Composition-Aware miRNA Target Discovery in Bulk Tumors

Many studies find the genes a microRNA (miRNA, a short RNA that silences messenger RNAs by pairing with them) regulates by looking across hundreds of patient tumors for mRNAs whose levels fall as the miRNA rises. The trouble is that a tumor biopsy is a mixture of cancer cells, fibroblasts, blood vessels and immune cells. If a miRNA made mostly by cancer cells is high in samples with many cancer cells, and an immune gene is low in those same samples, the two look anti-correlated even though no silencing happens. None of the widely used correlation, regression or causal target-prediction methods correct for this. The analogous problem in genetics was large: Geeleher et al. showed that eQTL models ignoring tumor purity had about a 46% false discovery rate in simulation, compared with about 3% for a purity-interaction model. They also found that only about 15% of conventional breast cancer eQTLs were specific to cancer cells.

We propose to carry that correction over to miRNA targeting and to extend it from a single cancer-versus-stroma split to many cell types. Cell-type fractions come from a single-cell RNA-sequencing (scRNA-seq) atlas used as a deconvolution reference. The model asks which cell type each silencing edge lives in. The harder problem is evaluation, and we address it with equal weight. Composition artifacts are not random noise. Epithelial miRNAs and mesenchymal genes really do have mutually exclusive expression, and some of those mesenchymal genes are true targets. A spurious composition-driven edge can therefore still "validate" against curated databases. Our evaluation is built so that it cannot be fooled this way.

Core Claim

Expression-based miRNA target prediction in bulk tumors is confounded by cell-type composition, and database overlap cannot detect this confounding. A per-cell-type interaction model with three properties removes most composition-driven edges while keeping true silencing, and it tells us which cell type each edge acts in. First, the model uses fractions estimated from an scRNA-seq reference. Second, it carries the uncertainty in those fractions forward. Third, it only tests a cell type for an edge if that cell type actually expresses the target. The necessary asymmetry is that standard droplet scRNA-seq captures only mRNAs with poly-A tails and is blind to mature miRNAs. So the mRNA side can be localized from the atlas, but the miRNA side must be localized statistically from bulk small-RNA data. Existing tools do neither. Single-cell co-sequencing, as used in CSmiR on 19 K562 cells, cannot yet reach clinical cohort scale.

Research Program

For a gene $g$ in tumor $s$, let $y_{gs}$ be log bulk expression, $m_s$ the log bulk level of the miRNA, and $\pi_{sk}$ the estimated fraction of cell type $k$. We fit

$$y_{gs} = \alpha_g + \sum_k \gamma_{gk}\,\pi_{sk} + \sum_{k \in E_g} \beta_{gk}\, m_s\,\pi_{sk} + \varepsilon_{gs},$$

The terms are defined as follows. $\gamma_{gk}$ absorbs how much of gene $g$ each cell type contributes. $\beta_{gk}$ is the silencing strength of the miRNA on $g$ within cell type $k$, where a negative value means repression. $E_g$ is the set of cell types in which $g$ is detectably expressed in the reference atlas. This is the Geeleher interaction structure, generalized from one purity term to roughly seven compartments: tumor epithelium, normal epithelium, fibroblasts, endothelium, T cells, B cells and myeloid cells. Restricting tests to $E_g$ removes most pointless tests, and those tests are the ones most prone to collinearity.

Fractions come from BayesPrism run with the breast cancer atlas of about 100,000 cells from 26 patients used in the deconvolution benchmark. That benchmark found BayesPrism and DWLS best at recovering fine-grained immune fractions. It also documented a "purity trap": above roughly 80–90% tumor purity, normal epithelium is mislabeled as cancer. To handle this, we refit the model over posterior draws of $\pi$ and combine the draws by multiple imputation. Edges whose sign flips across draws are down-weighted. Ridge penalties on $\beta$ keep the estimates stable where fractions are collinear. As a sensitivity analysis we also use SCDC, which combines several scRNA-seq references by minimizing bulk reconstruction error and so reduces the bias from depending on one platform.

Sequence information enters only at the final ranking stage. Candidate pairs are restricted to those with a predicted binding site, and the composition-corrected $\beta$ is combined with PRIMITI site scores. PRIMITI is a two-stage gradient-boosted model of binding sites and mRNAs that reports AUC up to 0.96 on CLIP and transfection test sets. Keeping sequence out of the core statistic lets us use seed enrichment as an independent check.

The application is TCGA breast cancer, which has matched mRNA and miRNA sequencing. This is the cohort used by hiddenICP-Pam50. That method treats the five PAM50 subtypes as "environments" for invariant causal prediction, but it treats each subtype as a pure population. The deconvolution benchmark shows that subtype systematically biases composition estimates, with triple-negative tumors over-predicting luminal progenitors. Our work accepts hiddenICP's goal of finding causal edges but changes what counts as the environment. As a secondary variant, we define environments by bins of continuous deconvolved composition instead of subtype labels.

Baselines

The naive reference point is bulk Pearson correlation between miRNA and mRNA across tumors, restricted to seed-matched candidates. The ensemble of Pearson, IDA and Lasso ranked by Borda count from Le et al. is the standard expression-based ensemble. On cell-line panels it beat every single method, with Wilcoxon p-values from 0.002 to 0.026. The strongest existing method for this exact cohort and question is hiddenICP-Pam50, which outperformed idaFast and jointIDA against transfection data and curated databases. The file gives no numeric precision for it: no reported result found on Lacuna for hiddenICP-Pam50 precision against miRTarBase or transfection data on TCGA BRCA. We will therefore rerun it ourselves on the identical sample set and the identical candidate universe. Context-MMIA multiplies negative-correlation scores by PubMed co-mention scores. We include it only in the database-overlap comparison, because literature scoring is circular with respect to curated databases. PRIMITI alone is the sequence-only reference.

We include three ablations. The first adds fractions as purely additive covariates, keeping $\gamma$ and removing $\beta$'s interaction; it tests whether the interaction matters. The second runs the full model on point-estimate fractions without posterior draws; it tests whether uncertainty propagation matters in the purity trap. The third drops the $E_g$ restriction. All methods receive the same samples, the same candidate pairs, the same fraction estimates where they use fractions, and a fixed grid of hyperparameters tuned only on simulated data.

Beating the baselines means two things. In simulation, our model must show lower false discovery at matched recall and correct cell-type attribution. On real data, it must show higher seed enrichment and higher transfection-validated precision among the top 1,000 edges. Database overlap alone does not count.

Why not just…?

Why not just use pure cell lines, such as NCI-60, where composition does not exist? Doing so discards exactly the edges that act in non-cancer cells. Geeleher et al. found that 15 of 24 breast cancer GWAS risk variants act through the microenvironment. CSmiR also found 6–37% rewiring even between individual cells of one line, so homogeneous cultures do not remove context dependence. Cell-line results remain a useful external check but do not answer where silencing happens in tissue.

Why not just add deconvolved fractions as additive covariates? This is a real threat for the narrow goal of removing spurious correlation, and it may remove much of it. It cannot, however, say which cell type carries the edge. It also leaves bias whenever silencing strength differs by compartment. The additive ablation sits side by side with the full model in the planted-edge simulation, and if it matches on false discovery we will report that honestly. The remaining contribution would then be cell-type attribution and the evaluation protocol.

Why not just deconvolve the mRNAs with BayesPrism first and then correlate the tumor-compartment expression with bulk miRNA? This two-stage approach is the most serious alternative. We run it as a baseline with all existing ensembles applied to the deconvolved matrices. Its weakness is structural rather than guaranteed. The miRNA side stays a bulk mixture because no single-cell miRNA reference exists. The deconvolved mRNA estimates are also shrunk toward the reference, which can create or cancel correlations with purity. If two-stage matches the joint model in simulation across the full purity range, we will adopt it as the recommended practice and concentrate on the benchmark.

Why not just trust sequence models with AUC near 0.96? Those AUCs measure binding on CLIP data or strong repression under transfection. They do not measure which sites are active in a given compartment of a given tumor, since binding does not guarantee silencing (Bader et al.). The sequence-only baseline is evaluated on the same planted and real checks.

Proposed Experiments and Decisive Checks

The first experiment is a semi-synthetic benchmark built from the breast cancer atlas, following the pseudo-bulk design of the deconvolution benchmark, with tumor purity spanning 5% to 95%. For chosen miRNA–gene pairs, we plant silencing of a known strength inside one named cell type. We also plant pure composition nulls, in which a simulated miRNA tracks one compartment's fraction but silences nothing. Bulk miRNA is generated as a fraction-weighted sum of cell-type levels. Metrics are false discovery rate at fixed recall, cell-type attribution accuracy, and both of these reported separately above 80% purity. This is the decisive test that database overlap cannot provide. A method that scores well on miRTarBase but calls the composition nulls fails.

The second experiment is on real TCGA data, where the cleanest check that avoids circularity is seed enrichment. We run the model without any sequence prior over all expressed genes. We then ask how strongly the top-ranked repressed genes are enriched for 7- and 8-mer 3′ UTR matches to that miRNA, compared with expression-matched background genes. True silencing should raise this enrichment. Removing composition artifacts should raise it too, because composition-driven anti-correlation carries no seed signal.

Two further checks follow. We measure precision against published miRNA transfection and knockdown profiles in breast lines, using the GEO resources hiddenICP used. We also test whether edges attributed to immune or stromal compartments involve targets and miRNAs plausibly expressed there, which requires the target to fall in $E_g$ and the miRNA's inferred compartment level to be non-negligible. Stability is measured by replication between random halves of the cohort and across PAM50 strata.

The main claim fails if, in both simulation and seed enrichment, the additive-covariate or two-stage baseline matches the joint model. It also fails if composition correction lowers seed enrichment on real data.