Compare codon strategies for a coding sequence
Ask: "Using rnafold, recode this coding sequence with each codon strategy and check each version for sequence liabilities." The example is the 720-nt avGFP coding sequence.
The assistant calls optimize_codons once per strategy, then scan_vector_liabilities on
each result. Every recoding is synonymous: the protein is unchanged, and the tool asserts
the translation is identical before returning. Real output (examples/mrna_strategies.json):
| Strategy | Codons changed | GC fraction | U fraction | G-quadruplex motifs |
|---|---|---|---|---|
| native | 0 | 0.392 | 0.265 | 0 |
| uridine_min | 127 | 0.403 | 0.132 | 0 |
| gc_max | 169 | 0.628 | 0.132 | 1 |
| gc_min | 126 | 0.271 | 0.264 | 0 |
| usage_max | not run |
Two things to notice:
gc_maxcreates a problem the others do not have. Raising GC to 0.63 produced a G-quadruplex motif at positions 162-181 on the reverse strand,GGGTGGGCCAGGGGACGGGG: four tracts of guanines that can fold into a structure that stalls polymerases and ribosomes. Recoding for one property can introduce another, and the scan is how you find out. It lists motifs and windowed energy discontinuities; it does not say what to do about them.usage_maxrefuses, and says why: it needs a codon usage table for your organism, none was supplied, and none ships with the tool. A strategy that cannot run says so instead of guessing a table.
Going further: how defined is the RNA ensemble?
The design_mrna_panel recipe in the assistant's prompt list does this as one pipeline:
generate the strategies, fold each on a CPU, score how defined each ensemble is, and cut. A
filter threshold is always yours; the assistant is told not to suggest one. Each number
comes with its control (the same measurement on shuffled sequences) and a note when your
threshold lies inside the replicate range.
Occupancy percentages from rna_ensemble are a report, not a score. For a long or weakly
structured RNA the sampled structures are nearly all different, and cluster percentages
then describe the clustering and not the RNA.
Tools used
optimize_codons, scan_vector_liabilities (free); for the pipeline version,
get_pipeline_schema, validate_pipeline, estimate_pipeline, submit_pipeline.
Last changed 2026-11-17