Learned mRNA coding-sequence designers are Pareto-suboptimal on their own objective

Numaro Autoresearch Team

Plain language

What this result means

This matters because learned sequence designers are often evaluated on deterministic proxy objectives before wet-lab validation. If a simple deterministic baseline dominates a learned method on its own proxy, the learned method is leaving objective value on the table. The result stays inside that proxy boundary.

The comparison uses translation correctness, MFE/EFE, CAI, uridine fraction, and forbidden motifs as deterministic scores.
The strongest warning is also in the result: better proxy values do not imply better expression in cells.
LinearDesign is treated as an honest wall. It is exact on its own MFE/CAI objective and is not dominated.

Visual notes

How to read the result

Bar chart of mRNA design Pareto domination counts for CodonRL, GEMORNA, EnsembleDesign, and LinearDesign. — **Domination counts**CodonRL and GEMORNA are dominated on their stated proxy axes; LinearDesign is not, because it is exact on its own MFE/CAI objective.

Bar chart of average proxy-objective improvements on dominated GEMORNA proteins. — **Average movement**On dominated GEMORNA rows, the proxy movement is not one-dimensional: MFE, CAI, uridine, and motifs all move in the allowed direction on average.

Result table

A deterministic baseline dominates CodonRL on 54/54 proteins and GEMORNA on 35/54 under the stated proxy objective.

Cell	Baseline	Numaro	Delta	Note
CodonRL	released soup55	54/54 dominated	all usable proteins	MFE, CAI, uridine
GEMORNA	repo sequences	35/54 dominated	proxy improvement	4-axis objective
EnsembleDesign	EFE specialist	1 dominated	mostly ties	honest wall
LinearDesign	exact MFE/CAI	0 dominated	not beaten	exact wall
Verifier	recorded dominations	90/90 pass	confirmed	translation and scores rechecked

Method

How it was found

The campaign combined deterministic seeds from exact or simple objectives with a ViennaRNA-in-the-loop local polish, accepting only candidates that re-scored as true Pareto improvements.

Reproduced the scoring conventions used by the compared tools.
Generated seeds from maximum-CAI, minimum-uridine, and LinearDesign-guided sequences.
Polished candidates while rechecking translation and all objective values.
Separated proxy-objective statements from biological-expression claims.

Verification

How it was checked

verify.py checks translation to the target protein, recomputes all deterministic scores, and rechecks each Pareto-domination claim from raw FASTA sequences.

Scope

What is not being claimed

This is a computational proxy result only. It is not an in-cell expression claim, not a safety claim, and not a claim that lower MFE is always biologically better.

References

Baseline sources

Citation

How to cite

Numaro Autoresearch Team. "Learned mRNA coding-sequence designers are Pareto-suboptimal on their own objective." Numaro Research Report NUMARO-2026-014, 2026.

@techreport{numaro2026MrnaCodonStructure,
  title = {Learned mRNA coding-sequence designers are Pareto-suboptimal on their own objective},
  author = {Numaro Autoresearch Team},
  institution = {Numaro},
  number = {NUMARO-2026-014},
  year = {2026},
  url = {https://numaro.tech/research/mrna-codon-structure-design-2026/}
}