NMD Efficiency as a Competing Removal Rate

Cells destroy messenger RNAs that carry a premature stop signal, a quality-control process called nonsense-mediated decay (NMD). Large genomic studies measure how well this works by counting sequencing reads. In a person carrying one mutant and one normal copy of a gene, they compare reads from the mutant copy with reads from the normal copy. This ratio, called allele-specific expression (ASE), has supported a much-cited conclusion: genes that are intolerant of mutations seem to have more efficient NMD. Such genes are called "constrained," and their essentiality is measured by gnomAD's LOEUF score.

The ratio, however, is not a rate. A transcript's steady-state abundance reflects two decay routes acting in parallel: NMD and the ordinary turnover every mRNA undergoes. The ratio therefore compresses non-linearly as NMD speeds up relative to normal turnover. Integrated silicon photonics faces the same measurement problem in active carrier sweep-out. There, reverse-biased diodes pull light-generated charges out of a waveguide before they absorb light. Light loss in such a waveguide is set by an effective carrier lifetime made of two parallel removal channels. Engineers read device performance off that lifetime, not off the raw loss. We propose to re-express NMD efficiency in that framework. We would calibrate each gene against its own NMD-escaping variants, which serve as a zero-NMD reference much as a zero-bias measurement does in the photonic case. We would then test whether the reported link between gene constraint and NMD efficiency survives once ordinary turnover is separated from decay.

Core Claim

The claim is that part of the apparent gene-to-gene and tissue-to-tissue variation in NMD efficiency is a measurement artifact. Gene-specific baseline turnover and allele-mapping biases distort the steady-state ratio. A two-channel model calibrated within each gene recovers an NMD rate that can be compared across genes. Write $k_N$ for the NMD removal rate of a stop-bearing transcript and $k_b$ for its gene's ordinary decay rate. Write $r$ for the observed mutant-to-normal read ratio for a decay-triggering stop, and $r_0$ for the same ratio when NMD is absent. Then at steady state

$$ r = r_0 \, \frac{k_b}{k_b + k_N}, \qquad \text{so} \qquad \frac{k_N}{k_b} = \frac{r_0}{r} - 1. $$

This is the direct counterpart of the sweep-out lifetime relation

$$ \frac{1}{\tau_{\text{eff}}} = \frac{1}{\tau_{\text{rec}}} + \frac{1}{\tau_{\text{sw}}}, $$

where $\tau_{\text{rec}}$ is the intrinsic recombination time and $\tau_{\text{sw}}$ is the field-driven extraction time. In both systems, the observable improves strongly while the added channel is slower than the intrinsic one. Once the added channel dominates, the observable saturates and stops distinguishing large removal rates from very large ones. Peucheret and colleagues describe sweep-out working on a sub-nanosecond timescale, which suppresses free-carrier absorption enough to reach a 20 dB phase-sensitive extinction ratio (Parametric Phase-sensitive and Phase-insensitive All-optical Signal Processing on Multiple Nonlinear Platforms). Xu and colleagues pair such diodes with low-loss waveguides and stress the trade-off between faster extraction and the extra loss the doped regions add (Progress in silicon-based reconfigurable and programmable all-optical signal processing chips).

The photonic literature contributes a measurement logic, not hardware. That logic is to report the removal rate, calibrate it against a reference where the added channel is switched off, and expect the observable to compress at high rates. We claim the NMD field has not adopted this logic.

The two reference points have to be defined carefully. The zero-NMD reference $r_0$ is estimated for each gene from stop codons that escape decay by position. These are stops in the last exon, in the final 50–55 nucleotides of the second-to-last exon, or very close to the start codon. Balasubramanian and colleagues characterize the first two classes as escaping (Using ALoFT to determine the impact of putative loss-of-function variants in protein-coding genes), and Yang and colleagues describe start-proximal escape (Connexin40 nonsense mutation in familial atrial fibrillation). Escaping alleles in the same gene share its mapping biases, any non-NMD effects of truncation on transcript stability, and its library context. They therefore play the role of the zero-bias measurement. The baseline rate $k_b$ comes from public transcriptome-wide mRNA half-life measurements. The research file lists no such dataset, so we flag this as an external input that we will assemble rather than cite.

Relation to Prior Work

Three papers motivate the question and supply the data machinery.

Kim and colleagues fit a 14-feature random-forest regression to ASE from somatic mutations in The Cancer Genome Atlas (TCGA) and validate it on germline cohorts (A spectrum of nonsense-mediated mRNA decay efficiency along the degree of mutational constraint). They report that constrained genes show stronger NMD, that NMD is stable across tissues, and that downstream exon count is the strongest predictor. They also leave open whether essential genes evolved efficient NMD or merely look that way. Our work reuses their ASE-on-TCGA design. Unlike them, we model the observed ratio as a saturating function of two rates rather than as the target itself.

Sun and Chen map variants across 48 GTEx tissues (Mapping genetic variants for nonsense-mediated mRNA decay regulation across human tissues). They split NMD regulation into the share of transcripts targeted for NMD ($\alpha$) and a decay-efficiency term ($\theta$), and find strong tissue specificity, especially in brain. Their $\theta$ is still a fraction inferred from steady state, with ordinary turnover absorbed into it. We separate $k_N$ from $k_b$ explicitly, which lets us ask whether tissue-specific efficiency partly reflects tissue-specific turnover. That bears directly on the conflict between their findings and Kim's tissue-invariance claim.

ALoFT (Using ALoFT to determine the impact of putative loss-of-function variants in protein-coding genes) classifies loss-of-function variants and uses NMD predictions as features. It shows why the question matters clinically: whether a transcript escapes decay decides whether a truncated, possibly toxic protein is made.

Baboo and colleagues show that much protein synthesis and turnover happens within minutes through exploratory translation (Most Human Proteins Made in Both Nucleus and Cytoplasm Turn Over within Minutes). This constrains our model. If NMD frequently acts within minutes while many mRNAs live for hours, many genes sit in the saturated regime, where the steady-state ratio is a poor readout.

The research file lists no work connecting the two directions. Its suggested bridges mostly do not survive scrutiny. One is photonic accelerators for NMD rules (What Is Next for LLMs? Next-Generation AI Computing Hardware Using Photonic Chips), whose authors themselves cite conversion overhead and weak precision. The other is nonlinear optical biosensing of truncated proteins. We discuss both below.

Research Program

First, we assemble paired allele sets from TCGA somatic mutations and from GTEx germline variants. For each gene, these sets contain stops that should trigger decay and stops that escape by position. Most genes will have few escaping variants. We therefore estimate $r_0$ with a hierarchical model that shrinks each gene toward the value for its transcript class. The model propagates uncertainty into the inferred $k_N/k_b$ and into the absolute rate $k_N$ obtained by multiplying by $k_b = \ln 2 / t_{1/2}$, where $t_{1/2}$ is the measured half-life.

Second, we run simulations in which ground-truth $k_N$ is independent of LOEUF but $k_b$ is correlated with it. These test whether the raw-ratio analysis spuriously recovers Kim's constraint–efficiency association, and whether our transform removes it. Rerunning the simulation with a true constraint–efficiency link checks the transform's statistical power.

Third, we refit the association between constraint and NMD efficiency on the real data using three readouts: the raw ratio, the within-gene-calibrated ratio, and the absolute rate. For GTEx, we repeat this per tissue to test whether brain-enriched NMD regulation persists after calibration.

Baselines

The naive reference is the raw ASE ratio regressed on LOEUF, with no calibration or rate transform. It reproduces the published analysis style, and any difference from it is our effect.

The strongest existing method is Kim and colleagues' 14-feature random forest, which explained 43.7% of total ASE variance. The authors equate this to about 95% of the explainable variance. We will retrain it on the same TCGA split, features, and tuning budget, and use it in two ways. On predictive skill, our model must match it within noise on held-out ASE; we do not expect to beat it on raw variance explained, and that is not the claim. On the scientific question, we compare LOEUF's partial dependence and importance rank (8th of 14 in their fit) before and after the target is replaced by the calibrated rate. We count a meaningful change as either of two outcomes. One is that the LOEUF–efficiency slope shrinks with confidence intervals excluding the original value. The other is that the slope persists with tighter intervals. Either result answers the open causal-direction question more sharply than the original fit does.

ALoFT's reported multiclass AUC of 0.97 serves as a downstream reference. We will check whether calibrated rates, added as a feature, change dominant-versus-recessive calls on its benchmark sets (HGMD, ClinVar, 1000 Genomes). Two ablations isolate the key ingredients. One applies the rate transform with a global rather than within-gene $r_0$. The other uses the within-gene $r_0$ without half-life scaling. These show whether the calibration or the turnover correction carries any effect.

Why not just…?

Why not just add mRNA half-life as a feature to the existing random forest? This is the most serious alternative and could settle much of the question cheaply. We will run it as a baseline. Its weakness is that a tree ensemble given half-life as a feature can absorb turnover without applying the saturating relation. Partial dependence of LOEUF in correlated feature sets is also hard to interpret. Most importantly, it cannot supply the within-gene zero-NMD reference that removes allele-mapping and truncation-stability biases. If the enriched random forest and our model give the same LOEUF conclusion, we will report that our contribution reduces to interpretability and calibrated uncertainty. If they disagree, the simulation study in the research program will show which one recovers the planted ground truth.

Why not just use Sun and Chen's split into targeted share and decay efficiency? That split separates which transcripts are targeted from how efficiently they decay. Both quantities, however, are inferred from steady-state fractions across individuals, so ordinary turnover remains folded into the efficiency term. We will compare our tissue-level rates against their tissue-specificity pattern in GTEx to see whether calibration removes part of it.

Why not just measure decay directly by blocking translation or knocking down the decay machinery? Translation-inhibition and minigene assays measure rates directly and are the gold standard. They are also low-throughput and cannot be run retroactively on cohort variants. We will treat any such measurements we can obtain as validation for a small set of genes, not as a competitor at genome scale.

Why not build a photonic accelerator or a nonlinear optical sensor instead? The file gives no evidence that NMD rule evaluation is computationally limited. The random forests already run on standard computers, and the photonic-AI review itself concedes precision and conversion bottlenecks. If a high-power optical sensor were wanted, nitride or lithium tantalate (Lithium tantalate photonic integrated circuits for volume manufacturing) avoids the need for sweep-out altogether. We therefore take only the measurement logic from sweep-out, not the device.

Decisive Checks

The first check is in simulation. The raw-ratio pipeline should report a spurious constraint effect when only turnover correlates with LOEUF, and the calibrated pipeline should not.

The second check, on TCGA, concerns how the constraint–efficiency slope behaves under calibration. If it collapses, Kim's "essential genes have stronger NMD" reading is substantially an artifact of turnover. If it persists, it is strengthened.

The third check, on GTEx, concerns tissue variation. If it shrinks under calibration, the apparent conflict between tissue-invariant and tissue-specific NMD is partly a measurement issue.

A further check concerns the very fastest-decaying genes, which sit in the saturated regime. They should show inflated uncertainty under our model and overconfident estimates under the raw ratio. This is the signature the sweep-out analogy predicts.