Does a mutation change the structure?
Two ways to ask, at two costs. Start with the free one.
The free, instant way: secondary structure on a CPU
Ask: "Using rnafold, does G10A change the predicted secondary structure of this tRNA? How sure are you?" with the 76-nt sequence from the quickstart.
The assistant calls make_point_variants, then fold_rna and rna_ensemble on the wild
type and the variant. Real output (examples/variant_cpu.json):
| Wild type | G10A | |
|---|---|---|
| Minimum free energy (kcal/mol) | -22.4 | -22.1 |
| Ensemble free energy (kcal/mol) | -24.13 | -23.93 |
wild type (((((((..((((........)))).(((((.......))))).....(((((.......))))))))))))....
G10A (((((((....((((((.((......)).)))))).((((....))))(((((.......))))))))))))....
The minimum-free-energy structure changes a lot (the D stem closed by position 10 is gone) while the energies hardly move: 0.3 kcal/mol. That mismatch is the finding. The model says the wild type already sits among several folds, so a point change can move the printed structure without moving the energy much.
In 1,000 sampled structures per sequence (rna_ensemble) the largest cluster holds 35.4% of
the wild type's samples and 45.5% of G10A's. Read those as a report about one energy model:
- The tool's own caveat: "Occupancies are frequencies in a sample of the Boltzmann ensemble under one energy model. They are not measurements of the molecule, and they are not a score or a ranking."
- Check
report.calculated.stability, the range of each occupancy under resampling. A decomposition that moves between resamples is telling you the clusters are not firm. - The model has no tertiary contacts, magnesium or modified bases, which a real tRNA relies on. The calculation says G10A weakens the D stem; it cannot say whether the real tRNA still folds into its L-shape.
The GPU way: several independent engines
Ask: "Using rnafold, compare G10A and A5G against the wild type." The assistant calls
compare_variants, which:
- folds the wild type and each variant on every engine that fits (up to 7 variants);
- returns straight away with the run id and an estimate (the assistant tells you to wait);
- costs one run per fold: 3 runs for 2 variants. A request your allowance cannot finish is refused before anything starts.
wait_for_job then returns, for each variant and each engine, which residues changed
base-pairing state against the wild type: each engine against itself. It also lists the
positions that changed in every engine, as a statement about the predictions.
A real run
The first 28 nucleotides of this tRNA (the part that holds the D stem), wild type and G10A, on six engines
(examples/compare_variants.json). The estimate was 6.1 minutes for the two folds; the comparison came back
when both had finished. For each engine, the number of residues whose base-pairing state differs between the
wild type and G10A (of 28):
| Engine | Residues that changed |
|---|---|
| boltz2 | 6 |
| chai1 | 15 |
| opendde | 19 |
| openfold3_apo | 17 |
| rf3 | 8 |
| rhofold | 4 |
How to read it:
- Every engine moves, but by very different amounts (4 to 19 residues), and no residue changes in all six engines. The engines do not agree on what G10A does. Which one is nearer the truth is not something this comparison can say.
- The engines also disagree with each other far more on the variant (27 of 28 residues where the six engines differ) than on the wild type (10 of 28). That is a reason for caution about any single prediction of this variant. It is not, by itself, a statement about the molecule.
- It sits alongside the free 2D result above, which says G10A loses the D stem. The 2D model and several 3D engines are different instruments, and here they do not tell one story.
The comparison names no consensus and does not rank the variants. If one engine moves and the others barely do, that is the interesting part: ask for the engines separately.
The run also leaves a group named "compare ..." holding the sequences, so it can be repeated. Asking for the same
comparison again the same day returns the first run and charges nothing; pass again=true to deliberately run it
again.
If a mutation does not match your sequence ("C20U: position 20 is G, not C"), the whole comparison is refused. That is the numbering telling you something: check it, usually a protein-style numbering against a linear one.
Tools used
make_point_variants, fold_rna, rna_ensemble (free); compare_variants,
wait_for_job (GPU). See the tool reference.
Last changed 2026-11-17