A Live, Single-Cell Dual-Luciferase Assay for Measuring How microRNAs Filter Transcriptional Noise

Messenger RNA levels predict protein levels only 40% to 60% of the time, so up to half of protein variance is governed by events after transcription (A high-stringency blueprint of the human proteome). The leading mechanistic explanation is that small regulatory RNAs called microRNAs bind the 3' untranslated region (UTR) of a transcript and repress its translation into protein. Theory says this repression should do something more interesting than shift the mean: it should reshape the distribution of protein counts across cells, compressing noise when the microRNA accelerates transcript decay and amplifying it when it acts as an activator (Post-Transcriptional Regulation of Noise in Protein Distributions during Gene Expression), and it should filter short-lived fluctuations in transcription so that noisy mRNA does not produce noisy protein (MicroRNA filters Hox temporal transcription noise to confer boundary formation in the spinal cord). Yet the experiment used thousands of times per year to validate a microRNA–target pair — the dual-luciferase reporter assay — pools roughly ten thousand to a hundred thousand cells and lyses them, returning a single averaged light ratio. It averages away exactly the quantity the theory is about.

The core claim of this proposal is that the dual-luciferase assay can be re-engineered into a non-lytic, single-cell, time-resolved instrument, and that doing so converts a ubiquitous validation tool into a measurement machine for the open problems at the intersection of stochastic gene expression and transcript-protein discordance: which genes transcribe in bursts versus steadily, which microRNAs compress versus amplify protein noise, and by what molecular mechanism each microRNA acts. In a second, computational stage, the parameters this instrument extracts per target become priors for models that predict protein abundance from RNA across time — the problem that current static models (scChord, scTEL) cannot solve because they ignore translation lags and post-transcriptional filtering.

Core Claim

Three theoretical results jointly define what is missing experimentally. First, single-burst protein distributions are geometric regardless of what happens in between, so one cannot tell bursting transcription from steady transcription — or locate the origin of noise — without pairing burst data with steady-state variance measured across regulator concentrations (Stumpf); the five-parameter protocol that breaks this nonidentifiability barrier was written down analytically but never validated in living cells (Post-Transcriptional Regulation of Noise in Protein Distributions during Gene Expression). Second, steady-state rates are mathematically incapable of distinguishing whether a microRNA blocks translation initiation, elongation, or transcript stability; only the relaxation time — how fast protein levels recover after a perturbation — separates these mechanisms (Dynamical modeling of microRNA action on the protein translation process). Third, promoter toggling between active and inactive states (the "Telegraph" regime) — not simple Poisson production — is what generates the over-dispersed, bursty mRNA counts that single-cell methods actually observe (Dadiani et al.). All three results demand the same instrument: many single cells, tracked continuously through a perturbation, with both the regulated and a reference readout in the same cell.

We claim that a live-cell dual-luciferase system provides it. Concretely, Firefly luciferase carries the candidate 3' UTR and Renilla luciferase is constitutively expressed as an internal control; instead of a lysis-based endpoint, both enzymes are read repeatedly from the same living cells with membrane-permeable substrates under low-light imaging. The per-cell, per-time-point quantity is the ratiometric signal Firefly/Renilla, which cancels cell size, transfection load, and imaging gain, exactly as in the bulk protocol, but retains cell-to-cell and minute-to-minute structure.

The working model is deliberately minimal. Let $M(t)$ be the transcript count of the reporter, produced by a Telegraph process whose promoter switches between active and inactive states, and let protein (enzyme) abundance $P(t)$ follow

$$\dot P \;=\; k_{\mathrm{tl}}\,M(t)\;-\;\gamma_p\,P,$$

where $k_{\mathrm{tl}}$ is the translation rate and $\gamma_p$ the protein decay rate. A microRNA enters either as added transcript decay (reducing the lifetime of $M$) or as reduced $k_{\mathrm{tl}}$ (pure translational repression). The two modes leave the same mean but different fingerprints: decay modulation ties protein noise to the transcriptional burst size, translational repression decouples it (Post-Transcriptional Regulation of Noise in Protein Distributions during Gene Expression), and they recover from perturbation on different timescales (Dynamical modeling of microRNA action on the protein translation process).

Research Program

The program has three stages. Stage one is the assay itself. Reporter constructs follow the established architecture — Firefly with the full-length 3' UTR or a mutated seed sequence downstream, Renilla constitutive — but integrated as a single copy into the genome rather than carried on high-copy episomal plasmids, which addresses the known criticism that plasmid copy numbers buffer or exaggerate endogenous kinetics. The readout is time-lapse bioluminescence imaging at single-cell resolution over 48 to 72 hours, sampling every few minutes, with the Firefly/Renilla ratio computed per cell per frame. The critical engineering risks are substrate toxicity and photon-limited signal at single-enzyme-copy sensitivity; these are the first things to characterize, and if single-cell resolution fails for weakly expressed constructs, the fallback is per-cell-family (micro-colony) resolution, which still preserves far more distributional structure than a 96-well bulk average.

Stage two is the perturbation grid designed around the theory. For each validated microRNA–target pair (drawn from the experimentally curated interactions in miRTarBase, prioritized by repression-strength predictions from MiRGraph), we record single-cell trajectories across a dose ladder of microRNA mimic and inhibitor, plus the seed-mutant control in which the binding site is disrupted. This grid maps directly onto the untested five-parameter protocol: means and variances of both transcript and protein readouts across regulator concentrations, which is what Post-Transcriptional Regulation of Noise in Protein Distributions during Gene Expression proved sufficient to identify the promoter regime (Poisson versus Telegraph) and the sign and strength of the noise filter. Relaxation is measured by washout of the microRNA inhibitor or induction pulses, giving the recovery-time signatures that Dynamical modeling of microRNA action on the protein translation process showed are the only way to discriminate mechanism. As a biological anchor, we replicate the mir-27/Hoxa5 feed-forward system in the embryonic stem cell motor-neuron differentiation protocol of MicroRNA filters Hox temporal transcription noise to confer boundary formation in the spinal cord, where 5–10% of motor neurons in mir-27 or Dicer mutants wrongly co-express both Hox proteins — a single-cell phenotype that the bulk assay is structurally blind to and our assay should see as excess tail mass in the ratio distribution.

Stage three is computational. We fit the Telegraph-plus-filter stochastic model to the trajectory ensembles by moment matching on means, variances, and autocorrelation times, inferring per-target parameters: burst frequency and size, filter mode (transcript decay versus translational repression), filter strength, and relaxation time. These parameters then become structured priors for RNA-to-protein prediction over time. Current predictors treat the RNA-to-protein map as a static snapshot — scChord transports RNA embeddings to protein space with conditional flow matching, PIKE-R2P adds protein-protein interaction priors, scTEL adds uncertainty via quantile loss — but none incorporates the lag and dampening that physically sit between transcript and protein. The clearest statement of the gap comes from dynamics: FLeCS models single-cell regulation as deterministic production-minus-decay ODEs on RNA alone, explicitly deferring stochastic terms and protein levels, and Cell-MNN learns linearized dynamics in RNA latent space while treating paired protein measurements as static endpoints. We will graft the inferred per-target filter parameters onto a temporal predictor that propagates RNA dynamics through a per-gene translational filter to predict protein abundance at a later time point, evaluated on longitudinal CITE-seq benchmarks.

Baselines

The do-nothing reference point is the assay as it exists today: the bulk dual-luciferase ratiometric readout pooled over roughly $10^4$–$10^5$ cells per well, run on our own constructs so the comparison is literal rather than literature-based. It is the strongest competitor for the validation question because it is the accepted gold standard for confirming direct microRNA–target binding, and our seed-mutant controls are its own logic. It cannot, by construction, report a Fano factor, a distribution shape, or a relaxation time — it returns one mean per well. Beating it means recovering, from the same wells, the distributional and kinetic quantities it averages away, and showing they agree with its mean.

The strongest existing method for the temporal prediction task is Cell-MNN, which models single-cell evolution with locally linear operators in RNA latent space and matches distributions across time points with Maximum Mean Discrepancy. It reports an average Earth Mover's Distance of 0.741 across the Embryoid Body, Cite, and Multi longitudinal benchmarks (versus 0.776 for OT-MFM), and it recovered gene-gene regulation at a 69% F1 score on the TRRUST reference versus 46% for SCODE. We will run it on the same three longitudinal benchmarks with the same cell counts and the same tuning budget, and additionally on the 7-day CITE-seq benchmark with protein endpoints held out, which is the setting it currently treats as snapshots. Beating it means lower Earth Mover's Distance on held-out protein distributions at future time points, not just RNA.

The strongest static RNA-to-protein predictor is scChord, which reports the lowest RMSE and highest Pearson correlation across CITE-seq benchmarks while training on 100,000 cells in approximately 27 minutes with a 229 MB GPU memory footprint (no reported numeric RMSE/PCC values found on Lacuna); scTEL reports 86.6% cell-type classification accuracy on GSE164378 versus 77.0% for sciPENN and 76.1% for Seurat. Both will be run on the same longitudinal CITE-seq data, given the frozen-time slice they were designed for, with matched compute. Beating them means lower prediction error on protein abundance at time $t_2$ given RNA at $t_1$ — the lagged setting they do not model.

Finally, the key ablation is our own model with the microRNA-filter prior removed: the same architecture propagating RNA dynamics to protein with a fixed, gene-independent translation lag. Any gain that survives comes from the per-target filter parameters, which is the proposal's central ingredient; any gain that vanishes means the improvement was just "any temporal model beats none," and we would report that honestly.

Why not just…?

Why not just use the bulk dual-luciferase assay as-is? Because the quantities this proposal targets are mathematically invisible at the population mean. The noise-shaping theory predicts Fano-factor changes of up to 0.5 relative to the mean under decay modulation (Post-Transcriptional Regulation of Noise in Protein Distributions during Gene Expression), and the developmental phenotype that motivates the whole program — 5–10% of motor neurons mis-specifying boundaries in mir-27 or Dicer mutants (MicroRNA filters Hox temporal transcription noise to confer boundary formation in the spinal cord) — is a tail of a distribution that a well average entirely hides. The bulk assay remains the right tool for confirming that a binding event is direct; it is simply the wrong instrument for characterizing what the binding does to noise.

Why not just do a time course of single-molecule RNA FISH, as Li et al. did? This is the most serious alternative, because the Hox study already demonstrated single-cell noise filtering in a real developmental context. But fixed-cell FISH is destructive at each time point: cells measured at $t_1$ are never measured at $t_2$, so relaxation times — the only signature that discriminates mechanism, per Dynamical modeling of microRNA action on the protein translation process — must be inferred across different cells rather than watched within one. Our assay's decisive advantage is the within-cell trajectory. If, in practice, bioluminescence photon noise makes per-cell ratios too coarse to beat cross-sectional FISH, the FISH time course becomes the better experiment and we would say so; the first months of the program are explicitly a head-to-head on the mir-27/Hoxa5 system where both can be run.

Why not just add more data to the existing static predictors? Because the failure is structural, not statistical. Steady-state RNA-to-protein mappings cannot distinguish translational-regulation mechanisms at all (Dynamical modeling of microRNA action on the protein translation process), point regression collapses multimodal protein states like bimodal CD4 to their conditional average (scChord), and scChord's Correlation Matrix Distance degrades as targets scale to the hundreds of proteins in GSE164378 precisely because it lacks regulatory priors. More cells sharpen a snapshot; they do not manufacture a time axis or a mechanism.

Why not just use optimal-transport trajectory methods on the time-series data? They are the established tool for snapshot trajectory reconstruction, but they scale quadratically with cell count and the OT-based variants crashed with out-of-memory errors at 250,000 cells on the inflated versions of the same benchmarks, while Cell-MNN matched its 0.741 average Earth Mover's Distance with linear scaling. More fundamentally, transport moves distributions without representing a mechanism, so it cannot tell a translational filter from a decay filter. We use it only as an evaluation metric (Earth Mover's Distance on held-out distributions), not as the model.

Decisive Checks

Four results would decide the proposal. First, the seed-mutant control: mutating the microRNA binding site must restore not only the mean ratio, as in the bulk assay, but also the wild-type distribution width and relaxation time — a three-way restoration that no bulk protocol has shown. Second, the nonidentifiability test: the five-parameter dose grid must classify known Poisson-like and Telegraph-like promoters correctly, using synthetic mixtures as ground truth before the biological targets. Third, mechanism discrimination: for microRNAs with independently established mechanisms, the relaxation-time signature must call transcript-decay actors differently from translational-repression actors, which is precisely what steady-state measurements provably cannot do. Fourth, the computational payoff: on the 7-day longitudinal CITE-seq benchmark, the filter-prior model must beat Cell-MNN's reported 0.741 average Earth Mover's Distance on held-out future protein distributions, and the ablation without the filter prior must lose most of that gain. If the assay works but the priors do not transfer — single-copy genomic reporters may behave differently from endogenous loci — the biological findings on noise filtering and mechanism identification stand on their own, and we would report the transfer failure as a genuine negative result rather than tune around it.