Reading microRNA competition in expression data to fine-map regulatory variants

When a genetic variant changes how strongly a microRNA binds one messenger RNA, the effect should not stay at that gene. MicroRNAs are short RNAs that dampen the expression of genes they bind, and each one is shared by hundreds of target genes. A variant that adds or removes a single binding site therefore changes how many copies of that microRNA remain free for every other target. This is an externality in the economist's sense: the outcome for one "match" depends on all the others. The weak, coordinated ripple it produces across the microRNA's other targets is a fingerprint. It says which microRNA is involved, and therefore which of many correlated candidate variants at a genetic locus actually alters a site for that microRNA.

We propose to compute this ripple from a competition model of microRNA–target binding and use it in two ways. First, we turn thousands of noisy distant-gene association tests into one test per locus and microRNA pair. Second, we feed the result into Bayesian fine-mapping as a data-derived prior over candidate variants. The target is the oldest problem in association genetics. Neighboring variants are inherited together (linkage disequilibrium), so a non-functional variant can look exactly as associated as the causal one next to it. Statistics on the target gene alone cannot separate them. Biology happening elsewhere in the transcriptome sometimes can.

Core Claim

For expression quantitative trait loci (eQTLs, genetic variants associated with a gene's expression level) whose causal variant acts through a microRNA site, the pattern of small distant ("trans") effects across that microRNA's other targets can be predicted from a binding-competition equilibrium. That pattern identifies the responsible microRNA with far fewer tests than gene-by-gene trans scanning. Upweighting candidate variants that disrupt sites for the identified microRNA should improve fine-mapping resolution over sequence-annotation priors alone. This should hold most strongly for non-canonical sites inside coding sequence that standard seed-matching misses.

The competition structure is the key ingredient. Write $F_m$ for the free concentration of microRNA $m$, $M_m$ for its total abundance, $T_k$ for the abundance of target transcript $k$, and $K_{ks}$ for the dissociation constant of site $s$ on transcript $k$. At equilibrium

$$ M_m = F_m + \sum_k T_k \sum_{s \in \text{sites}(k,m)} \frac{F_m}{F_m + K_{ks}} . $$

A variant that changes one $K_{cs}$ at a cis gene $c$ shifts $F_m$, which shifts the repression of every other target $k$. The predicted trans signature $\sigma_m \in \mathbb{R}^{G}$ over $G$ genes is the derivative of each target's bound fraction with respect to that perturbation. It is non-linear because it depends on site multiplicity and saturation, not just on target-set membership. This is the regime that Matching Games with Additive Externalities explicitly sets aside: that paper formalizes stability when each match receives fixed additive spillovers from the others, and leaves interacting, non-linear spillovers open. A microRNA pool split among competing targets under a capacity constraint is a many-to-one fractional matching with exactly such non-additive externalities. Its stable fractional assignment coincides with the mass-action equilibrium above. We use the matching view for its structure: which perturbations propagate, and how sensitivity scales with competitor load. We do not use it for game-theoretic solution concepts. Strategy-proofness concerns such as the impossibility result of Mechanism design and implementation theoretic perspective for interference coupled wireless systems have no analogue for molecules and play no role.

Research Program

The pipeline has three stages.

Stage 1: site catalogs. For each candidate variant $j$ in a locus, we record the set $\mathcal{M}_j$ of microRNAs whose sites it creates or destroys. We build this from two sources:

This paper motivates the proposal: it shows that seed-only catalogs miss much of the functional binding.

Stage 2: one test per locus and microRNA. For each lead eQTL locus and each $m \in \cup_j \mathcal{M}_j$, we estimate the lead variant's genome-wide trans effect vector. We then score its alignment with $\sigma_m$ using a weighted inner product. The equilibrium needs microRNA abundances, which our expression data may not measure. We therefore treat $M_m$ as a latent scale with a weak prior and integrate over it, so the signature's direction is learned from target-site counts and transcript abundances.

This stage replaces roughly $G$ trans tests per locus with $|\mathcal{M}|$ tests, typically tens. That shrinks the multiple testing burden at its source rather than through a more permissive correction.

Gene expression is strongly intra-correlated. Beware of counter-intuitive levels of false discoveries in datasets with strong intra-correlations showed across about 620,000 null datasets that Benjamini–Hochberg can flag up to 20% of RNA-seq features in individual null realizations. This paper constrains the design. All per-pair p-values are calibrated against genotype-permuted nulls that preserve expression correlation, never against analytic nulls. We report the realized false discovery proportion distribution, not only its mean.

Stage 3: microRNA-aware fine-mapping. The calibrated evidence $w_m$ becomes a prior inclusion weight $\pi_j \propto \exp(\sum_{m \in \mathcal{M}_j} w_m)$. This prior goes into Sum of Single Effects regression. That method, from A simple new approach to variable selection in regression, with application to genetic fine-mapping, writes the effect vector as a sum of single-variant effects and returns credible sets. We use the adaptive-shrinkage variant from SuSiE 2.0, which absorbs weak background polygenic effects; these two papers supply the method. Stage 2 and Stage 3 use disjoint halves of the samples, with roles swapped and results averaged, so the prior is not fit to the same noise it then rewards.

The prior is microRNA-specific and learned from data. SparsePro and BTS integrate functional annotations into fine-mapping, but their annotations are fixed tracks such as chromatin or binding marks. Our prior is instead re-estimated per locus from trans evidence. A variant destroying a miR-320a site gains weight only at loci whose distant footprint looks like miR-320a competition.

The structured eQTL lassos are the closest earlier attempt to borrow strength across outputs. A Two-Graph Guided Multi-task Lasso Approach for eQTL Mapping and Structured Input-Output Lasso penalize LD blocks and co-expression modules jointly. Both fix their output groups in advance from co-expression. We replace those groups with mechanism-derived signatures that carry sign and magnitude, and we keep the credible-set guarantees that penalized regression does not provide.

Baselines

Naive reference. Marginal single-variant eQTL testing with the lead variant reported as causal. This is the practice that Rong Chen et al. and de los Campos et al. show is inflated by LD tagging. It defines the floor for resolution.

Strongest existing method. SuSiE-ash with a uniform prior, run from the SuSiE 2.0 code on identical genotypes, expression, covariates and number of single effects $L$. SuSiE 2.0 reports a 1.5–3x reduction in false discovery rate over standard SuSiE in complex polygenic simulations. Beating it means one of two things:

  • higher power at matched empirical false discovery rate;
  • smaller credible sets at matched coverage on simulated microRNA-mediated loci, with no loss of coverage on loci without a microRNA mechanism.

We also run KnockoffZoom as a frequentist competitor with model-free false discovery rate control. Its height result of 3,284 UK Biobank discoveries is not on our task. No reported result was found on Lacuna for KnockoffZoom applied to eQTL fine-mapping, so we generate our own numbers.

Annotation baseline. SuSiE-ash with a SparsePro-style static prior that upweights any variant in any predicted microRNA site, using the same site catalog. No reported result was found on Lacuna for SparsePro on microRNA-site annotations. This baseline gets identical sequence information and differs only in lacking trans evidence.

Ablations isolating the key ingredient. The first replaces $\sigma_m$ with the unweighted target-set indicator, i.e. a mean shift across targets. The second replaces it with a first-order additive-externality approximation in the spirit of Brânzei et al., ignoring saturation and site multiplicity. The third drops the crosslinking-derived non-canonical sites. All variants share the same sample split, permutation budget and $L$.

Why not just…?

Why not just use a static microRNA-site annotation prior? Sequence alone cannot say which of a dozen microRNAs with sites in a locus is active in the tissue. A static prior therefore upweights every site-disrupting variant equally, including proxies that happen to hit an irrelevant site. The annotation baseline above tests this directly. If it matches our method on credible-set size, the trans evidence adds nothing and we would say so.

Why not just aggregate trans effects over each microRNA's target set, with no equilibrium model? This is the real threat. Target-set enrichment is simpler and may capture most of the signal if saturation effects are small at physiological stoichiometry. The first ablation settles it. If the equilibrium signature does not beat the unweighted set score in power for microRNA identification at calibrated error, the competition model is ornamental. We would then publish the simpler set-score prior as the method and report the matching view as an unnecessary detour.

Why not just rely on SuSiE-ash or KnockoffZoom, which already handle LD? Within a tight haplotype block, the cis gene's expression contains no information that separates perfectly correlated variants. Both methods then correctly report wide credible sets or coarse-resolution discoveries. The only way to narrow the set is to add external information, and our claim concerns which external information helps. On loci without microRNA mechanisms our prior should be near-flat and recover SuSiE-ash. We check this explicitly to make sure we lose nothing there.

Proposed Experiments

Simulation. We use UK Biobank and GTEx genotypes with expression simulated by simxQTL from SuSiE 2.0. We plant causal variants that perturb a microRNA site and propagate trans effects through the equilibrium at realistic effect sizes, deliberately including very small ones. Background polygenic effects and null loci are mixed in. Metrics are:

  • power at empirical false discovery rate 0.05;
  • credible-set size and coverage;
  • posterior inclusion probability calibration;
  • the 95th and 99th percentiles of realized false discovery proportion across replicates, following Kanduri et al.

Null calibration. We run plasmode nulls built by permuting genotypes against real GTEx expression. Our per-pair test must hold nominal error in the tails, not just on average.

Real data. We fine-map GTEx eQTLs and, as a smaller check, the 1000 Genomes Yoruba chromosome 20 expression data. Validation comes from external evidence the method never sees. The main check is enrichment of high-inclusion-probability variants in Argonaute hybrid sites from Fields et al., reported separately for coding-sequence sites. We also test agreement with the destabilization direction predicted by the reporter data of A massively parallel 3′ UTR reporter assay, where GC content correlated with mRNA level at r = −0.639. Those reporters were measured in mouse T cells, so we treat this only as a directional consistency check.

Decisive checks. The first question is detectability. In GTEx sample sizes, can any locus–microRNA pair pass permutation-calibrated significance? If the realistic trans effects of single-site variants are below detection, the proposal fails cleanly, and that is itself a useful negative result about competition effects. The second question is whether the equilibrium signature beats the unweighted target-set score in the ablation. The third is whether real-data credible sets shrink without losing enrichment for crosslinked sites relative to SuSiE-ash.