Is It Repression or Just Changing Cell Mix? Cell-Type-Resolved MicroRNA Target Inference in Time Courses

MicroRNAs are short RNAs that switch genes down after transcription by pairing with matching sequences on messenger RNAs. Because sequence matching alone predicts thousands of sites that do nothing in living cells, researchers confirm targets with expression data. The usual test is that when a microRNA rises, its targets should fall. Time courses look like the ideal place to run this test, since a microRNA ramping up during differentiation should drag its targets down. Our claim is that in most tissue and culture time courses this signal is badly contaminated. As cells differentiate, the mix of cell types shifts. A microRNA made by an expanding cell type will then anti-correlate over time with any gene made by a shrinking cell type, whether or not it ever touches that gene's RNA. Bulk sequencing averages over the mix, so it cannot tell the two cases apart. Time makes this worse, because composition drift is monotone and lines up with nearly every monotone gene trend.

We propose to measure how much of the temporal evidence for microRNA targeting is a composition artifact. We also propose an estimator that scores a microRNA–target pair only by evidence that survives inside a single cell type. Within one cell type at one time point, the predicted targets of an active microRNA should sit below carefully matched non-target genes, and that gap should widen as the microRNA's activity rises. A shift in cell mix cannot produce that pattern. In two sentences: bulk time-course anti-correlation largely reflects shifting cell types rather than microRNA repression. Scoring targets by within-cell-type depletion relative to matched controls, tracked over time, should recover validated targets with far fewer false positives.

What exists and what is missing

Two lines of work set up the problem without solving it. The first models regulatory networks that change over time but contain only transcription factors, the proteins that switch genes on by binding DNA. SimiC fits a fused-lasso network for each ordered phenotype or time point, with a penalty that keeps neighbouring networks similar. Its own authors' framing leaves microRNAs out entirely. VMPLN shows that clustering cells first and inferring networks second fails when cell types overlap, and it estimates cell identities and networks jointly instead. It is static and also has no microRNAs. We reuse VMPLN's lesson that composition and regulation must be estimated together, and SimiC's idea of smoothing regulatory effects across ordered time points.

The second line studies microRNA targets but ignores either cell mixing or time. CIDER builds causal Bayesian networks anchored on curated transcription-factor–microRNA and microRNA–target databases. It improves recovery on bulk NCI-60 and breast cancer data, but it assumes homogeneous populations, which is exactly the assumption we argue breaks in time courses. CSmiR and Scan infer cell- or sample-specific microRNA networks from paired single-cell microRNA and mRNA sequencing, but only on 19 K562 cells and with no time structure. Nowosiad et al. come closest to our setting. They measured microRNAs in single cells of fetal human cortex at two gestational windows and paired this with argonaute-2 CLIP, a biochemical assay that captures microRNA-bound RNA. Their CLIP was done on bulk tissue, however, and the analysis used co-expression modules rather than a target-scoring model.

Together these papers motivate the gap. No method scores microRNA targets in a way that is explicitly protected against composition drift over time. The sequence-based filters supply our prior: seed pairing, duplex stability and conservation, as used by Dong et al. and the multi-database consensus of Wei et al.. Those same papers also show that sequence alone over-predicts, which constrains us to treat sequence as a prior rather than as evidence.

Core Claim and Model

Standard single-cell RNA sequencing barely captures mature microRNAs, so we do not require microRNA measurements in single cells. Instead we infer microRNA activity the way transcription-factor activity is inferred from a factor's targets. metaTF and the regulon methods surveyed with it show that a regulator's activity is better read from its targets than from its own transcript. For a microRNA, the readout is coordinated depletion of seed-matched genes relative to control genes. The controls are matched on expression level, 3′ untranslated-region length and nucleotide composition, which are the known confounders of site counts.

For gene $g$ in cell type $k$ at time $t$, we model the mean count as

$$\log \mu_{gkt} = \alpha_{gk} + f_{gk}(t) - \sum_{m} s_{gm}\, \beta_{gm}\, A_{mkt},$$

with the following terms:

  • $\alpha_{gk}$ is a per-type baseline.
  • $f_{gk}(t)$ is a smooth per-type trend, which absorbs transcriptional change.
  • $s_{gm}$ is a sequence prior for microRNA $m$ on gene $g$, built from seed type, site count and conservation.
  • $\beta_{gm}\ge 0$ is the pair-specific repression we want to score, partially pooled across genes.
  • $A_{mkt}$ is the microRNA's activity in type $k$ at time $t$, smoothed across adjacent time points in the spirit of SimiC.

Identification comes from the contrast between targets and matched controls within the same $(k,t)$ cell, not from co-variation across samples. A pure composition shift changes the proportions but leaves this within-type contrast untouched. When only bulk data are available, the bulk count is modeled as $\sum_k p_{kt}\,\mu_{gkt}$, with proportions $p_{kt}$ estimated jointly against a single-cell reference. This is the same "sum to the observed bulk" constraint used by Count Bridges. Bulk small-RNA sequencing, where available, enters as a weak constraint that $\sum_k p_{kt} A_{mkt}$ tracks measured microRNA abundance.

Research Program

The work has three parts.

Audit. We take single-cell time courses, collapse them into pseudo-bulk samples, and run standard bulk temporal anti-correlation. We then ask what fraction of the top-ranked seed-matched pairs show any within-type depletion in the very same cells. This needs no external ground truth, because the same data are analyzed at two resolutions.

Semi-synthetic benchmark. We plant repression of known strength into a subset of seed-matched genes within chosen cell types of a real time course. We then vary how much composition drifts, and measure how every method's false-positive rate grows with drift.

Real-data validation. We score real time courses against curated validated interactions and against argonaute-2 CLIP sites. An optional extension adds the inferred microRNA activities as repressive regulators alongside transcription-factor activities in a time-indexed network. We then test whether they improve prediction of held-out later time points.

Baselines

Sequence-only prior. The naive reference is ranking pairs by $s_{gm}$ alone, combining seed, stability, conservation and multi-database consensus as in Wei et al.. It costs nothing and shares our prior, so any method that does not beat it is adding no information from expression.

Bulk temporal anti-correlation. The standard expression check is anti-correlation computed on pseudo-bulk or bulk samples. We run it with Scan's correlation and proportionality estimators (Scan), restricted to seed-matched pairs.

CIDER on the same bulk series. The strongest published expression-based competitor is CIDER, run on the same bulk time series with the same curated anchor edges. It is the strongest competitor because it already uses causal structure and curated knowledge to cut false positives. Its reported gain on TCGA breast cancer was nine validated miR-31 targets with transcription-factor–microRNA knowledge versus three with expression alone. The best single-cell microRNA network method, CSmiR, reported a 47.09% average validation rate against miRTarBase on 19 K562 cells (CSmiR). No reported result was found on Lacuna for microRNA target recovery in any time-course benchmark, so we will establish these baseline numbers ourselves.

Composition-adjusted regression. The baseline we take most seriously regresses bulk expression on the deconvolved proportions $p_{kt}$, then computes anti-correlation on the residuals.

Fair comparison and ablations. All methods see identical samples, the identical candidate pair set and the identical single-cell reference. Each gets the same budget for hyperparameter search, selected on a held-out set of planted pairs in the semi-synthetic data and on a held-out half of validated interactions in real data. Two ablations isolate the key ingredient. One removes the matched-control contrast, scoring by raw within-type target level against activity. The other removes the time smoothing on $A_{mkt}$.

What counts as winning. Beating the baselines means higher area under the precision–recall curve on planted pairs, with the gap widening as drift increases, together with higher CLIP-site and validated-target enrichment in the top-ranked pairs on real data. The comparison is made at matched list sizes.

Why not just…?

Why not just regress out deconvolved cell proportions and keep using correlation? This is the real threat, and we treat it as one. Adjustment removes the between-type artifact at the bulk level. It still estimates a single pooled effect, though, and therefore loses power when a microRNA acts in only one minor or expanding type, which is exactly the regime where developmental microRNAs matter. Its residuals also inherit any error in the deconvolution. The semi-synthetic benchmark settles the question directly by planting effects in one type and varying that type's abundance. If adjusted correlation matches our model on planted-pair recovery and on CLIP enrichment in real data, we will report that. The contribution would then shrink to the audit and a recommendation to always adjust, rather than a new estimator.

Why not just use single-cell microRNA–mRNA co-sequencing, as CSmiR does? It measures what we infer, but the only dataset in this literature is 19 cells from one cell line with no time axis (CSmiR). That data cannot support a time-course analysis across several cell types. Where small co-sequenced sets exist, we will use them as a check on our inferred activities rather than as the inference engine.

Why not just trust sequence and conservation filters? These filters encode where a microRNA could bind, not whether it represses in a given cell type at a given time. Their high false-positive rate is the motivating problem of the whole field. The sequence-only baseline quantifies exactly how much expression evidence adds beyond them.

Proposed Experiments

Semi-synthetic test. We use the human CD34+ blood stem and progenitor time course from the Open Problems longitudinal benchmark, which covers days 2, 3, 4, 7 and 10, where composition genuinely drifts as cells differentiate. We plant multiplicative repression of 10–30% into randomly chosen seed-matched genes within one cell type. We scale drift by resampling cells per day, then build pseudo-bulk and score all methods. The decisive check is the false-positive rate of bulk anti-correlation as drift grows, with real time-course drift as the reference point. Our prediction is that false positives grow sharply for bulk anti-correlation and only mildly for the within-type estimator. The prediction follows from the model: drift enters the bulk correlation through $p_{kt}$ but cancels in the within-$(k,t)$ target-versus-control contrast.

Audit on the same real data, no planting. We report what fraction of the top pseudo-bulk anti-correlated seed pairs show within-type depletion at all.

Developing human cortex. We use the Nowosiad et al. data (GSE107468), with a fetal cortical single-cell reference drawn from the atlas underlying GET. Two checks follow. First, the inferred per-type activities $A_{mkt}$ should reproduce that study's single-cell qPCR enrichment of microRNAs in specific cell types and gestational windows. That measurement is independent of our expression-based inference, which makes it a strong test. Second, top-ranked pairs should be enriched for their argonaute-2 CLIP sites relative to every baseline.

Validated interactions and a time-course secondary. We evaluate recovery of curated interactions from miRTarBase and TarBase in both systems, holding out half for tuning. As a secondary test, the pig preadipocyte series (Zhao et al.) checks whether bulk-only inference with a deconvolution model still beats CIDER when no single-cell data from that tissue exist.

Extension: microRNA activity in a temporal network. We add the inferred microRNA activities to a time-smoothed transcription-factor network and test whether held-out day-10 RNA prediction improves over transcription factors alone.

What would falsify the core claim. Two results would count against it. One is finding that real time-course drift produces few artifactual pairs in the audit. The other is composition-adjusted regression matching the within-type estimator on both planted recovery and the independent single-cell qPCR and CLIP checks.