Equi-mRNA

The genetic code is degenerate: most amino acids are encoded by more than one codon. GCU, GCC, GCA, and GCG all specify alanine: different codons encoding the same amino acid. (Synonymous substitutions are not inert: they can still affect expression and stability, which is part of why the symmetry is worth modelling explicitly.) A model that treats mRNA as ordinary text has to learn that equivalence from data, separately, for every motif it ever appears in.

Equi-mRNA encodes the structure directly instead. Synonymous codons are mapped to cyclic subgroups of SO(2), so codon substitutions that preserve amino-acid identity become rotations the model is equivariant to by construction. Geometry the architecture already knows is geometry it doesn't have to spend parameters memorizing.

25M coding sequences drawn from 56M RefSeq entries (20–512 codons, canonical bases only), pretrained on 32 NVIDIA H100 GPUs. Evaluated across six biologically-driven benchmarks, each split 70/15/15 train/validation/test:

Against baselines several times its size, Equi-mRNA holds its own. All figures below are the paper's reported results on its evaluated tasks, relative to its vanilla baselines: computational benchmarks, not measured therapeutic outcomes.

Score vs. parameter count

One scatter per dataset: model size (log scale) against score. Only models Table 2 reports a parameter count for can appear here; RNA-FM, CALM, and mRNA-FM don't state one.

Ranked by task

Table 1 of the paper: the top-performing configuration from each rotation strategy, on the 1M-sequence ablation subset. Fuzzy θ (soft, learned codon-rotation distributions with a Stiefel basis and the equivariance loss) was carried forward to full-scale training as Equi-mRNA.

Sequences generated by autoregressive codon sampling, scored against true suffixes by Fréchet BioDistance (lower is better) across sampling temperatures. Vanilla GPT-2 stays flat and unstable; both Equi-mRNA variants decline sharply as temperature rises, with the 5M model reaching ≈76 at T=1.0, a ≈33× gap from vanilla at full training scale. The paper's own headlined ≈4.3× gain comes from the equivariant prior alone, in the controlled 1M-sequence ablation.

One sign the representation tracks real biology, rather than a benchmark artifact, is what falls out of it unsupervised, fine-tuned on the human coding transcriptome (GRCh38). The learned codon rotations correlate with GC-content bias (Pearson r = 0.98, R² = 0.97, p < 10⁻¹¹) and with tRNA gene-copy abundance (Spearman ρ = −0.69, p < 10⁻⁶): two independent, well-characterized biological signals the model was never directly trained to predict. These associations are evidence of biological interpretability; they do not establish therapeutic validation.

Mehdi Yazdani-Jahromi, Ali Khodabandeh Yalabadi, Ozlem Ozmen Garibay, University of Central Florida.

Mehdi Yazdani-Jahromi is a DeepBio Scientific co-founder. All three authors list the University of Central Florida on the publication; this note does not make a separate ownership or licence claim.

Peer-reviewed at NeurIPS 2025 (39th Conference on Neural Information Processing Systems). The full paper (architecture, training data, and evaluation) is on arXiv.

The representation behind four programs in the co-scientist:

• Expression prediction
• Stability assessment
• mRNA generation
• Therapeutics design