Functional Counts: Reading Real Pathway Suppression from Fragmented Single-Cell Transcripts

When biologists suppress a signaling pathway, they usually do it with reagents that attack RNA — microRNAs (short regulatory RNAs that bind many messenger RNAs and either block their translation or cut them up), or engineered RNA-cleaving enzymes. The goal is a coordinated, multi-node knockdown: hit several genes in one cascade at once so the pathway switches off rather than leaking. The problem is that our main measurement tool, single-cell RNA sequencing, cannot tell a working messenger RNA from a corpse. When a transcript has been sliced into non-functional fragments, those fragments are still reverse-transcribed and counted as if the gene were still on. CaRPool-seq quantified how badly this distorts the readout: across combinatorial RNA-targeting perturbations, target surface proteins fell by a mean of 76.5%, while target mRNA counts in the same cells dropped by only about 15% — a five-fold understatement of the actual suppression (Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq).

The distortion compounds at every scale. Single-cell counts are >90% zeros from technical dropout, so any signal that survives is sparse and noisy. In patients, measurements are usually bulk averages over mixed cell types, which hides whether suppression is working in the malignant subclones that matter or only in bystander cells. And every suppression signature we learn in cell-culture screens is silently miscalibrated before we ever try to transfer it to living tissue, which is one unexamined reason the in vitro-to-in vivo translation gap is so persistent (DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction). This proposal builds a count model that fixes the first error at its source, then traces whether the fix changes what we conclude at the other three.

Core Claim

Every observed count for a target gene should be modeled as a mixture of two latent states: intact, translatable mRNA and non-functional cleavage or degradation fragments. The intact fraction follows the known biophysics of coordinated suppression — multi-site binding that saturates rather than scales linearly. Write the observed count $y_{gi}$ for gene $g$ in cell $i$ as the sum of intact and fragment counts,

$$y_{gi} = y^{+}_{gi} + y^{-}_{gi}, \qquad y^{+}_{gi} \sim \mathrm{Poisson}\!\big(\lambda_g \, s_i \,(1-\rho_{gi})\big),$$

where $\lambda_g$ is the baseline expression rate, $s_i$ is a cell-specific size factor, and $\rho_{gi} \in [0,1)$ is the probability that a molecule of gene $g$ in cell $i$ is being suppressed. The repression probability is not a free parameter: it follows the saturating dose–response established empirically for multi-site post-transcriptional regulation,

$$\rho_{gi} = \rho_{\max}(m)\,\frac{(u_i\, d_g)^{h}}{K_g + (u_i\, d_g)^{h}},$$

where $u_i$ is the latent activity of the regulator (or regulator cocktail $m$) in cell $i$, $d_g$ is the binding-site density on gene $g$ including clustered sites in coding regions and 3' UTRs, and $h$, $K_g$, and $\rho_{\max}(m)$ are anchored to the saturation curves measured on more than 14,000 designed 3' UTR sequences, which showed that repression rises non-linearly with site count and then saturates at a ceiling set by miRNA identity and intracellular concentration (Ilya Vainberg Slutskin et al.; Tannia Gracia et al.). The claim is that this "functional-count" model, calibrated once against paired protein readouts, lets RNA-only single-cell, bulk, and spatial data report protein-level pathway suppression — and that it changes downstream conclusions: how strong suppression really is, whether it acts uniformly across a tumor or is confined to subclones, and whether in vitro signatures survive transfer into patients.

This is distinct from the closest existing work. bayesReact infers latent miRNA activity from the rank positions of target motifs among expressed genes and is robust to dropout, but it assumes every counted molecule reflects a functional transcript, treats regulators independently, and uses a rigid seed-match model (bayesReact: expression-coupled regulatory motif analysis detects microRNA activity across cancers, tissues, and at the single-cell level); under active cleavage its activity estimates are biased toward zero precisely where suppression is strongest. Count Bridges supplies the discrete-integer generative machinery and an EM deconvolution of bulk into single-cell profiles, but it has no mechanism, no perturbation model, and no notion that a count can be debris (Count Bridges enable Modeling and Deconvolving Transcriptomic Data). CaRPool-seq discovered the fragment discordance but offered no corrective model. C3TL transfers perturbation effects across cellular contexts but assumes an additive linear shift on a manifold, which contradicts the documented cooperative saturation of multi-site repression, and it pseudo-bulks cells, erasing the subclonal structure we want to see (Causal Cellular Context Transfer Learning (C3TL): An Efficient Architecture for Prediction of Unseen Perturbation Effects).

Research Program

The first stage is calibration on data where protein truth exists. The CaRPool-seq AML screen profiled 26 regulators and 158 pairwise combinations across more than 31,000 cells with paired CITE-seq (an assay that counts surface proteins alongside RNA), giving exactly the anchored pairs needed to fit the fragment model and its decay parameters (Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq). The Mixscale screens add scale — 1,626 perturbations across six cell lines and five signaling contexts over 2.6 million cells, with continuous cell-level perturbation scores that solve the all-or-nothing knockdown-dilution problem (Systematic reconstruction of molecular pathway signatures using scalable single-cell perturbation screens) — and the scPerturb harmonized collection adds 44 datasets across modalities for held-out generalization (scPerturb: Harmonized Single-Cell Perturbation Data). Fitting follows the count-model tradition of avoiding Gaussian transformations of zero-inflated counts (A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data; Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-seq data), with the added guard that self-consistency constraints prevent the autoencoder-style hallucination of biological signal documented in unconstrained imputation (Leveraging data-driven self-consistency for high-fidelity gene expression recovery). For tractability at millions of cells, inference uses Laplace-style approximations rather than full sampling, following the >278-fold speedup bayesReact achieved that way.

The second stage is combinatorial logic. Coordinated suppression behaves like an AND-gate when distinct miRNAs converge on one transcript (Tannia Gracia et al.) and shows genuine cooperative synergy only at high site density, not additive add-up (Claudia Coronnello et al.). The saturating $\rho$ function encodes this, and the CaRPool-seq epistasis findings — KDM1A–HDAC3 synergy, EP300 dominance — provide protein-anchored ground truth to test whether the model recovers interaction structure that additive shift models miss.

The third stage deploys the model RNA-only where no protein readout exists: an EM deconvolution in the spirit of Count Bridges, but with the functional-count likelihood, applied to bulk TCGA tumors and spatial spots. The scientific question is Gap 2 made concrete: is post-transcriptional suppression of an oncogenic axis uniform across a tumor, or confined to the rare aggressive subclones that bulk averaging is known to hide — as with tamoxifen-selected stem-like cells misread as uniform population shifts (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) and pro-tumorigenic macrophage subsets invisible in bulk (Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression). Reference atlases come from the curated cancer single-cell collection (over 2.7 million cells across 15 cancer types) used by POPPY (Cross-modal transfer learning for mapping bulk transcriptomes at cellular level), with the caveat POPPY itself flags — states missing from the atlas cannot be attributed — handled by validating on single-cell patient cohorts directly rather than only on bulk.

The fourth stage tests translation directly: suppression signatures calibrated purely in vitro are applied to the COMBAT clinical single-cell COVID-19 atlas and 10x Visium colitis sections, where Mixscale already showed that causal perturbation signatures can reveal cell-type-restricted pathway failures that observational gene sets conflate (for example, IFNβ versus IFNγ). Our signatures carry an extra, testable piece of information: a predicted protein-level effect size, which the Mixscale-style transcriptional signatures do not.

Baselines

The naive reference point is what the field does now: read target-gene mRNA log-fold-change from the scRNA-seq count matrix and call that the knockdown. CaRPool-seq already measured how badly this fails — mean protein reduction of 76.5% against an mRNA drop of roughly 15% — so this baseline starts roughly five-fold biased on the calibration task, and we will quantify its correlation with held-out protein change to put a number on it.

The strongest existing competitor is bayesReact (bayesReact: expression-coupled regulatory motif analysis detects microRNA activity across cancers, tissues, and at the single-cell level), the only published method that infers regulator activity directly from sparse single-cell expression while tolerating dropout zeros. It will be run on the identical CaRPool-seq and Mixscale splits, with the same compute and tuning budget, using its published Laplace approximation for speed. No reported result found on Lacuna for bayesReact's accuracy against protein-level readouts, so the head-to-head is genuinely open; beating it means materially higher correlation with held-out CITE-seq protein reduction and correct recovery of synergistic versus dominant pairwise interactions that its independent-regulator design cannot express. For the bulk and spatial deployment, Count Bridges (Count Bridges enable Modeling and Deconvolving Transcriptomic Data) is the strongest deconvolution baseline — it outperformed CIBERSORTx and MuSiC on bulk deconvolution (no reported numerical value found on Lacuna for that margin) — and it will be run with its fragment-naive likelihood on the same TCGA and Visium inputs; beating it means recovering cell-type-specific suppression activities whose held-out protein or outcome correlates exceed what intact-everywhere counts support. For patient outcome prediction, the reported numbers to beat are contextual rather than identical-task: SubNetDL achieved a median AUROC of 0.74 across ten TCGA cancer–drug pairs (A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer) and DeepSADR reached AUC = 0.957 for Sorafenib on 555 TCGA samples against a next-best of 0.811 (DeepSADR); both operate on coarse pathways and bulk vectors, and neither models post-transcriptional suppression, so the fair comparison is whether per-subclone suppression activity adds predictive signal on top of their inputs.

The key ablation removes the latent intact/fragment split, collapsing the model to a standard single-state count model with the same saturation prior. If protein-anchored prediction does not degrade, the fragment mechanism — the proposal's core ingredient — is not earning its place, and the simpler model should win. A second ablation replaces the saturating $\rho$ function with an additive linear shift in the C3TL style to isolate the value of the biophysical nonlinearity on the combinatorial pairs.

Why not just…?

Why not just measure protein directly with CITE-seq? For new experiments, this is a real competitor, and it is why the calibration stage uses CITE-seq as ground truth rather than as a rival. But it cannot be the answer for the science we want: CITE-seq covers a limited panel of surface proteins, not arbitrary pathway nodes; it cannot be retrofitted to the 44 archived scPerturb datasets, the COMBAT patient atlas, or 10x Visium spatial sections; and the entire point is to make the RNA-only measurements we already have at scale interpretable in protein-equivalent units. If, however, the model's protein predictions fail on held-out CITE-seq targets outside its calibration panel, that failure would argue the fragment model is overfit to surface-marker physics, and we would report it as such.

Why not just pseudo-bulk the cells and skip all the single-cell machinery? This is the most serious alternative, because C3TL showed pseudo-bulking gives a stable, denoised signal that matches or beats massive foundation models at a fraction of the cost (C3TL). But averaging answers a different question. The subclonal-confinement stage of this proposal is unanswerable after pseudo-bulking: tamoxifen resistance looked like a uniform population shift in bulk while being selection of rare stem-like cells (Integrative analyses of gene expression and DNA methylation profiles in breast cancer cell line models of tamoxifen-resistance), and only single-cell perturbation signatures caught the T-cell-specific IFNβ failure in critically ill COVID-19 patients (Mixscale). Fragment bias also survives averaging, since debris inflates counts in every cell. The decisive experiment is direct: compare pseudo-bulked versus per-cell-type suppression estimates on the same patient cohorts and test which one correlates with outcome and with spatial niche identity.

Why not just use bayesReact and move on? bayesReact is included as the strongest baseline precisely because if it matches the functional-count model on protein-anchored prediction, the simpler method should win. We expect it to fall short for a structural reason: it scores motifs by their position among expressed genes, so fragments of a suppressed transcript — which remain counted — push the target's rank back up and attenuate the inferred activity exactly when suppression is strongest. Its independent-regulator and seed-only design also cannot express the AND-gate logic of converging miRNAs (Tannia Gracia et al.) or the coding-region clustered sites that rival 3' UTR targeting (Michael Schnall-Levin et al.). If the head-to-head shows otherwise on the combinatorial pairs, that result would itself be informative and we would concede the calibration stage.

Why not just apply a fixed correction factor to mRNA counts? A single scalar (say, divide observed knockdown by five) is tempting given the 76.5%-versus-15% figure, but the MPRA measurements rule out any constant factor: repression depends non-linearly on miRNA identity, intracellular concentration, and site density, and saturates at a ceiling (Ilya Vainberg Slutskin et al.). The discordance is therefore gene- and context-specific. The scalar correction is implemented as an ablation anyway; if it matches the full model on held-out pairs spanning low and high site densities, the saturation machinery is unnecessary and we will say so.

Proposed Experiments and Decisive Checks

Experiment 1 calibrates and tests protein prediction. Train on a random split of the CaRPool-seq regulator set and its 158 pairs; evaluate on held-out targets and pairs, predicting per-cell protein reduction from RNA counts and guide identity, scored against the measured CITE-seq values. The naive mRNA log-fold-change baseline is known to sit roughly five-fold off (76.5% versus ~15%); success is a substantial correlation improvement over both the naive baseline and bayesReact, with the fragment-state ablation showing the gain comes from the latent split.

Experiment 2 tests combinatorial recovery on the pairs: the saturating model should reproduce the documented KDM1A–HDAC3 synergy and EP300 dominance at the protein level, while the additive-shift ablation should fail specifically on high-site-density, convergent-regulator targets — a direct test of whether C3TL's linear manifold assumption is compatible with cooperative biophysics.

Experiment 3 tests sparsity limits. Following the scPerturb sampling design and its E-distance framework (a distributional distance between perturbed and control "cell clouds"), we downsample to the documented minimum of 300 cells per perturbation and 1,000 UMI counts per cell and below, checking whether functional-count suppression estimates stay calibrated where naive fold-changes degrade (scPerturb).

Experiment 4 localizes suppression in patients. Apply the EM-deconvolved functional-count model to TCGA bulk tumors and the cancer single-cell atlas, asking whether suppression of candidate oncogenic axes is uniform or subclone-confined, and whether subclone-specific suppression activity predicts outcomes beyond what subclonal mutation architecture provides (SubNetDL's median AUROC of 0.74 on TCGA pairs is the reference number; DeepSADR's Sorafenib AUC of 0.957 on 555 samples bounds what bulk-only, pathway-coarse transfer can already achieve).

Experiment 5 is the translation test and the decisive check of the whole program. Take suppression signatures calibrated only in vitro, apply them RNA-only to the COMBAT single-cell atlas and 10x Visium colitis sections, and ask two things: whether predicted protein-level suppression in specific cell types matches the known pathway failures (the IFNβ failure in critically ill T cells is the anchor case), and whether the fragment-aware signatures separate cascades that observational gene sets conflate, as Mixscale showed perturbation signatures can. If the in vitro-calibrated fragment model transfers, we have a calibrated ruler for coordinated suppression in living tissue; if it does not, the discordance itself becomes a measurable, mechanistic component of the translation gap rather than an unexplained failure — and either outcome is a result worth having.

References

  1. Efficient combinatorial targeting of RNA transcripts in single cells with Cas13 RNA Perturb-seq
  2. DeepSADR: Deep Transfer Learning with Subsequence Interaction and Adaptive Readout for Cancer Drug Response Prediction
  3. Unraveling the determinants of microRNA mediated regulation using a massively parallel reporter assay
  4. Urinary Exosomes Contain MicroRNAs Capable of Paracrine Modulation of Tubular Transporters in Kidney
  5. bayesReact: expression-coupled regulatory motif analysis detects microRNA activity across cancers, tissues, and at the single-cell level
  6. Count Bridges enable Modeling and Deconvolving Transcriptomic Data
  7. Causal Cellular Context Transfer Learning (C3TL): An Efficient Architecture for Prediction of Unseen Perturbation Effects
  8. Systematic reconstruction of molecular pathway signatures using scalable single-cell perturbation screens
  9. scPerturb: Harmonized Single-Cell Perturbation Data
  10. A fast and efficient count-based matrix factorization method for detecting cell types from single-cell RNAseq data
  11. Modeling expression ranks for noise-tolerant differential expression analysis of scRNA-seq data
  12. Leveraging data-driven self-consistency for high-fidelity gene expression recovery
  13. Novel Modeling of Combinatorial miRNA Targeting Identifies SNP with Potential Role in Bone Density
  14. 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
  15. Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression
  16. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
  17. A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
  18. Unusually effective microRNA targeting within repeat-rich coding regions of mammalian mRNAs