A gene's messenger RNA level is a poor guide to how much of its protein a cell makes. Even careful bulk mass spectrometry finds that mRNA tracks protein only 40–60% of the time (A high-stringency blueprint of the human proteome). MicroRNAs are one well-known cause. These are short non-coding RNAs that bind a short "seed" match in a transcript's tail (the 3' untranslated region) and suppress translation or speed up decay. Each microRNA has hundreds of possible targets. So the repression it applies to any one target should depend on how many other target transcripts the cell is carrying: many competing targets spread the microRNA thin and let each target make more protein. This competition idea is the usual explanation for why the same microRNA looks oncogenic in one tissue and tumor-suppressive in another (miR-17/20 sensitization of breast cancer cells to chemotherapy-induced apoptosis requires Akt1; Aberrant Expression of miR-1301 in Human Cancer). It has almost always been argued from bulk cell lines, where each cell's own target pool is averaged away.
We propose a direct test using data that already exists. CITE-seq measures RNA and a panel of surface proteins in the same single cell. For a surface protein whose transcript carries a seed match for microRNA $m$, we ask whether the cell's protein-per-transcript rises with that same cell's summed abundance of other transcripts carrying $m$'s seed, within a fixed cell type. If competition is real at physiological levels, the association should be positive and specific to the true seed. It should vanish when seed assignments are shuffled. Its sign and size should be predictable for proteins the model never saw. It should also be hidden, or reversed, when cells are pooled into pseudobulk. ChIP-seq, which maps where proteins physically bind DNA, has a supporting role: it anchors which promoters drive each microRNA's primary transcript, so that host-gene RNA can serve as a checked proxy for microRNA supply.
Core Claim
Within a cell type, part of the protein-given-RNA discordance for seed-bearing surface proteins is predicted by a small number of per-microRNA parameters. These parameters describe microRNA supply and competing-target load, both computed from the same cell's transcriptome. Because the parameters belong to microRNAs and not to proteins, the model makes predictions for proteins it was not trained on. A generic RNA-to-protein predictor cannot do that without paired training data for those proteins. The claim is falsifiable. It fails if seed-shuffled controls explain residuals equally well, or if held-out-protein predictions are at chance.
Relation to Prior Work
Two generative models supply the first-stage method. totalVI embeds paired RNA and surface-protein counts in a shared latent space. It models protein as a two-component mixture so that antibody background can be separated from true signal. scChord predicts protein from RNA with a count-noise-absorbing autoencoder followed by optimal-transport flow matching. It recovers bimodal CD4 and heavy-tailed CD45RA distributions. Both models treat discordance as a statistical mapping and contain no regulator. scChord's authors report worse Correlation Matrix Distance as the protein panel grows and suggest pathway priors are needed. We do not compete with these models on raw imputation accuracy. We use a denoiser of this kind to define a background-corrected discordance residual, and then ask whether a mechanistic, seed-structured term explains that residual.
A second group motivates the question. The single-cell ceRNA atlases LnCeCell 2.0 and LnCeVar 2.0 (ceRNA meaning "competing endogenous RNA," transcripts that soak up microRNAs) catalog sponge networks at cell-type resolution, but only from transcript counts, with no protein readout. Target-level studies such as MiR-372-3p Functions as a Tumor Suppressor in Colon Cancer by Targeting MAP3K2 and MiR-361-3p regulates ERK1/2-induced EMT via DUSP2 mRNA degradation in pancreatic ductal adenocarcinoma show protein-level repression and lineage-specific target engagement in bulk. We supply the missing single-cell, protein-level test of the competition mechanism these studies invoke.
Two benchmarking efforts constrain the design. The NeurIPS 2021 multimodal challenge and Open Problems both found that shallow networks and linear models often match deep ones on noisy single-cell prediction. Our main outcome is therefore not "higher correlation than a big model." It is a specificity test against shuffled seeds and a transfer test to held-out proteins, where flexible predictors have no structural advantage. The network-biology approach of Linking Proteomic and Transcriptional Data through the Interactome and Epigenome connects ChIP-derived transcription-factor activity to bulk proteomics. We borrow its use of ChIP as wiring, not as a measurement, and move it to single cells.
Research Program
The first stage defines discordance. We fit totalVI within each dataset. For each cell $i$ and protein $p$, we take the residual $r_{ip}$: the background-corrected log protein minus the protein expected from that cell's denoised RNA for gene $p$. This expectation is fit separately within annotated cell types so that cell identity is not counted as discordance.
The second stage builds microRNA covariates from sequence and the transcriptome alone. Seed matches are counted directly in annotated 3' UTR sequences for immune-relevant microRNA families. For each cell we compute two quantities. The first is a competing-pool load $L_{im}$: the site-weighted sum of normalized abundances of all transcripts carrying $m$'s seed, excluding gene $p$. The second is a supply proxy $S_{im}$: expression of $m$'s host or primary transcript. ChIP-seq enters here. ENCODE-style transcription-factor and active-promoter ChIP peaks (ChIP-seq guidelines and practices of the ENCODE and modENCODE consortia) are used to fix which transcription start site drives each primary transcript. In the NeurIPS Multiome data, which pairs RNA with chromatin accessibility but has no protein, we check that host-transcript RNA tracks accessibility at those ChIP-supported promoters across cells. MicroRNAs whose supply proxy fails this check are dropped rather than modeled.
The third stage fits a titration-inspired model of the residual:
$$r_{ip} = \sum_{m} n_{pm}\,\beta_m\,\frac{L_{im}}{L_{im} + \kappa_m S_{im}} + \epsilon_{ip}.$$
Here $n_{pm}$ is the number of seed sites for $m$ in gene $p$'s 3' UTR, $\beta_m$ is the maximal derepression for microRNA $m$, and $\kappa_m$ sets the pool size at which competition bites. The ratio comes from a simple sequestration picture in which a fixed microRNA pool is shared across target sites. The prediction is that $\beta_m > 0$: residuals rise as the pool grows relative to supply. Only two parameters per microRNA are learned. They are shared across all proteins, which is what allows prediction for unseen proteins.
Baselines
The naive reference is a per-protein linear regression of log protein on the same-cell RNA of its own gene within cell type. This model implies zero residual structure. Any explained residual variance above it is the quantity of interest.
The strongest generic competitor is a shallow multilayer network, three to five hidden layers as in the NeurIPS challenge winners, predicting each protein from the full transcriptome plus cell type. Such a network could in principle learn sponge effects implicitly. It gets the same training cells, the same splits, and a tuning budget matched to our model by search trials.
We also run totalVI and, where code permits, scChord as full RNA-to-protein predictors. The file reports that scChord achieved the lowest RMSE and highest Pearson correlation on GSE100866 and GSE164378, but no reported result found on Lacuna for the numeric RMSE or PCC values. We will therefore rerun it ourselves on identical splits and not quote figures.
Three ablations isolate the key ingredient. The first reassigns seeds to random heptamers matched for 3' UTR position and GC content. The second keeps true seeds but permutes $L_{im}$ across cells within cell type, breaking the cell-level link. The third removes $S_{im}$. Beating the baselines means two things. On held-out proteins, the seed model must explain significantly more within-cell-type residual variance than the shuffled-seed model, across repeated shuffles. On proteins it was trained on, it must add explained variance on top of the shallow network's residuals. We do not require it to beat the shallow network on total protein accuracy.
Why not just…?
Why not just use a shallow network on the full transcriptome, which already absorbs any sponge effect? For in-sample accuracy, this is a real threat, and we expect it to match or beat us. The benchmark evidence (NeurIPS 2021 challenge) makes that likely. But a network trained per protein says nothing about a protein it has no paired data for, and its weights do not separate competition from cell-state covariation. The held-out-protein test and the residual-on-residual test settle whether mechanism adds anything. If the network's residuals contain no seed-specific structure, we will report that competition is not detectable at CITE-seq resolution.
Why not just call the correlation cell-state confounding, for example activation raising both 3' UTR-bearing transcripts and surface protein? This is the most serious alternative. The matched shuffled-seed control addresses it directly: generic activation would raise residuals equally for fake seeds. Fitting within cell type and adding the totalVI latent as a covariate further absorbs continuous state.
Why not just analyze pseudobulk or bulk proteomics, which is cheaper and less noisy? Across cell types, target pool, supply, and protein all differ together, so a bulk association mixes mechanism with composition. We will compute the same coefficient on cell-type pseudobulks as an explicit check. The masking hypothesis predicts it will be unstable or reversed relative to the within-cell estimate.
Proposed Experiments and Decisive Checks
Training and main tests use GSE164378 (about 161,000 PBMCs, 228 proteins). The mouse SLN-all dataset (111 and 208 protein panels) serves as a cross-species replication, since seed families are conserved. The NeurIPS CITE-seq bone marrow data provide a third tissue, and the NeurIPS Multiome data validate the ChIP-anchored supply proxy. GSE100866's 13 proteins are too few for held-out-protein tests and serve only as a sanity check.
Proteins are split into training and held-out sets by gene family, so that paralogs with shared 3' UTR features do not leak across the split. The primary metric is held-out-protein within-cell-type residual variance explained relative to the shuffled-seed distribution. Secondary metrics are sign agreement of $\beta_m$ between human and mouse, and the within-cell versus pseudobulk coefficient comparison.
The result is decisive if the true-seed model beats at least 95% of seed shuffles on held-out proteins in two of the three datasets, and the leading microRNA families show consistent positive $\beta_m$ across species. A null on all of these, given more than 100,000 cells, would be informative too. It would bound competing-target effects on surface-protein output as below what paired single-cell data can detect, and it would weaken ceRNA explanations of context-dependent microRNA function in immune cells.
References
- A high-stringency blueprint of the human proteome
- miR-17/20 sensitization of breast cancer cells to chemotherapy-induced apoptosis requires Akt1
- Aberrant Expression of miR-1301 in Human Cancer
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- LnCeCell 2.0: an updated resource for lncRNA-associated ceRNA networks and web tools based on single-cell and spatial transcriptomics sequencing data
- LnCeVar 2.0: an updated resource and web tools for genomic variations disrupting ceRNA networks from single-cell/spatial transcriptomics data
- MiR-372-3p Functions as a Tumor Suppressor in Colon Cancer by Targeting MAP3K2
- MiR-361-3p regulates ERK1/2-induced EMT via DUSP2 mRNA degradation in pancreatic ductal adenocarcinoma
- Multimodal single cell data integration challenge: results and lessons learned
- Defining and benchmarking open problems in single-cell analysis
- Linking Proteomic and Transcriptional Data through the Interactome and Epigenome Reveals a Map of Oncogene-induced Signaling
- ChIP-seq guidelines and practices of the ENCODE and modENCODE consortia