When a biologist wants to know whether a promoter or enhancer is actually firing, the standard tool is a reporter assay: the regulatory DNA of interest is placed upstream of a light-emitting enzyme (usually firefly luciferase), the construct is delivered into cells in a well, and the light emitted after lysis is measured relative to a constitutive control reporter (Renilla luciferase), so that transfection efficiency and cell death cancel out. What this ratiometric number cannot tell you is which cells in the well were doing the transcribing. In practice, reporter assays are run on mixed populations — tumor-immune co-cultures, primary tissue cultures, organoids — and the resulting light is a population average. This is the same averaging problem that motivated the entire single-cell genomics field: bulk RNA-seq blurred averages hid the cell types SARS-CoV-2 actually infects (COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets), the macrophage subsets driving hepatocellular carcinoma (Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression), and the rare stem-like cells behind tamoxifen resistance in breast cancer (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties). A low reporter signal in a mixed well is ambiguous in exactly the way bulk RNA-seq is ambiguous: it could mean uniform weak activation, or strong activation confined to a minority subpopulation.
Computational genomics has built sophisticated machinery for undoing this averaging for sequencing counts, but none of it touches reporter assays. Count Bridges enable Modeling and Deconvolving Transcriptomic Data generatively deconvolves bulk and spatial RNA-seq into integer single-cell counts, but only for sequencing readouts. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level maps bulk tumor transcriptomes into single-cell spaces, but only for endogenous transcript abundance. Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy does measure a live transcriptional reporter, but in a clonal Jurkat line where every cell is the same, so no deconvolution is needed or attempted. Nobody has built the inverse problem for the physical reporter readout itself. That is the gap this proposal fills.
Core Claim
I claim that mixed-well reporter readouts are mathematically identifiable mixtures of per-cell-type transcriptional activity, provided two ingredients are supplied jointly: a panel of reporter constructs whose activity profiles are discriminative across the cell types present, and a noise model that describes how molecular reporter activity becomes measured photons. Given both, an Expectation-Maximization inference can recover, from a small arrayed panel of wells, the cell-type fractions, the per-cell-type activity of each reporter construct, and — critically — whether a nominally single cell type contains an activated minority state. The deliverable is a method I will call ReporterBridge: a generative model of mixed-well dual-reporter output plus a design procedure for the reporter panel that makes the inverse problem well-posed.
The observation model is compact. In a well co-transfected with control reporter, let $n^{\mathrm{exp}}_r$ and $n^{\mathrm{ctrl}}_r$ be the detected photon counts from the experimental reporter $r$ and the Renilla control, $s_r$ an unknown reporter- and well-specific gain (enzyme expression per copy, substrate kinetics, plate-reader scaling), $\pi_k$ the fraction of cell type $k$, and $a_{rk} \ge 0$ the latent activity of reporter $r$ in cell type $k$:
$$y_r = \log\frac{n^{\mathrm{exp}}_r + 1}{n^{\mathrm{ctrl}}_r + 1}, \qquad n^{\mathrm{exp}}_r \sim \mathrm{Poisson}\Big(s_r \sum_{k=1}^{K} \pi_k\, a_{rk}\Big)$$
The ratiometric readout $y_r$ cancels well-level confounders exactly as in standard practice (Measuring Transcriptional Activity with Dual Luciferase Assays), but the residual reporter-specific gain $s_r$ is precisely the kind of platform-scaling nuisance that Statistical inference for cell type deconvolution shows must be modeled explicitly — treating it as known inflates false positives and breaks identifiability when scaling correlates with expression. Notably, the apparent tension between discrete count models and continuous optical output resolves cleanly at the detector: photon arrivals are integer Poisson events, so the count-based machinery of Count Bridges transfers, provided the enzyme-substrate transduction stage (which I model as the gain $s_r$ with saturation) is separated from detection noise. To capture subpopulation structure within a cell type, each $a_{rk}$ is itself modeled as a two-component mixture (resting/activated), which turns the rare-subpopulation question into a model-selection question answerable from the posterior.
Research Program
Three components will be built. First, the reporter panel. The panel must form an activity matrix whose columns (cell types) are not collinear — this is the identifiability condition, analogous to MEAD's structural conditions for deconvolution. I will design it with the CODA machinery from Machine-guided design of cell-type-targeting cis-regulatory elements, which already produced over 51,000 synthetic CREs whose activity in K562, HepG2, and SK-N-SH cells is empirically measured, and whose forward model Malinois predicts empirical MPRA activity with Pearson correlation $r = 0.88\text{--}0.89$ — accurate enough to optimize panels in silico before synthesis. That paper optimized CREs for on/off specificity across isolated cell lines; I will repurpose the same optimizer with a new objective (posterior conditioning of the mixture inverse problem, not just specificity margin), which is a genuinely different use of the tool. GET (A foundation model of transcription across human cell types) can seed candidate CREs for primary-cell settings where no MPRA data exist, since it predicts cell-type-specific activity from accessibility across 213 cell types.
Second, the inference engine: an EM scheme over $(a, \pi, s)$ with the Poisson likelihood above, retaining posterior uncertainty via slack-posterior sampling in the style of Count Bridges rather than point estimates, and sandwich-type confidence intervals for $\pi$ in the style of MEAD. MEAD's own finding that marker-gene filtering can degrade accuracy guides the panel-size choice: I will weight the full panel inversely to noise variance rather than hand-curating a few "specific" reporters.
Third, the validation platform. Ground truth for per-cell-type activity is the hard part, and I will get it two ways: calibration wells containing single sorted cell types (recovering $a_{rk}$ directly), and matched wells carrying a fluorescent version of the same reporter read out at single-cell resolution by live imaging, following the live-GFP readout format of Clarke et al.'s NF-$\kappa$B assay. An honest limitation I will confront rather than hide: transient episomal reporters decouple regulatory sequences from native chromatin context (Measuring Transcriptional Activity with Dual Luciferase Assays). Because ReporterBridge estimates $a_{rk}$ empirically from matched wells rather than assuming it from sequence models, chromatin-context distortion affects the absolute activity values but not the mixture decomposition; a lentivirally integrated panel will be built as a secondary check.
Baselines
The naive reference point is the standing practice: the raw well-level firefly/Renilla ratio interpreted as "the population's" transcriptional activity (Measuring Transcriptional Activity with Dual Luciferase Assays). This is what essentially every published dual-luciferase experiment reports, so every claim of this proposal must be stated as a delta against it — specifically, cases where the raw ratio is flat (no detectable change) while the deconvolved estimate identifies a strongly activated 5% subpopulation.
The strongest existing competitor is Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data), the state-of-the-art generative deconvolution method for integer molecular data, which has already outperformed CIBERSORTx and MuSiC in recovering cell-type proportions from bulk RNA-seq. No reported result found on Lacuna for the numerical Wasserstein-2 or Energy Score values of Count Bridges on its own benchmarks, so the comparison will be run head-to-head on identical simulated and wet-lab reporter data rather than by quoting published figures. I will adapt it by feeding discretized photon counts and running its EM with projection-guided sampling, giving it the same data, the same panel, the same compute budget, and the same tuning effort as ReporterBridge. Count Bridges will likely win when its noise assumptions hold; ReporterBridge should win when reporter-specific gains and enzyme saturation dominate — and the saturated-gain wells are exactly where naive count models mis-scale. Beating it means lower Wasserstein-2 distance between recovered and true activity-proportion joint distributions on the simulated benchmark, and lower mean absolute error on spike-in fractions, at matched panel size.
A third baseline is classical linear deconvolution (a CIBERSORTx/MuSiC-style regression applied to the reporter activity matrix instead of a transcriptome matrix), which isolates the value of the generative noise model. Fourth, an experimental baseline: FACS-sort the cell types, run conventional dual-luciferase on each purified fraction. Finally, the key ablation replaces the discriminatively designed panel with a panel of generic promoters matched in size, isolating the panel-design ingredient; on the file's evidence, this ablation should collapse performance toward the naive ratio, since panel conditioning, not estimator sophistication, is the binding constraint. As a prediction-only reference, I will also compare against sequence-model predictions of activity (GET; FIRE-style scores), for which Deep learning-assisted genome-wide characterization of massively parallel reporter assays reports that FIRE achieves AUROC 0.75 and AUPR 0.20 on functional-variant discrimination versus MpraNet's AUROC 0.85 and AUPR 0.43 — off-the-shelf scores clearly trail assay-driven inference.
Why not just…?
Why not just run scRNA-seq or multiome on the tissue? For static abundance questions this is a fair alternative, and where it suffices, this proposal loses. But Clarke et al. (Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy) showed that during T-cell activation many regulated signaling proteins show no significant mRNA change, because the regulation is post-translational and the fluxes are transient — a reporter is the only readout that sees them, and POPPY (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level) documents that scRNA-seq is too expensive and logistically complex for routine clinical workflows. The decisive experiment is Experiment 3 below: matched wells measured by both scRNA-seq and ReporterBridge, with the prediction that discordant cases (reporter active, transcript unchanged) concentrate exactly in contact-activated states.
Why not just FACS-sort first and run dual-luciferase on each fraction? This is the most serious threat, and for pure cell-line mixtures it will work as well as ReporterBridge — I will say so in the paper. It fails in three documented ways: enzymatic dissociation destroys the cell–cell contact that induces the transcription under study (the central finding of COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets and of Clarke et al.'s co-culture system); sort-induced stress directly perturbs transcription and reporter translation (Measuring Transcriptional Activity with Dual Luciferase Assays); and rare plastic states cannot be gated by known surface markers (A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers, Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties). Experiment 4 settles this empirically: where activity is contact-dependent, sorted-then-assayed readouts should diverge from in-mixture deconvolution, and the live-imaging ground truth adjudicates which is right.
Why not just apply CIBERSORTx or MuSiC to the reporter matrix? Because both were shown to fail even on their home turf: Count Bridges outperformed them on bulk RNA-seq proportion recovery, and MEAD (Statistical inference for cell type deconvolution) showed that methods treating estimated proportions as perfectly known produce invalid confidence intervals and inflated false-positive rates. The reporter setting adds the platform-gain nuisance $s_r$ that linear regression cannot absorb. The linear baseline stays in the comparison precisely to quantify this.
Why not just predict activity in silico with GET or off-the-shelf scores? GET predicts cell-type-specific expression well ($r = 0.94$ on unseen astrocytes) but is a static accessibility-based predictor with no notion of a physical well, gain, or photon noise — it cannot say what a mixed lysate read. I use GET and Malinois as priors for panel design, which is where their strength lies, not as substitutes for measurement; the FIRE-versus-MpraNet gap (AUROC 0.75 versus 0.85) is the file's direct evidence that precomputed scores miss what functional assay readouts capture.
Proposed Experiments and Decisive Checks
Experiment 1 (simulation with real ground truth). Build simulated mixed wells from the CODA empirical MPRA activities of the 51,000+ synthetic CREs in K562, HepG2, and SK-N-SH, with realistic plate-reader photon noise, enzyme saturation, and gain variation. Evaluate recovery of $(\pi, a)$ by Wasserstein-2 distance and Energy Score (following Count Bridges' evaluation protocol), mean absolute error on fractions, and recall of an activated 5% minority state. This is the same simulation-from-sorted-populations logic as the Kassandra database used by POPPY, transplanted to reporter readouts.
Experiment 2 (two- and three-line co-cultures). Physically mix the three cell lines at known fractions, run the arrayed reporter panel, and check recovery against single-type calibration wells and matched live single-cell fluorescent imaging down to 1% fractions. Success is pre-registered: mean absolute fraction error below 0.05 at three cell types, and correct identification of the activated cell type in 90% of spiked wells where the raw ratio shows no directional change.
Experiment 3 (rare resistant subpopulation). Spike tamoxifen-resistant breast cancer cells (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties) into sensitive cells at 2–10%, with a reporter on a resistance-associated regulatory program. The claim worth testing is that ReporterBridge detects the subpopulation-specific activation that bulk assays conflate with a uniform population shift — the exact failure Lin et al. documented for bulk profiling — while matched scRNA-seq either confirms (abundance also shifts) or reveals discordance (activity-only signal), both informative.
Experiment 4 (contact dependence). Replicate the Jurkat–A375 co-culture geometry of Clarke et al. with an NF-$\kappa$B reporter, and compare in-mixture deconvolved activity against dissociate-sort-assay. If the two agree, the FACS alternative wins on simplicity and I will report that; if they diverge where imaging shows contact-dependent activation, the case for in-situ deconvolution is made.
The decisive negative check is identifiability: I will characterize analytically when the reporter activity matrix becomes collinear (closely related subclones, per the subclonal-architecture concerns of A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer) and empirically map the failure boundary, so the method ships with an honest domain of validity rather than an implicit promise of universality.
References
- COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets
- Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Signalome-wide mapping of the NFκB pathway in T-cells reveals novel targets for immunotherapy
- Statistical inference for cell type deconvolution
- Machine-guided design of cell-type-targeting cis-regulatory elements
- A foundation model of transcription across human cell types
- Deep learning-assisted genome-wide characterization of massively parallel reporter assays
- A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer