The genetic code is redundant: most amino acids are spelled by several interchangeable three-letter codons, and which synonym a cell prefers — captured by the statistic called relative synonymous codon usage (RSCU) — changes translation speed, accuracy, and even how the protein folds. A recent line of work turns this redundancy into geometry: Equi-mRNA represents each synonymous codon as a rotation angle on a circle and aggregates angles across a transcript with a Fourier transform. Wireless engineering, meanwhile, has an old and precise result about rotations and errors. In a base station with hundreds of antennas, uncorrelated hardware noise cancels as the array grows — wanted signals add coherently so their power grows like $N^2$, while independent distortions add incoherently and grow only like $N$ — but phase errors, small rotations that accumulate step by step, refuse to cancel: Björnson et al. proved such drift tolerates only $\mathcal{O}(\log N)$ growth before the link collapses.
These two facts have never been examined together, and the intersection hides a concrete, testable weakness. Fourier pooling over codon rotations implicitly assumes the angles stay coherent across hundreds of positions — exactly the assumption that MIMO theory identifies as fragile. We claim that codon-level language models inherit a phase-drift problem: pooling removes uncorrelated per-codon noise but cannot remove angular drift that accumulates along the sequence, producing a length-dependent error floor that more parameters and more training data within the current range will not lift.
Core Claim
Let $\theta_t$ be the rotation angle a codon language model assigns to the codon at position $t$, and let $\Phi_T = \phi_1 + \cdots + \phi_T$ be the accumulated drift, a random walk of small per-position angular errors with per-step variance $\sigma_\phi^2$. Pooling $T$ unit vectors then loses signal exponentially in sequence length:
$$\Big| \mathbb{E}\Big[\tfrac{1}{T}\sum_{t=1}^{T} e^{i\theta_t}\Big] \Big| \;\approx\; e^{-\mathrm{Var}(\Phi_T)/2}, \qquad \mathrm{Var}(\Phi_T) = T\,\sigma_\phi^2.$$
Uncorrelated per-token noise leaves the variance constant, so pooling averages it away; drifting phase makes the pooled signal decay as $e^{-T\sigma_\phi^2/2}$. This is the sequence-level twin of two wireless results. In over-the-air computation, Chen, Xing, and Xu showed that with imperfect channel knowledge a single-antenna receiver hits a hard mean-squared-error floor (approximately 0.08 at channel-error variance 0.2, regardless of transmit power), which only array averaging removes — cutting error by nearly two orders of magnitude from 1 to 100 antennas. And in coarse-quantization uplink, Jacobsson, Durisi, and Coldrey showed that when quantization distortion is uncorrelated across antennas, even 1-bit converters support 16-QAM, and 3-bit converters reach about 90% of infinite-resolution performance for 64-QAM with zero-forcing (versus about 71% for 1-bit). Coarse per-symbol precision is fine when errors are independent; the open question is what happens when they accumulate. Our claim has three parts: drift is measurably present in rotational codon encodings; it, not data scarcity, explains part of the length bottleneck (CodonTranslator, for instance, caps at a 2,048-token context with no length-generalization analysis, and no reported result found on Lacuna for its Dynamic Time Warping similarity scores); and correcting it yields a falsifiable prediction about natural codon bias itself.
Research Program
The program has four parts, each with a clear supplier of technique. First, theory: we port Björnson's error taxonomy (Massive MIMO Systems with Hardware-Constrained Base Stations, with the energy-efficiency motivation of Circuit-Aware Design of Energy-Efficient Massive MIMO Systems) into the sequence domain, separating additive per-codon noise (which pooling suppresses) from accumulated angular drift (which it cannot), and deriving when each dominates as a function of transcript length. Second, measurement: we audit drift in an Equi-mRNA-style model, defining drift operationally — the circular variance of angles assigned to the same codon in matched amino-acid contexts, as a function of position — so the claim is falsifiable even though the angles are learned rather than physical. Third, method: drift-corrected pooling. We pool the Fourier transform over windows and periodically re-anchor the rotation frame, subtracting the estimated accumulated drift. The external phase reference is the organism's RSCU profile, computable directly from a reference corpus with no expression data — the same intrinsic, reference-free philosophy as the Frare index for rare codon pairs (Wu et al.). Window size and drift magnitude are tuned without clean targets using a Stein-type unbiased risk estimator, the technique BEACHES developed for denoising DFT-domain channel vectors in $\mathcal{O}(B \log B)$ operations. Fourth, biology: the drift picture makes a prediction. Equi-mRNA found that its learned angles correlate with tRNA abundance, so the angle plausibly encodes translation timing; timing offsets over a long reading frame accumulate like oscillator phase drift, and codon identity is known to correlate with backbone conformation (codon-specific Ramachandran plots), so drift has structural consequences. If the usable timing "phase budget" grows only logarithmically with length, then across natural genes the strength of codon bias (measured by ENC, following the RSCU/ENC methodology of Chen et al.) should saturate or scale near-logarithmically with ORF length, not linearly.
This differs from the closest existing work in a specific way. Equi-mRNA builds the rotational representation but assumes equivariance holds at arbitrary depth, pretrains only on sequences of 20–512 codons, and never models cumulative drift. AirComp eliminates error floors with antenna arrays but touches no biology. BEACHES supplies risk-estimated denoising in a Fourier domain, but for physical channels. No paper in the Lacuna corpus bridges the two directions at all; we reuse the MIMO error taxonomy as a diagnostic and design principle, not as metaphor.
Baselines
The do-nothing reference is the released Equi-mRNA — a 15-million-parameter model pretrained on 25 million RefSeq mRNAs of 20–512 codons — evaluated unchanged on longer inputs, alongside classical length-agnostic RSCU-vector statistics of the kind Vicente used as lineage fingerprints for viral detection, which give a non-neural floor. It is the strongest competitor on this question because it owns the representation we are auditing: it reports up to roughly 10% accuracy improvements on six downstream property-prediction tasks over standard transformer baselines, 4.3-times-better generation realism under Fréchet BioDistance, and about 28% better functional preservation, all at 15M parameters while outperforming the 50M HELM and 82M CodonBERT. Beating it means: on transcripts longer than 512 codons, the drift-corrected variant must retain the accuracy margin over vanilla transformers that the released model enjoys at short lengths, while the released model's margin measurably erodes. The strongest existing codon generator is CodonTranslator, a 150M-parameter decoder-only transformer trained on 62 million CDS-protein pairs across 2,163 species with held-out unseen species; since no reported result found on Lacuna for its DTW or similarity scores, we will run it ourselves on its own unseen-species and long-gene strata under equal tuning budget. Evaluation runs through mRNABench, whose 59 tasks and homology-aware splits matter because random splits inflate accuracy — its 10M-parameter Orththus+MLM is SOTA or near-SOTA on 6 of 10 datasets while 700 times smaller than the 7B Evo2, though no reported result found on Lacuna for its per-task scores. Finally, the key ablation isolates our ingredient: windowed Fourier pooling without pilot anchoring versus with it, plus a robustness comparison of simple Polar pooling against structure-exploiting attention aggregation under injected noise, mirroring Björnson's finding that MMSE combining is more impairment-sensitive than the simpler maximum-ratio combining. All variants are trained on the same RefSeq data (the 1M-sequence ablation subset for iteration, the full corpus for the final run), at matched parameter counts and matched tuning budget via equal random-search trials.
Why not just…?
Why not just train on longer sequences? CodonTranslator already uses 2,048-token contexts and reports no evidence that this fixed the length problem; and mRNABench documents that annotated mRNA data is sparse and biased toward a few cell lines, so long labeled transcripts are exactly the scarcest resource. But this is a genuine threat, not a strawman: if the failure is data scarcity rather than representational drift, more long-sequence training at equal compute would match our correction. Our length-extrapolation experiment — train at 20–512 codons, evaluate at 512–2048 — is designed to separate the two, and if plain long-context training wins at equal compute, we will report that the drift story does not explain the bottleneck.
Why not just use standard codon encoders without rotations? Equi-mRNA's own results argue against abandoning the rotational encoding: it beats CodonBERT and other standard transformers by up to about 10% accuracy and 4.3x on Fréchet BioDistance at one-third the size. The interesting question is whether that advantage survives length, which is precisely our experimental stratum. The classical RSCU/ENC statistics are length-robust but discard all context, so they serve as a floor, not a competitor.
Why not just apply standard length-extension tricks? RoPE rescaling, sliding-window attention, and plain windowed pooling are the right first baselines and we run all of them. They address position, not semantics: none of them re-reference the rotation frame, which is the specific operation our theory says is needed. The ablation of windowed pooling with versus without RSCU-referenced anchoring is the decisive comparison — if unanchored windowing matches anchored windowing at every length, the drift mechanism adds nothing and the proposal's premise fails.
Proposed Experiments
We first audit drift. Training the Equi-mRNA architecture on the 1M-sequence RefSeq subset, we measure circular variance of assigned angles for matched codons as a function of position, testing random-walk growth (variance linear in position) against saturation. We then verify the pooling-decay equation by injecting synthetic noise — uncorrelated additive versus correlated angular — and checking that prediction error follows the exponential law. Second, length extrapolation: on mRNABench's 59 tasks with homology-aware splits, stratified by transcript length, plus generation realism via Fréchet BioDistance on synthesized transcripts beyond 512 codons, we compare the released model, our anchored variant, RoPE-rescaled and sliding-window variants, and plain windowed pooling. Third, the MRC/MMSE analogue: under injected correlated drift, we test whether attention-based aggregation degrades faster than simple Polar pooling, as Björnson's combining analysis predicts. Fourth, the biological scaling test: across the 2,163-species RefSeq corpus (and independently on organelle genomes, e.g., Astragalus chloroplasts and Pentatomidae mitogenomes), we regress codon-bias strength against ORF length; the MIMO-derived prediction is near-logarithmic saturation, and we additionally check whether codon-optimization gains plateau with gene length using the mRNAutilus property predictors, which were trained on roughly 13,000 human transcript half-lives and 1,282 human plus 995 mouse ribosome-profiling datasets (mRNAutilus).
Decisive Checks
The proposal succeeds if three things hold: measured angular drift grows with position and quantitatively predicts the length-dependent decay of pooled representations; RSCU-referenced anchoring restores long-transcript performance to within the margin Equi-mRNA reports at short lengths, beating the unanchored ablation; and natural codon bias scales near-logarithmically with gene length across independent lineages. It fails cleanly — and we will say so — if drift is bounded, if unanchored windowing suffices, or if bias scales linearly with length, in which case the MIMO analogy is decorative and the honest conclusion is that only the diagnostic, not the correction, was worth building.
References
- Equi-mRNA: Protein Translation Equivariant Encoding for mRNA Language Models
- Massive MIMO systems with hardware-constrained base stations
- Over-the-Air Computation in OFDM Systems With Imperfect Channel State Information
- Throughput Analysis of Massive MIMO Uplink With Low-Resolution ADCs
- Circuit-aware design of energy-efficient massive MIMO systems
- Computational identification of rare codons of Escherichia coli based on codon pairs preference
- BEACHES: Beamspace Channel Estimation for Multi-Antenna mmWave Systems and Beyond
- Codon-specific Ramachandran plots show amino acid backbone conformation depends on identity of the translated codon
- K-Mer Analyses Reveal Different Evolutionary Histories of Alpha, Beta, and Gamma Papillomaviruses
- Machine Learning for detection of viral sequences in human metagenomic datasets
- CodonTranslator: a conditional codon language model for codon optimization across life domains
- mRNABench: A curated benchmark for mature mRNA property and function prediction
- Comparison of the Complete Chloroplast Genomes in the Astragalus
- Differences between phytophagous and predatory species in Pentatomidae based on the mitochondrial genome
- Multi-Objective-Guided Generative Design of mRNA with Therapeutic Properties