Reading miRNA Activity Per Cell Type and Time From Spliced and Unspliced RNA

Tiny regulatory RNAs called microRNAs switch genes down after those genes are transcribed. They do this by binding the finished messenger RNA and marking it for destruction. Biologists usually say a microRNA "targets" a gene when two things are true: the two move in opposite directions across samples, and a lab test called a luciferase reporter assay shows the microRNA can silence a piece of the gene's sequence in a dish. Both kinds of evidence are weak in the same way. They cannot say whether the silencing actually happens inside a particular cell type at a particular moment. The expression data are usually bulk measurements averaged over many cell types. The reporter assay shows what the microRNA can do, not what it does at natural levels. And an opposite-direction trend often comes from indirect effects through transcription rather than from direct binding.

We propose a direct readout of post-transcriptional silencing that works inside ordinary single-cell RNA sequencing, even though that technology does not capture microRNAs themselves. Every cell's data contain two versions of each gene: unspliced reads, which still carry introns and reflect new transcription, and spliced reads, which reflect finished mRNA. A microRNA acting on mature mRNA lowers the spliced count and leaves the unspliced count alone. An indirect, transcriptional effect lowers both. We will measure this spliced-minus-unspliced shift for genes that carry a given microRNA's binding sequence, and compare it against matched genes that lack the sequence. Computed within each cell type, division state and real time point, this gives a map of where and when each microRNA is active. We then test that map in living cells with paired fluorescent reporters that can be read one cell at a time.

Core Claim

Post-transcriptional silencing leaves a signature that bulk averaging and transcriptional cascades both hide: mature mRNA falls without a matching fall in new transcription, and it falls specifically in genes carrying the microRNA's binding sequence. The claim is that this signature can be measured per cell state and per time point from existing single-cell time courses. The resulting activity map should predict endogenous silencing, as confirmed by live-cell reporters, better than correlation-based single-cell network methods. The key ingredient, subtracting the unspliced change, should account for a measurable share of that gain.

For microRNA $m$, cell group $c$ (a cell type crossed with a division state) and time $t$, let $\delta_g(c,t)$ be the change from a reference time in log spliced counts minus the change in log unspliced counts for gene $g$, computed on pooled pseudobulk counts. The activity score is

$$A_m(c,t) = \overline{\delta_g}_{\,g \in S_m} - \overline{\delta_g}_{\,g \in C_m},$$

where $S_m$ is the set of genes with conserved binding sites for $m$ in their 3′ untranslated region (the tail of the mRNA after the protein-coding part), and $C_m$ is a control set matched on untranslated-region length, AU content and expression level. A negative $A_m$ means the microRNA's targets lose mature mRNA faster than new transcription explains, beyond any global change in mRNA stability. Pooling hundreds of target genes and many cells is what makes noisy intronic counts usable. Significance comes from permuting gene labels within the matched strata.

Where This Sits

Three works motivate the gap. CSmiR infers cell-specific microRNA–mRNA networks from 19 co-sequenced K562 leukemia cells. It finds only 63–94% network similarity between cells of one line. Scan derives sample-specific networks from bulk and single-cell data using 27 inference methods. Both depend on measured microRNA levels, which rarely exist at single-cell scale, and on association statistics. Both are validated only against static databases. Splitting-averaging Bayesian networks show that regulation changes between epithelial and mesenchymal states, but they use bulk cell-line data split into two categories. Our method needs no microRNA measurement, uses real time points instead of binary states, and replaces association with a mechanistic contrast.

Two works constrain the design. The miR-4463 study shows that inverse expression often reflects indirect signaling, which is exactly the confound that unspliced subtraction targets. MBSTAR shows that many sequence-matched sites are inaccessible decoys. Our score is therefore an aggregate activity measure, and we do not claim any one gene is a real target from it alone.

The confound structure and the main application data come from CellDivider. That work sorted activated CD4+ T cells into six dye-dilution division states and showed that 23% of eQTL genes (genes whose expression depends on a nearby genetic variant) have division-dependent effects that are hidden when cells are grouped only by activation time. It did not look at microRNAs. We borrow its predictor of division state from the transcriptome so that we can condition on division.

The miR-874-3p/ATF3 study is the closest example of pairing single-cell data with reporter validation. Its reporter assay shows that the microRNA is capable of silencing ATF3 in a dish. It does not show that silencing occurs in the disc cell clusters, and the design is static. Our reporters are read in the target cell state across time.

Research Program

We will build the activity estimator on spliced and unspliced counts, using transcriptome-predicted division state to define groups. We will also build a bulk variant that splits bulk time-course counts into cell-type pseudobulks with Count Bridges, a generative model that keeps counts as integers, before applying the same contrast. The bulk variant answers whether clinical-style longitudinal bulk data can yield cell-type-resolved activity at all.

The validation layer uses dual-color reporters. One fluorescent protein carries a wild-type 3′ untranslated-region fragment of a predicted target. A second color carries the same fragment with the binding site mutated. The per-cell ratio of the two colors reads endogenous microRNA activity at natural microRNA levels. Combined with dye dilution and time sampling, it gives a live, cell-resolved, time-resolved answer to whether silencing occurs. This addresses the gap between what a microRNA can do and what it does, and it avoids the destructive sampling problem of sequencing.

Baselines

The naive reference point is correlation on bulk time courses: negative Pearson correlation between microRNA and mRNA across bulk time points, filtered by binding-sequence prediction. It is the most common practice. It gets the same samples and the same TargetScan prior.

The strongest existing method for cell-resolved inference is CSmiR-style cell-specific network inference with a TargetScan prior. Where microRNA levels are unavailable, this is replaced by Scan's best-performing distance-based method on single-cell data, per Scan. The reported reference is CSmiR's average validation rate against miRTarBase of 47.09% (CSmiR). No reported result was found on Lacuna for Scan as a single aggregate number. Because a database-overlap rate is not the same quantity as reporter-confirmed occurrence, we will rerun CSmiR on our data and score it with the same reporter and ground-truth metrics, rather than compare against the published figure. For sequence-only prediction, MiRGraph reports an AUROC of 0.8706 and an AUPRC of 0.8109 on miRTarBase. We include it as a context-free prior to show what any context-blind predictor cannot resolve: cell type and timing. All methods get the same cells, groupings, binding-site sets and a matched hyperparameter search budget.

The critical ablation is our estimator with the unspliced term removed. It then becomes a plain binding-site enrichment test on spliced changes, with the same controls. Beating the full baselines while failing to beat this ablation would mean the mechanistic contrast adds nothing. A second ablation drops conditioning on division state, which tests whether CellDivider's confound matters for microRNAs. Counting as beating the baselines means higher AUROC on the ground-truth benchmark described below, and a higher fraction of predicted (microRNA, cell state, time window) calls confirmed by reporter ratio, with bootstrap confidence intervals that do not overlap.

Why not just…?

Why not just correlate microRNAs with targets across a bulk time course? Bulk profiles mix composition changes with within-cell changes. CellDivider shows that expanding dividing cells hide a large share of regulatory effects in T cell activation. Indirect transcriptional effects also produce inverse correlations (miR-4463). The designed benchmark below measures this directly: the transfected population is a minority, so bulk correlation should be diluted and mistimed.

Why not just deconvolve the bulk data with CIBERSORTx or Count Bridges and then correlate? Deconvolution recovers composition, and Count Bridges reports outperforming CIBERSORTx and MuSiC. But deconvolved per-type profiles still cannot tell transcriptional from post-transcriptional change unless intronic signal is kept. We test this as our bulk variant, with and without unspliced counts. If deconvolution plus correlation matches our score, the unspliced term is not needed in the bulk setting, and we would say so.

Why not just run a binding-site enrichment test on spliced changes per cell type? This is the real threat. Binding-site enrichment already partly screens out indirect effects, and intronic counts in droplet data are noisy and partly artifactual. Our ablation settles it. Our expectation is that the gain concentrates in time windows where the targets are also changing transcriptionally, such as early activation, and is small elsewhere. The analysis will report gains stratified by the size of the unspliced change.

Why not just run luciferase assays on the top predictions? Those assays confirm capability in an overexpression setting, not occurrence in a specific cell state. That is why we read ratiometric reporters by division state and time.

Proposed Experiments

The first experiment is a benchmark with a known answer. Two fluorescently labeled cell lines are co-cultured. Only one line is given a microRNA mimic, or in a second arm an inhibitor, at a known time. Single-cell and bulk RNA are sampled at six time points. The truth is known: which cell type, which onset time, and a transfection-only reference set of responding genes. The metrics are AUROC for detecting activity in the right (cell type, time) cells, error in estimated onset time, and precision of the top-ranked genes against the reference responders. We vary the transfected fraction from 5% to 50% to quantify masking.

The second is a cross-check on the K562 co-sequencing data, if the raw reads allow counting of intronic reads. This asks whether inferred per-cell activity for abundant microRNAs anti-correlates with their measured levels. With 19 cells, it is only a sanity check.

The third is the main application, on the CellDivider activation data (86,760 training cells across four donors). It depends on raw reads being available for spliced and unspliced counting; otherwise we will generate a matched time course. We will map activity for expressed microRNA families, including the miR-17/92 cluster that CSmiR linked to cell-cycle modules, across division states and time. We will then test ten predicted windows of activity and ten predicted windows of inactivity in freshly activated T cells using dual-color reporters, and we will add antisense inhibitors to confirm that the reporter ratio depends on the microRNA.

Three outcomes would be decisive. The unspliced contrast must beat the enrichment ablation on the benchmark at low transfected fractions. Reporter ratios must separate predicted-active from predicted-inactive windows. Conditioning on division state must change at least some calls in ways the reporters confirm. If the ablation matches the full method, the paper becomes a negative result about intronic signal, with the reporter data still standing as a resource.