Fold a sequence and get the structure

Ask: "Using rnafold, fold this and give me the structure file." Sequence: a 28-nucleotide fragment of yeast tRNA-Phe, GCGGAUUUAGCUCAGUUGGGAGAGCGCC.

The assistant calls study_sequence, which starts the fold at once and returns:

  • the job id and an estimate: for this length, about 3 minutes on six engines;
  • a free 2D preview from the CPU, so there is something to look at while it runs.

Then wait_for_job returns a short summary when it is done (call it again if it says still running). A real run of this fold (examples/fold_summary.json) took 169.8 seconds.

What the summary says

What ran. Six engines, alphabetically: boltz2, chai1, opendde, openfold3_apo, rf3, rhofold. Each one reports how many structures it produced, and its runtime.

Where they agree. 18 of the 28 residues have the same base-pairing state in all six engines; 10 differ, in six regions: 1-2, 11, 14-15, 20-21, 24 and 26-27. Four engines (boltz2, chai1, opendde, rf3) give the same pairing state at every residue: 10-13 and 22-25 paired, the rest unpaired. rhofold differs at six residues (11, 24, and 14, 15, 20, 21); openfold3_apo differs at four (1, 2, 26, 27), where it also pairs the two ends.

How far apart they sit. Superposed on boltz2, the reference engine for this fold:

EngineRMSD, all residues (A)
chai12.09
opendde1.53
openfold3_apo12.60
rf37.59
rhofold6.56

These are distances to boltz2, not quality scores. Two engines can agree on the pairing and still place the residues differently, which is what rf3 does here: its pairing matches boltz2's at every residue, yet it sits 7.6 A from boltz2.

Whether it was in training. The closest PDB match is a 2025 entry at 100% identity. The date that matters is the earliest close relative: 6TNA, deposited 1978, before the training cutoff of boltz2 (2023-06-01), chai1 (2021-01-12) and openfold3 (2025-06-30). The agreement above may be recall of a tRNA the engines have seen, not a prediction for this fragment.

The platform deliberately names no consensus structure and no best engine. You have six candidate structures.

Getting the coordinates

Ask for one engine: "Give me the boltz2 PDB." The assistant calls get_structure(job_id, "boltz2"), which returns the file as text.

  • All six engines use chain A.
  • A 28-nt structure is about 49 KB of text, roughly 20,000 tokens of your assistant's context. For anything longer, ask for trace_only: one atom per residue (C1' for nucleotides, CA for protein), about twenty times smaller.
  • A file too large to return is not truncated: a partial PDB would silently drop residues. The tool says so and offers trace_only or a higher limit.
  • Most chat assistants cannot save a file for you. Ask it to show the text, or to put it in an artifact or a code block you can copy.

Longer sequences

Fold time is not linear in length: about 145 seconds up to ~40 nt, about 9 minutes at 200 nt, about 19 minutes at 291 nt. Beyond that nothing has been measured, so study_sequence still runs it and says "no time can be quoted". estimate_fold is free and tells you how many engines fit.

Tools used

study_sequence, wait_for_job, fold_summary, get_structure, estimate_fold.

Last changed 2026-11-17