A dual-luciferase reporter assay is the standard experiment for proving that a microRNA — a roughly 22-nucleotide RNA molecule that cells use to silence genes after transcription — directly represses a specific messenger RNA. The biologist builds a plasmid encoding two light-producing enzymes: firefly luciferase, whose transcript carries the candidate target sequence (the 3' untranslated region, or 3' UTR, of the gene of interest), and Renilla luciferase, which is left untouched as an internal control. Dividing firefly light by Renilla light cancels well-to-well noise and isolates the repression caused when the microRNA binds the inserted sequence. The assay is almost always treated as a measurement of a passive system: the cell is a reaction vessel, and the construct is a probe.
But the cell is not a passive vessel, and this is where a second research tradition becomes useful. The silencing machinery — the RNA-induced silencing complex, RISC, which carries each microRNA to its targets — exists in finite amounts, so a plasmid transfected at unnaturally high copy number soaks up microRNAs that would otherwise regulate the cell's own transcripts, a effect the field calls sponging. Cells can also answer over hours to days by adjusting microRNA abundance. The measured luminescence ratio therefore depends not only on what the designer built but on how the cell responds to it. That is exactly the shape of a leader-follower, or Stackelberg, game: one player commits to a decision, and a second player, who cannot undo the commitment, best-responds to it. We propose to treat reporter construct design as such a game, to build the corresponding bilevel optimization, and to test in real dual-luciferase assays whether anticipating the cell's response produces better constructs than the best static design rules available.
Core Claim
The construct designer is the leader, committing to a design vector $d$: the number of microRNA binding sites, each site's binding free energy $\Delta G$ and local structural accessibility $\Delta\Delta G$, the plasmid copy number, and the promoter strength. The cell is the follower: after the commitment, it settles into an attractor state $s^*(d)$ — the free abundance of each relevant microRNA and the allocation of RISC across all competing sites, endogenous and synthetic. The follower is adaptive but myopic: it reacts to the immediate perturbation (fast competitive binding, slower homeostatic adjustment of microRNA supply) and cannot plan around the designer. Myopia is exactly what a leader can exploit, a structure established for biological opponents by Steering Evolution Strategically, which treats an evolving population as a follower that optimizes only immediate fitness while the treater plans ahead. Our claim is that this structure is present, measurably, at the scale of a single transfection, and that design tools which freeze the cell cannot see it.
Concretely, the design problem becomes a bilevel program:
$$\max_{d \in \mathcal{D}}\; u\big(d, s^*(d)\big), \qquad s^*(d) = \arg\min_{s \in \mathcal{S}(d)}\; H(s; d),$$
where $H$ is the cell's instantaneous homeostasis error (how far native target repression sits from its pre-perturbation state) and the leader's utility $u$ is minus the log-error between the achieved firefly/Renilla ratio and a target repression level, minus a penalty $\lambda$ on the worst log fold-change imposed on any endogenous target. The penalty term has a direct ancestor in Vorobeychik et al., who showed a publisher can generalize data so a re-identification attacker's expected payoff drops to zero — our analog is a construct whose collateral dysregulation of every native target falls below the cell's own biological noise floor, a "no-dysregulation equilibrium."
The sharpest way to state why this is not a re-description of existing practice: the MPRA study showed, across 14,151 designed 3' UTR variants in four cell lines, that intracellular microRNA abundance alone explains 81% of the variance in repression, and fitted a Gradient Boosting Regression (GBR) reaching a Pearson correlation of $R = 0.84$ for protein repression. That model conditions on the cell state as measured before the perturbation. A single-construct dual-luciferase assay at realistic plasmid loads changes that state. The gap between the conditioned-on state and the post-load state $s^*(d)$ is precisely what the bilevel model predicts and the static model cannot. The MPRA authors themselves note that repression for miR-92a saturates near four sites — a signature of resource limitation that a fixed-context regression documents but cannot design around.
Where This Sits in the Literature
Three lines of work converge here, each supplying something different. Sandholm's extensive-form treatment games supply the conceptual template — a planning leader versus a myopic biological follower — and explicitly call for empirical validation in controlled cell systems; it never formalizes post-transcriptional regulation, microRNAs, or reporter systems, and we supply exactly that instantiation. The privacy games supply the computational machinery: Vorobeychik et al. demonstrate that bilevel search over combinatorial action spaces is tractable via lattice pruning (their heuristic ran roughly 17 times faster than exact backward induction, 15.8ms versus 268.1ms, and found the optimum 99.47% of the time), and the multistage genomic masking game shows a leader can share ~86% of data at under 1% expected risk by anticipating a follower's economics — both are static-data settings, and we port their solver philosophy to a living system. The MPRA paper supplies the calibration data and the static competitor; it treats the cell as a fixed measurement context, which is the assumption we relax. Two further papers constrain the design rather than motivate it: Kazıklı et al. proved that linear signaling strategies remain optimal for a Stackelberg sender under quadratic costs and Gaussian sources, which justifies restricting our leader's policy class to low-dimensional design parameterizations instead of searching arbitrary sequences; and MBSTAR documents that transcripts contain many sequence-compatible but non-functional decoy sites (its site-level precision against PAR-CLIP peaked at an F-score of only 0.337), which forces our follower model to operate at the level of individual site occupancies rather than transcript-level labels.
Research Program
We build three things. First, a follower model with two layers: a fast titration layer computing competitive binding equilibrium of a finite RISC pool across all sites as a function of affinities and abundances, and a slow adaptation layer in which the cell myopically adjusts microRNA abundance to reduce the homeostasis error $H$. Both layers are calibrated on the MPRA library (14,151 oligonucleotides across K562, MCF7, HEK293, and HepG2), which varies site multiplicity, $\Delta G$, $\Delta\Delta G$, and cell-line microRNA context — the exact factors our state variables encode. Second, a bilevel solver that nests the follower attractor inside a leader search over $\mathcal{D}$, using lattice-style pruning in the spirit of Vorobeychik et al. to keep the combinatorics manageable. Third, a wet-lab testbed: the designer's construct is a standard firefly/Renilla dual-luciferase plasmid, and collateral damage is read out through an orthogonal fluorescent sentinel (an mNeonGreen/mCherry dual reporter, the architecture used in the MPRA) carrying a native, literature-validated 3' UTR — for example the JAK2 3' UTR targeted by miR-381-3p in hepatic ischemia-reperfusion work Serum miR-381-3p or the SHMT1 3' UTR targeted by miR-944 in bladder cancer miR-944 inhibits malignant progression. A qRT-PCR time course of microRNA abundance at 24, 48, and 72 hours directly measures whether the slow adaptation layer is real.
Baselines
The naive reference point is field-standard practice: clone the native 3' UTR, or a single canonical seed-matched site, into the plasmid at the usual transfection dose, with no model at all. We also include its greedy cousin — the maximal design, as many high-affinity sites as fit — which is what any designer does when told "make repression strong" without a model of the cell's answer. These anchor the two ends of the design spectrum and, in the sentinel channel, will show whether naive designs cause the collateral dysregulation the theory predicts.
The strongest competitor is static model-based design using the GBR of the MPRA study, which reported $R = 0.84$ on protein repression and is, to our knowledge, the only quantitative predictor of repression magnitude for designed 3' UTRs in the literature. We rank the same candidate construct pool by GBR-predicted repression and pick its top design for each repression setpoint. To keep the comparison honest, both methods draw candidates from the same pool, defined by the same interaction-prediction front end (site candidates filtered by RFMirTarget, whose Random Forest reached an AUC of roughly 0.96, and cross-checked against MiRGraph, which reported an AUROC of 0.8706 on miRTarBase), and both receive the same in-silico evaluation budget and the same wet-lab replicates, cell line, and transfection protocol. Candidate site sets are fixed before either method sees them, so no method gets a better menu. Beating this baseline means two things: matching or exceeding $R = 0.84$ on held-out MPRA variants and achieving a target repression level in vitro with at most half the sentinel dysregulation of the GBR-optimal design.
The key ablation is the same bilevel optimizer with the follower frozen: the leader optimizes against the pre-transfection cell state $s_0$ rather than the post-adaptation state $s^*(d)$. This isolates anticipation itself as the ingredient, separate from any benefit of our model class or search procedure. A second ablation keeps only the fast titration layer and drops the slow homeostatic layer, isolating whether the exploitable myopia lives in the hours-scale adaptation or only in instantaneous resource competition.
Why not just…?
Why not just optimize the GBR regression model directly? This is the most serious alternative, and we do not argue it away — we test it. The GBR already encodes multiplicity, $\Delta G$, and $\Delta\Delta G$ effects (including the 4-fold accessibility effect and the 57-fold dynamic range reported in the MPRA study), and it is the reigning quantitative model. Its structural limit is that it conditions on pre-perturbation microRNA abundance, while its own finding — abundance explains 81% of repression variance — implies that any construct which shifts abundance moves the prediction off-manifold. Pooled MPRAs keep each variant at low copy and so mostly stay on-manifold; single-construct dual-luciferase assays at high copy do not. If the head-to-head experiment shows GBR-optimal designs match our setpoint accuracy and sentinel profile anyway, the game layer adds nothing at these doses and we will report that boundary honestly.
Why not just run an active-learning campaign and let the data speak? One could screen a few thousand constructs by design-of-experiments and skip modeling entirely. The file's own evidence cuts against this: Vorobeychik et al. faced exactly this curse of dimensionality in combinatorial leader action spaces and found exact search intractable without structure, and the space here (multiplicity × affinity × accessibility × copy number × promoter) is comparably large. Active learning spends wet-lab budget rediscovering, construct by construct, a follower response function we can calibrate once from the existing MPRA data and reuse for every future design. It is also silent on counterfactuals — what the cell would have done under a load you never tested — which is the entire point of a leader model.
Why not just write down standard titration ODEs? Partly right, and we say so: the fast layer of our follower model is standard competitive-binding kinetics, and if that were the whole story the "game" would be an ordinary simulation-based optimization. The scientific bet is the slow layer — that the cell's microRNA supply responds myopically over the assay's one-to-three-day window, making it a genuine follower in Sandholm's sense rather than a fixed buffer. The titration-only ablation is designed to settle this, and the qRT-PCR time course settles it even more directly: if microRNA abundance does not move under reporter load, the slow layer is falsified, the problem collapses to static titration design, and the correct conclusion is a precise negative result — a bound on the regime (pooled low-copy assays) where bilevel thinking is unnecessary and the regime (high-copy single-construct assays) where only fast titration matters.
Proposed Experiments and Decisive Checks
The first experiment is in silico. We hold out a fraction of the MPRA library, fit the follower model on the rest, and predict repression for held-out variants. Success is a held-out correlation at or above the GBR's reported $R = 0.84$, with the decisive margin coming from cross-cell-line transfer: train on three cell lines, predict repression in the fourth, where a model with an explicit abundance state should transfer and a context-conditioned regressor should degrade. No reported cross-line transfer result exists on Lacuna for the GBR, so this comparison would be new evidence either way.
The second experiment is the in vitro head-to-head. For a fixed repression setpoint (say, five-fold repression of the designer construct in HEK293T), we transfect the GBR-optimal design, the Stackelberg-optimal design, and the naive single-site and maximal designs, in the same plates with the same replicate count, each alongside the native-UTR sentinel reporter. Primary endpoint: achieved firefly/Renilla ratio versus setpoint. Secondary endpoint: sentinel log fold-change. The Stackelberg design wins if it hits the setpoint within a pre-registered tolerance while holding sentinel dysregulation to at most half that of the GBR design, and the naive designs are expected to show the largest collateral effects, confirming the phenomenon the model targets.
The third experiment is the dose–response and falsification check. We titrate plasmid amount across an order of magnitude for each design and measure where repression switches on — titration systems are known to act as thresholds, and a leader that models the follower can place that threshold where it wants. Alongside this, the qRT-PCR time course tests the homeostasis hypothesis directly. The proposal is falsifiable in a clean way: if abundance is flat and titration-only ablation matches the full model everywhere, then the leader-follower formulation earns nothing at assay timescales, and the honest deliverable is a calibrated static titration design tool plus a documented boundary condition for when game-theoretic design begins to matter. If abundance does move, and the bilevel designs dominate, we will have shown — for the first time in a system small enough to fully measure — a designed leader exploiting a living follower's myopia, with the dual-luciferase assay serving not as a passive probe but as the arena of the game itself.
References
- Steering Evolution Strategically: Computational Game Theory and Opponent Exploitation for Treatment Planning, Drug Design, and Synthetic Biology
- A Game Theoretic Framework for Analyzing Re-Identification Risk
- Unraveling the determinants of microRNA mediated regulation using a massively parallel reporter assay
- Using game theory to thwart multistage privacy intrusions when sharing data
- Quadratic Privacy-Signaling Games and the MMSE Information Bottleneck Problem for Gaussian Sources
- MBSTAR: multiple instance learning for predicting specific functional binding sites in microRNA targets
- Serum miR-381-3p: Diagnostic Role and Mechanisms in Liver Transplant Ischemia-reperfusion Injury
- miR-944 inhibits malignant progression of bladder cancer through ATIC/AKT/FOXO3 A axis mediated by SHMT1
- RFMirTarget: Predicting Human MicroRNA Target Genes with a Random Forest Classifier
- MiRGraph: A hybrid deep learning approach to identify microRNA-target interactions by integrating heterogeneous regulatory network and genomic sequences