The dual-luciferase reporter assay — a test-tube experiment where a gene's regulatory tail is fused to a light-producing enzyme, and a microRNA's binding dims the light — is the accepted way to prove that a microRNA directly silences a particular target gene. Yet the assay is run in a single homogeneous cell line, flooded with synthetic microRNA at concentrations far above anything a cell in the body ever sees, and read once at a fixed endpoint. Meanwhile, every property the assay ignores — heterogeneity across cell types, change over time, physiological abundance — is exactly where drug targets discovered in culture go on to fail in patients. Liu et al. made the point sharply two decades ago: in twelve real brain tumors, most microRNA–mRNA pairs with strong negative correlation were indirect downstream ripples, not the direct targets that overexpression assays in dishes had suggested.
The core claim of this proposal is that the luciferase assay is not the problem — it is an underused measurement. What the assay actually measures is a cell-type-invariant physical efficacy: how strongly a microRNA's binding sites in a given 3' untranslated region can repress, at saturating microRNA dose, when the seed region is intact. What fails in vivo is not the binding physics but the stoichiometry: how much microRNA, how much target mRNA, and how many competing RNA transcripts share that microRNA, in a specific cell type, at a specific moment. Both of these are measurable, and the transfer between them is a learnable, testable function. We propose to build and validate that function — a reporter-calibrated, stoichiometric model that predicts, per cell type and per time point, whether a luciferase-validated interaction will actually repress its target inside a heterogeneous living tissue.
Core Claim
Call the model the repression-transfer model. It predicts the log fold change of target mRNA $i$ in cell type $c$ at time $t$ as
$$\Delta_{i,c,t} \;=\; \beta_{i,m}\; \sigma\!\left( \log \frac{A_{m,c,t}\, S_{i,m}}{T_{i,c,t} + \rho\, C_{i,c,t} + K_{i,m}} \right)$$
where $\beta_{i,m}$ is the maximum repression of microRNA $m$ on target $i$, taken from the plateau of a luciferase dose–response curve rather than a single saturated reading; $A_{m,c,t}$ is the deconvolved abundance of microRNA $m$ in cell type $c$ at time $t$; $S_{i,m}$ is a sequence-based binding-strength prior; $T_{i,c,t}$ is the target's own mRNA abundance; $C_{i,c,t}$ sums competing binding sites (the "sponge" load of other transcripts), weighted by $\rho$; $K_{i,m}$ is the half-maximum constant read off the same luciferase dose–response curve; and $\sigma$ is a logistic squashing. The model has roughly a handful of fitted parameters — deliberately small, because geneRNIB found that simple correlation and gradient-boosting methods beat large foundation models on regulatory-network inference, and a skeptical reviewer should not have to trust a neural network for the claim to hold.
The decisive novelty is what feeds the likelihood. bayesReact already infers microRNA activity from expression rankings at single-cell and time-course resolution — it tracked miR-124-3p rising during neuronal development — but its binding model is naive seed matching, with no physical reporter measurement anywhere in the machinery. TargetNet predicts functional sites from sequence with high recall but no cellular, temporal, or in vivo context. DeepSADR transfers between cell lines and patients, but only for small-molecule drugs, only on static bulk profiles, and with no post-transcriptional layer. Our proposal is the first to put reporter-measured repression efficacy, as a dose–response curve, inside a generative model whose remaining inputs are cell-type- and time-resolved abundances, and to test the composite against endogenous in vivo data.
Research Program
The program has three components. First, we build a reporter panel: for a few hundred microRNA–target pairs spanning canonical and non-canonical sites, we run dual-luciferase assays across a titration of mimic concentrations, with the standard firefly/Renilla ratiometric readout and seed-mutant controls. This yields, per pair, both a plateau $\beta$ (maximum direct repression) and a sensitivity constant $K$. Measuring the dose–response rather than one saturating point is itself a methodological contribution — it converts the assay's chief artifact, non-physiological overexpression, into the very quantity the model needs. Second, we assemble the stoichiometric inputs: single-cell reference atlases and bulk profiles from GTEx, TCGA, and time-course data, deconvolved with the strongest available tools. POPPY maps bulk clinical transcriptomes into single-cell representations and improved cohort composition correlation by up to 714%, and Count Bridges recovers single-cell count distributions from bulk and spatial spots while respecting the discrete, integer nature of counts — we use these as the abundance source rather than building a new deconvolver. For the binding prior $S_{i,m}$ we use TargetNet-style scores, which capture non-canonical sites that rigid seed rules miss. Third, we fit and test the transfer model end to end on endogenous data, holding out tissues, cancer types, and time points.
Baselines
The naive reference point is seed-match prediction plus bulk Pearson correlation: take TargetScan-style predictions of canonical seed matches and correlate bulk microRNA and mRNA expression across tissues. This is what most target papers in the file effectively do, and its ceiling is documented — on the miRAW benchmark of validated functional sites, TargetScan achieved an F1 of 0.4712 and recall of 0.3945, versus 0.7739 and 0.9411 for TargetNet's alignment-gap network (TargetNet). We will run it on the identical GTEx/TCGA splits with the same deconvolution inputs, so any gain is attributable to the model, not the data.
The strongest competitor is bayesReact, the only existing method that is simultaneously single-cell-aware, temporal, and microRNA-aware. It infers activity from expression rankings with a probabilistic generative model, and it demonstrated temporal tracking of miR-124-3p in spinal cord development; however, no reported result found on Lacuna for bayesReact's quantitative accuracy at predicting target repression in held-out tissues, so the head-to-head numbers will come from our own runs. To keep the comparison fair, we give bayesReact the same deconvolved abundances, the same held-out splits, the same compute, and the same tuning budget (its Laplace approximation mode, which gave it a 278-fold speed-up, plus full Hamiltonian Monte Carlo where feasible), and we evaluate both methods on predicting observed target-mRNA down-regulation, not on recovering activity patterns, which favors its own training objective.
Two further comparisons isolate the proposal's ingredients. One ablation swaps the reporter-derived $(\beta, K)$ pair for the naive seed-match prior bayesReact uses — same model, same data, only the physical calibration removed; if performance does not drop, the reporter panel is decorative and the idea fails its own decisive test. A second ablation collapses cell types and time points into bulk averages while keeping the reporter calibration intact, isolating the heterogeneity-and-time conditioning. Finally, following the geneRNIB finding that simple models often win, we run Pearson correlation and GRNBoost2 on the deconvolved per-cell-type profiles with equal tuning; beating the transfer model would falsify our claim that stoichiometric structure, not just better-resolved correlations, carries the signal. Success is defined concretely: the transfer model should predict observed log fold change of target mRNAs in held-out tissues and time points with materially higher Spearman correlation than every baseline, and should recover the negative-correlation fat tails of Liu et al. — 35 miRNA–mRNA pairs with $r < -0.9$ at $P < 5 \times 10^{-21}$ — while correctly flagging correlated but indirect pairs as non-targets.
Why not just…?
Why not just run single-cell RNA sequencing on the tissue directly? Because single-cell expression data measure abundance and correlation, not binding. Standard single-cell data are too sparse to read microRNA activity at all — this is precisely why bayesReact needed a dedicated motif model — and Liu et al. showed that the strongest endogenous correlations are mostly indirect ripples, which only seed-mutant reporter controls can separate from direct binding. Cost also matters: as POPPY notes, single-cell sequencing remains impractical for routine longitudinal clinical monitoring, so bulk-plus-deconvolution is the realistic scaling path. To be honest, this is not an either/or: our model consumes single-cell data wherever it exists, as in the four-time-point mouse embryonic stem cell course used by bayesReact, and treats it as the abundance source, not a rival.
Why not just take TargetScan predictions and correlate them with longitudinal bulk RNA-seq? Because both halves are known to be weak. The prediction half: TargetScan's canonical seed matching misses most functional non-canonical sites, with recall of 0.3945 on miRAW (TargetNet). The correlation half: PUMA found that tissue-specific microRNA regulation is not driven by differential microRNA expression — a microRNA with static expression across the 38 GTEx tissues can regulate different processes in each — so raw correlation cannot see it; and bulk time-course correlation conflates genuine kinetic change within a cell type with shifts in cell-type composition, a limitation the temporal-profiling literature flags repeatedly. Our head-to-head on identical data will settle whether the stoichiometric structure adds anything over this baseline, and we commit to reporting it either way.
Why not just deconvolve with CIBERSORTx or MuSiC and correlate per cell type? This is the closest serious alternative, and we treat it as a real threat. The evidence against it is that standard continuous deconvolution mishandles discrete count structure — Count Bridges outperformed CIBERSORTx, MuSiC, and STDeconvolve on recovering single-cell count distributions — and that even perfect deconvolution yields abundance tables, not repression: it leaves the mapping from "microRNA present, target present" to "target repressed" as an unstated correlation, with no role for dose sensitivity or competing-site load. The decisive experiment is cheap to describe: we run deconvolve-then-correlate and the transfer model on the same deconvolved abundances from the same tools. If the correlation version matches the transfer model, the reporter calibration and stoichiometric term add nothing and the simpler route wins; that is a finding we would accept and report.
Why not just run the luciferase assay in primary cells at physiological concentrations? This fixes stoichiometry for one cell type at one moment, and we include it as a control arm, not a substitute. It cannot see heterogeneity across the cell types in a tumor or tissue, cannot see change over time, and cannot see the systemic and microenvironmental defenses that neutralized pro-apoptotic microRNAs like miR-135a-5p and miR-141 in vivo in Wang et al.. The dose–response panel at least extracts the concentration dependence that single-point assays discard, which is exactly what lets the model extrapolate to endogenous abundance rather than to whatever concentration a pipette happens to deliver.
Proposed Experiments and Decisive Checks
Evaluation proceeds in three tiers, all on data in hand. Tier one is prediction on endogenous bulk: fit on GTEx tissues and TCGA cancers with cell types and time held out, predicting observed target down-regulation; compare against all baselines above. Tier two is the temporal test: on the four-time-point mouse embryonic stem cell differentiation course used by bayesReact, the model must predict the time course of target repression, not just its average — a direct answer to the gap between static endpoints and dynamic regulatory cascades. Tier three is replication and extension of Liu et al.: on modern TCGA glioblastoma samples with matched bulk miRNA and mRNA data, recover their fat-tail negative-correlation pairs, and check whether the model's direct-binding predictions — but not the indirect correlated pairs — survive in deconvolved cell-type-resolved profiles.
The decisive wet-lab check closes the loop. Before any measurement, we register predictions from the fitted model for a small panel of pairs: some where the model says luciferase-validated repression transfers to a given cell type because stoichiometry favors it, and some where it says the same validated pair fails because the cell type's competing-site load $C$ soaks up the microRNA. We then run dual-luciferase assays at physiological mimic concentrations in sorted sub-populations versus mixed cultures, with seed-mutant controls to keep the physical interpretation honest, and, for one tissue model, the zebrafish in vivo pooled transgenesis platform of this work, which enables pooled reporter screening inside a living organism. The proposal stands or falls on one number: does the model's predicted per-cell-type, per-time repression match the measured firefly/Renilla ratios substantially better than the seed-match-correlation baseline does on the same pairs? If yes, we have a calibrated bridge from the easiest validation experiment in molecular biology to the conditions under which its results actually matter.
References
- Detection of a MicroRNA Signal in an In Vivo Expression Set of mRNAs
- geneRNIB: a living benchmark for gene regulatory network inference
- bayesReact: expression-coupled regulatory motif analysis detects microRNA activity across cancers, tissues, and at the single-cell level
- TargetNet: functional microRNA target prediction with deep neural networks
- DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- PUMA: PANDA Using MicroRNA Associations
- Effects of miR-135a-5p and miR-141 on proliferation, invasion and apoptosis of colorectal cancer SW620 cells
- Library transgenesis in zebrafish through delayed site-specific mosaic integration for in vivo pooled screening of transgenes