Hundreds of cancer papers report that a circular RNA, a stable loop-shaped RNA, acts as a "sponge." In this account it soaks up a microRNA, a short RNA that represses genes, and so frees that microRNA's target gene to make more protein. Almost all of this evidence comes from bulk tissue or cell-line lysates, where thousands of cells are ground together and averaged. That averaging creates a specific trap. Suppose a circular RNA and a target gene are both simply higher in one cell type. Then any set of samples that differ in how much of that cell type they contain will show the two rising together. The pattern looks exactly like sponging even if, inside every individual cell, the circular RNA does nothing to the target.
We propose to measure how much of the published sponging evidence this composition artifact can explain. We then replace the bulk correlation test with a sharper one. Within a single cell type, does sponge abundance track the part of the target's protein level that its own messenger RNA (mRNA) does not explain? That leftover part, which we call the protein residual, is where a translational derepression signal should appear. It is also exactly what current RNA-to-protein translation models are built to absorb rather than expose.
Core Claim
Our claim is that ceRNA sponging (competitive endogenous RNA sponging, the circular-RNA-soaks-up-microRNA mechanism) should be tested as a within-cell-state dependence between sponge abundance and target protein residual. Under that test, a substantial share of bulk-derived circRNA→miRNA→target axes will turn out to be explained by cell composition. The statistical reason is the law of total covariance. Write $S$ for sponge abundance, $T$ for target abundance, and $k$ for the cell type of a cell. Across pseudobulk samples, the observed covariance splits into two parts:
$$\mathrm{Cov}(S,T) = \mathbb{E}_k\big[\mathrm{Cov}(S,T \mid k)\big] + \mathrm{Cov}_k\big(\mathbb{E}[S\mid k],\, \mathbb{E}[T\mid k]\big).$$
Only the first term reflects anything that could be sponging. The second term is pure cell-type co-specificity, and its contribution to between-sample bulk correlation grows with how much the cell-type proportions vary across samples. Bulk ceRNA studies cannot separate the two terms.
What is new is not another RNA-to-protein model. It is the use of such models as calibrated null predictors. The departure of measured protein from the protein predicted from mRNA within a cell type becomes the test statistic for a post-transcriptional mechanism.
Where This Sits
Four bodies of work motivate the idea. First, the ceRNA literature reports axes such as circASH1L sponging miR-515-5p to derepress RRM2 and CDCA7 in ovarian cancer (CircASH1L…), circSEPT9 sponging miR-637 to lift LIF repression (circSEPT9…), and circ_0061140 acting through miR-136/CBX2 (Circ_0061140…). All of these rest on bulk lysates and on mRNA or Western blot readouts. The research file itself flags that whether single-cell circRNA copy numbers can titrate enough microRNA is untested.
Second, heterogeneity work shows that averaging hides minority cell states. One example is the rare stem-like cells behind tamoxifen resistance, which bulk profiles miss (Lin et al.). Another is the liver cancer stem cells in which an m6A-modified circRNA acts (circRAPGEF1). These cases cut both ways. A real sponge restricted to a rare subset would be diluted in bulk, while a spurious one would be manufactured by composition.
Third, the discordance literature reports that mRNA correlates with protein only 40% to 60% of the time (Adhikari). That leaves room for microRNA-mediated translational control.
Fourth, and supplying the method, are the paired RNA–protein models. totalVI (totalVI) models CITE-seq RNA and surface protein jointly and separates ambient antibody background. CITE-seq is a method that measures RNA and a panel of antibody-tagged surface proteins in the same cell. scChord (scChord) uses a noise-absorbing variational autoencoder plus flow matching, and recovers bimodal protein distributions such as CD4 that mean regression smooths away. scTranslator (scTranslator) pre-trains a Transformer to map transcriptomes to more than 14,000 proteins.
These models aim to predict protein as well as possible from RNA. We instead need them to predict protein only from the target's own RNA and the cell's state, leaving out the sponge. Their errors then carry the signal. Deconvolution methods such as Count Bridges (Count Bridges) and POPPY (POPPY) constrain the design. They show that cell composition can be recovered from bulk data. They also show that this recovery works only on linear polyadenylated mRNA and depends on how complete the reference atlas is.
PTM crosstalk enters as a competing explanation rather than a co-equal topic. PTMs are post-translational modifications, chemical tags added to proteins after they are made. A protein can fall below what its mRNA predicts because it is ubiquitinated and degraded, not because a microRNA represses it. Somatic mutations at lysines that can carry both acetylation and ubiquitination are frequent in tumors (Bader et al.), so degradation-driven residuals are expected. We therefore use site predictions from COMPASS-PTM (COMPASS-PTM) and MeToken (MeToken) as covariates for degradation susceptibility. A sponging claim must survive conditioning on them.
Research Program
The work has three components. The first is a composition audit. We build pseudobulk samples from single-cell atlases by drawing cells under proportions sampled from a Dirichlet distribution, a standard distribution over mixing proportions. We tune its spread to match the composition variability seen in tumor cohorts. For each published axis we can name from the file, we use the circRNA's host gene as an explicitly crude proxy for the sponge, together with the target mRNA. We compute what fraction of the expected bulk correlation comes from the between-type term above. We also create null pairs by permuting expression within each cell type, which keeps type means but destroys within-type dependence. Testing these pairs gives the false-positive rate of bulk correlation tests. Finally, we plant synthetic derepression effects confined to subpopulations making up 1% to 30% of a type, which measures power.
The second component is a within-type protein residual. For each target protein on a CITE-seq panel, we fit a translator conditioned on the target's own mRNA and a cell-state embedding, with the sponge proxy and microRNA-target genes held out. We then express each cell's observed protein as a probability integral transform, the cell's quantile under the predicted distribution. Using quantiles rather than raw differences matters because scChord-style models predict multimodal distributions, and a simple subtraction would misread bimodal cells.
The third component is a dependence test. Within each cell type, we regress the target's residual on sponge abundance with composition fixed. A second, label-free test asks whether residuals of proteins that share predicted seed sites for one microRNA covary more than residuals of matched non-target proteins. Shared covariation of this kind would be the footprint of microRNA activity varying from cell to cell. The PTM covariates enter both regressions.
Baselines
The naive reference point is the standard ceRNA analysis: Pearson correlation between sponge and target mRNA across bulk or pseudobulk samples. We run it on the same pseudobulks with identical sample counts. On null pairs, our test must show a calibrated false-positive rate near nominal where the naive test is inflated. Only then does the audit have a point.
The strongest competitor for the composition question is proportion-adjusted bulk correlation. Here we estimate cell-type proportions from the bulk profile and test the partial correlation of sponge and target given those proportions. We run it three ways: with true proportions as an oracle, with Count Bridges deconvolution, and with POPPY. POPPY reports up to 73% better cross-cohort patient-representation integration by silhouette score, and up to a 714% increase in composition-similarity correlation over baselines (POPPY). For Count Bridges, no reported result was found on Lacuna for exact Wasserstein or energy values; the file states only that it outperforms CIBERSORTx and MuSiC.
Our within-type test beats this baseline only if it detects planted subpopulation-restricted effects at lower prevalence for the same false-positive rate. We also test the case where the reference atlas omits the relevant subtype. Our test is not expected to win against the oracle-proportion version on effects that span a whole type.
For the residual, the strongest translator is scChord. Its paper reports the lowest RMSE and highest Pearson correlation on GSE164378 against totalVI, sciPENN and scTranslator, but no reported numerical RMSE or PCC was found on Lacuna for it. scTranslator reports a 10.7% cosine-similarity gain over sciPENN and cTP-net, and similarity above 0.83 when gene and protein sets are not aligned (scTranslator). We train totalVI, scChord and a ridge regression on identical splits with equal tuning budgets. We judge them on whether their residuals are calibrated: quantiles should be uniform on held-out cells of the same type.
The key ablation replaces the translator residual with the raw log protein-to-mRNA ratio. This isolates whether the calibrated null model is what buys power. A second ablation drops the PTM covariates, to see how many apparent sponging hits are better explained by degradation-prone lysines.
Why not just…?
Why not just adjust bulk correlations for deconvolved cell proportions? This is the real threat. It is cheap and uses existing cohorts. With accurate proportions it removes the between-type term exactly for effects that span a whole cell type. It falls short in two ways. It cannot see dependence confined to an unannotated subpopulation, because POPPY and similar methods depend on reference completeness. And it still reads only mRNA, so it misses translational effects. The pseudobulk power curves settle this directly. If the oracle-adjusted test matches ours at every planted prevalence, we will report that the audit plus adjustment is sufficient and drop the single-cell residual from the recommended protocol.
Why not just test sponge–target mRNA correlation within cell types in ordinary single-cell RNA-seq? This drops the protein requirement. However, microRNAs repress translation as well as mRNA stability, and the reported 40–60% mRNA–protein correspondence (Adhikari) implies that an mRNA-only test misses part of the effect. Single-cell RNA dropout further weakens this test. We run it as a parallel arm. The comparison is how many planted translational effects, which change protein but not mRNA, each arm recovers. A further practical limit is that poly-A-captured single-cell data largely miss circular RNAs, a limitation shared by every arm here and discussed below.
Why not just use predicted protein from scTranslator as the readout? A protein predicted from the full transcriptome is a function of RNA. If the sponge proxy is among the inputs, any association is circular, and if it is not, the residual against prediction is the quantity we actually want. Predicted protein can therefore serve only as the null model, never as the readout.
Proposed Experiments and Decisive Checks
The composition audit uses GSE164378, which has more than 161,000 PBMCs with 228 proteins, and the 3CA atlas, which has more than 2.7 million cells from 1,631 samples across 15 cancer types. 3CA provides the realistic tumor cell-type means needed to compute composition-explained correlation for published axes such as circASH1L/RRM2 and circ0060467/GPX4 (circ0060467). Host-gene proxies are a known weakness, because circRNA and host expression can decouple. We will therefore report audit results as bounds and not as verdicts.
The residual and seed-coherence tests run on GSE164378, GSE100866 (about 8,600 cord blood cells, 13 proteins) and the murine SLN data (111 and 208 proteins). Surface panels contain few microRNA targets, so these tests are a feasibility and calibration study.
The decisive check is a targeted cell-line experiment, because no dataset in the file pairs single-cell circRNA with protein. In the ovarian cancer line used for the circASH1L claim, we would measure per cell the back-splice-junction signal of circASH1L by junction-specific in situ probes, RRM2 mRNA, and RRM2 protein. We would combine this with circRNA knockdown. The sponging hypothesis predicts three things: a within-population, dose-dependent coupling between circASH1L and the RRM2 protein residual; loss of that coupling on knockdown; and circRNA copy numbers within an order of magnitude of the microRNA's own target-site pool. A composition-artifact or degradation explanation predicts no within-population coupling, or coupling that disappears after conditioning on ubiquitination-site covariates. Either outcome is informative. The first gives the field a stoichiometrically grounded protocol for ceRNA claims, and the second shows how much of the existing evidence base needs revisiting.
References
- CircASH1L inhibits ferroptosis and enhances cisplatin resistance by sponging miR-515-5p to regulate cell cycle-related CDCA7/RRM2 in ovarian cancer cells
- The circRNA circSEPT9 mediated by E2F1 and EIF4A3 facilitates the carcinogenesis and development of triple-negative breast cancer
- Circ_0061140 knockdown inhibits tumorigenesis and improves PTX sensitivity by regulating miR-136/CBX2 axis in ovarian cancer
- Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance indicate a potential role of cells with stem-like properties
- m 6 A‐Modified circRAPGEF1 Interaction with IGF2BP3 Promotes Hepatocellular Carcinoma Progression via Reprogramming Aspartate Metabolism
- A high-stringency blueprint of the human proteome
- Joint probabilistic modeling of paired transcriptome and proteome measurements in single cells
- scChord: A Probabilistic Manifold Rectification Framework for RNA-to-Protein Translation
- A pre-trained large generative model for translating single-cell transcriptome to proteome
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Frequent mutations in acetylation and ubiquitination sites suggest novel driver mechanisms of cancer
- Learning the PTM Code through a Coarse-to-Fine, Mechanism-Aware Framework
- MeToken: Uniform Micro-environment Token Boosts Post-Translational Modification Prediction
- Circ0060467 sponges miR-6805 to promote hepatocellular carcinoma progression through regulating AIFM2 and GPX4 expression