-Phe with default parameters the MFE structure is the cloverleaf, with a probability of six percent. The centroid of the ensemble keeps only the acceptor stem, and at 25 °C a different structure becomes the MFE, with alternative D and anticodon arms. What holds the real molecule in its shape is a network of tertiary contacts and magnesium ions, neither of which is in the model.
Why the model produces it. The nearest neighbour model contains stacking energies, loop penalties and the standard base pairs at 37 °C in 1 M NaCl. It does not contain pseudoknots, divalent ions, proteins, modified bases unless asked, or tertiary contacts. Where those decide the fold, the ensemble says what the sequence would do on its own in a test tube of salt water. Everything beyond that has to come from somewhere else, e.g. from probing data, from a homologue with a solved structure, or from an experiment.
What it costs. Months spent on a question the method cannot answer, and a claim that no better parameter set can rescue. This is the error that most often becomes a correction rather than a revision.
What prevents it. Knowing the list of what is not in the model before starting, and asking whether the RNA at hand is on it. The closing section of the RNAfold tutorial spells the list out.
7. Numbers compared across parameter sets or versions
What it looks like. A free energy computed in 2019 set beside one computed with ViennaRNA 2.8.0. Pseudo energies from a probing-constrained fold compared with unconstrained ones. Results from two versions of the package pooled into one table.
Why the model produces it. Energies are only comparable within one parameter set and one method. ViennaRNA 2.8.0, released on 6 November 2026, corrected the units of the G-quadruplex parameters, three corrections for modified bases, Wuchty-style suboptimal folding for two dangle settings, and the loop decomposition used by RNAeval. A number from before and a number from after are not the same quantity.
What it costs. A resubmission when the numbers do not reproduce, or, if nobody checks, a difference between two versions of the software interpreted as biology and published.
What prevents it. The package version and the parameter file in the methods section of every paper, and a recomputation when upgrading. The 2.8.0 release note lists which results moved.
What the arithmetic says
Each item above costs more than the check that would have caught it, usually by a factor of a hundred or more. The check is a second command, or a second look at output that already exists. Done deliberately, before anything is ordered, that check has a name, an RNA structure review, and it costs a small fraction of any item above. The expensive mistakes in RNA structure work happen when the most probable of many structures is read as the structure the molecule has, and an order form is signed on that reading.
Sources: FWF standard personnel costs, rates for applications from January 2026 · NC_045512.2, SARS-CoV-2 isolate Wuhan-Hu-1, positions 1 to 265, 1 to 400 and 1 to 600 · Rangan R, Zheludev IN, Hagey RJ, Pham EA, Wayment-Steele HK, Glenn JS, Das R. RNA genome conservation and secondary structure in SARS-CoV-2 and SARS-related viruses. RNA 26:937–959 (2020). doi:10.1261/rna.076141.120 · Manfredonia I, Nithin C, Ponce-Salvatierra A, Ghosh P, Wirecki TK, Marinus T, Ogando NS, Snijder EJ, van Hemert MJ, Bujnicki JM, Incarnato D. Genome-wide mapping of SARS-CoV-2 RNA structures identifies therapeutically-relevant elements. Nucleic Acids Res 48:12436–12452 (2020). doi:10.1093/nar/gkaa1053 · Huston NC, Wan H, Strine MS, de Cesaris Araujo Tavares R, Wilen CB, Pyle AM. Comprehensive in vivo secondary structure of the SARS-CoV-2 genome reveals novel regulatory motifs and mechanisms. Mol Cell 81:584–598.e5 (2021). doi:10.1016/j.molcel.2020.12.041 · ViennaRNA 2.8.0, RNAfold, RNAsubopt and RNAplfold man pages · ViennaRNA 2.8.0 release notes.