What rnafold will not do

These are choices, not missing features. They are why a result reads the way it does.

It will not give you one answer

A fold runs several independent engines and reports where they agree and where they disagree. It does not average them into a consensus structure, name a best engine, or combine anything into a score. Engines are listed alphabetically. If you see a distance between two engines, it is a distance to a named reference engine, not a measure of which is closer to the truth.

It will not say a target is promising, a variant is good, or a design should go ahead. Those are your judgement, and a number from us would only hide the reasoning.

Agreement is not evidence

Engines can share a bias, and they can all be recalling a structure they were trained on. Every fold reports the earliest close relative in the Protein Data Bank and which engines' training cutoffs it predates.

Example, from a real fold of yeast tRNA-Phe: the closest PDB match is a 2025 entry, but the earliest relative at 100% identity is 6TNA, deposited in 1978, before every engine's training cutoff. Agreement between those engines may be recall of a tRNA they have seen, not a prediction about this molecule. The result says so next to the numbers.

It checks what you tell it

If you say a molecule is a G-quadruplex, the result checks the predicted coordinates for G-tetrads and tells you if there are none. A declared fold is checked, not trusted.

It refuses rather than shrinks

  • A mutation whose wild-type letter does not match your sequence is refused, and the whole comparison with it. Running a smaller set would give you a different comparison from the one you asked for.
  • A sequence too long for two engines is not folded by one: one engine's opinion is not a comparison. It is declined before anything is spent.
  • A request your allowance cannot finish is refused up front, with the numbers.
  • A time estimate it cannot back is not made up: for lengths nobody has measured it says "no time can be quoted".

It says what it did not model

Secondary-structure calculations use one nearest-neighbour energy model at 37 °C on the unmodified sequence. No tertiary contacts, no magnesium, no base modifications. A predicted structure is a prediction, not a measurement, and the tools say so in caveats on every result that needs it.

What this means in practice

When your assistant answers, ask it to show the disagreement, the earliest relative and the caveats. A good answer contains them. If it gives you one confident structure and no caveats, ask it to read the tool's caveats again.

Last changed 2026-11-17