RNA immunoprecipitation (RIP) uses an antibody to pull a chosen RNA-binding protein out of a cell lysate, then sequences the RNAs that come with it. It is how biologists argue that a protein physically holds a given RNA, rather than affecting it through a chain of indirect steps. When RIP is run on real tissue, the lysate mixes many cell types, such as neurons with glia or tumor cells with immune and stromal cells. The readout then averages over all of them. Researchers usually summarize the resulting list of bound genes with functional enrichment analysis, a statistical test asking whether genes from a curated pathway appear in the list more often than chance.
This note argues that the two steps together produce a specific, predictable artifact. Genes expressed mainly in the cell types where the binding protein is abundant look strongly "bound." Enrichment then reports the marker pathways of those cell types as if they were the protein's targets. The problem gets worse when two conditions differ in cell composition, because a change in the number of immune cells becomes a "differential binding" result.
We propose to model tissue RIP counts as a mixture over cell types. The model uses a single-cell reference and variation in composition across samples to estimate how strongly the protein occupies each gene's transcripts within each cell type. Enrichment is then run separately for each cell type, against that cell type's own expressed genes. The practical output is a statement like "the protein binds this pathway's transcripts in hepatocytes, and the disease change sits in macrophages." Tissue RIP cannot currently support statements of that kind.
Core Claim
The standard tissue RIP pipeline computes a per-gene ratio of immunoprecipitated reads over input reads, thresholds it, and tests the target list against a genome-wide background. Its central problem is that the IP-over-input ratio for a gene is a weighted average of cell-type-specific occupancies. The weights are the share of that gene's transcripts contributed by each cell type. Genes whose transcripts come mostly from cell types that express the binding protein therefore inherit high ratios whatever their intrinsic binding behavior. Pathway-level summaries amplify this, because curated pathways are full of cell-type markers.
We claim that three things together turn tissue RIP into cell-type-resolved binding maps and cell-type-resolved enrichment:
- a count-level mixture model fit across samples of varying composition;
- a prior that suppresses occupancy in cell types that do not express the binding protein;
- per-cell-type enrichment against expressed-gene backgrounds.
We also claim these maps remove the composition-driven false pathways that current practice reports.
The motivation comes from the heterogeneity literature. Bulk profiles hid which cells SARS-CoV-2 actually infected in COVID-19 tissue atlases. They also hid disease-linked macrophage and exhausted T-cell subsets in hepatocellular carcinoma, which only single-cell data revealed in Global immune characterization of HBV/HCV-related hepatocellular carcinoma. Neither study concerns RNA–protein binding, but the same averaging logic applies directly.
Tissue RIP studies are where it bites:
- Brain. RIP-seq of Prdx1 in brain reports that the enzyme binds and stabilizes a network of anti-inflammatory mRNAs. From bulk brain lysate, such a finding cannot separate intrinsic binding preference from the fact that inflammatory-response transcripts are concentrated in microglia. We do not claim the finding is wrong, only that the current pipeline cannot test it.
- Other mixed tissues. The same ambiguity applies to CLIP maps of TDP-43 used to explain disease in TDP-43 loss induces cryptic polyadenylation in ALS/FTD, and to intestinal binding claims such as IGF2BP3-stabilized CLDN11 mRNA, whenever the input is a cell mixture.
Relation to Prior Work
Three deconvolution lines supply the machinery.
- Count Bridges treats bulk deconvolution as expectation-maximization over integer counts with Poisson dynamics. We reuse its integer-count view, because RIP libraries are sparse and a Gaussian approximation distorts low-count targets.
- SCDC weights several single-cell references by how well they reconstruct the bulk data. We adopt this to limit protocol bias in the reference.
- POPPY maps bulk transcriptomes into a single-cell foundation-model embedding using the Cell Ontology.
All three deconvolve expression. None of them models a second assay whose signal is expression multiplied by a cell-type-specific occupancy, which is the structure RIP introduces. None of them carries the result into enrichment testing.
On the RIP side, two tools address signal quality and sequence rules. omniCLIP calls binding sites with a Bayesian model of CLIP reads but treats the sample as one population. PaRPI and sequence-to-signal CLIP prediction predict binding from RNA and protein sequence across protocols and batches. They learn transferable sequence rules, but they have no notion of which cell type a transcript came from. We use PaRPI-style predictions only as a baseline and as an optional prior, not as the estimator.
The limitations the file lists for RIP explain why the gap exists. Profiling requires 10^6–10^7 cells per replicate, so most binding maps come from cell lines such as K562 and HepG2. Tissue applications average over tumor, immune and stromal cells.
The research file I had access to was truncated and contained no work that deconvolves RIP or CLIP signal by cell type. The claim of novelty is therefore bounded by that search.
Research Program
For gene $g$ in sample $s$, we model immunoprecipitated counts as
$$y_{gs} \sim \mathrm{Poisson}\Big(\lambda_s \sum_{c} p_{cs}\, x_{gc}\, \theta_{gc}^{(k_s)}\Big),$$
where the symbols are:
- $p_{cs}$ is the proportion of cell type $c$ in sample $s$, estimated from the matched input library by SCDC-style multi-reference deconvolution;
- $x_{gc}$ is the reference expression of gene $g$ in cell type $c$;
- $\theta_{gc}^{(k)}$ is the occupancy of gene $g$'s transcripts in cell type $c$ under condition $k$;
- $\lambda_s$ is a library-size factor.
The input library follows the same model with $\theta \equiv 1$. The input therefore anchors composition, and the IP library identifies occupancy.
Occupancy is identifiable per gene when the number of samples with distinct compositions exceeds the number of cell types. With fewer samples, we tie $\theta_{gc}$ to a shared gene effect plus sparse cell-type deviations. We place a shrinkage prior pulling $\theta_{gc}$ toward zero in cell types where the reference shows little expression of the binding protein itself. Fitting uses expectation-maximization over the latent split of each gene's counts among cell types, in the spirit of Count Bridges. Uncertainty comes from the posterior of that split.
Enrichment is then done per cell type:
- Rank genes expressed in cell type $c$ by posterior occupancy.
- Test pathways against a background restricted to those expressed genes.
- For differential binding between conditions, test $\theta^{(1)}_{gc}-\theta^{(0)}_{gc}$ rather than the raw ratio change.
The deliverables are the estimator, a mixture benchmark with known ground truth, and a diagnostic. The diagnostic scores any published tissue RIP enrichment result by how much of it is predicted from cell-type marker overlap alone.
Baselines
Naive reference point. This is the pipeline as practiced: per-gene IP over input fold enrichment, a significance threshold, and a hypergeometric test against a genome-wide background. For differential questions, it compares fold enrichment between conditions. It runs on exactly the same libraries. Beating it means recovering pure-cell-type target lists and pathways with higher precision, and reporting near-zero pathways under a composition-only null where it reports many.
Expressed-gene background. A stronger conventional practice restricts the enrichment background to genes expressed in the tissue, which removes abundance bias. It gets identical thresholds and target lists.
Strongest competitor: covariate-adjusted enrichment. This keeps the tissue-level target list. It weights the enrichment null by each gene's cell-type specificity scores, computed from the same single-cell reference, so that marker-heavy pathways are not favored by chance. It is the strongest competitor because it uses the same reference information we use, needs no mixture model, and directly targets the composition artifact at the pathway level. We give it the same reference, the same SCDC proportions, and the same tuning budget on held-out mixtures.
Sequence-based predictor. We run PaRPI predictions for the binding protein, weighted by cell-type expression, to produce "expected" cell-type binding. This tests whether sequence rules alone recover cell-type occupancy.
Ablations. Two ablations isolate our key ingredients. The first removes the binding-protein-expression prior. The second removes per-cell-type backgrounds while keeping the mixture model.
Reported numbers. I found no reported result on Lacuna for any of these baselines on cell-type-resolved RIP target recovery or composition-null enrichment false positives. The benchmark below establishes the first numbers.
Why not just…?
Why not just normalize IP by input, as everyone already does? Input normalization divides out a gene's total abundance in the lysate. It does not divide out which cells the transcripts came from. A gene expressed only in cell types lacking the binding protein will show a low ratio, and a marker of protein-rich cells will show a high one, even if intrinsic binding is identical. The composition-only null experiment below tests this directly. If the naive pipeline reports few spurious pathways there, the premise of this proposal fails and we would say so.
Why not just adjust the enrichment test for cell-type specificity? This is a real threat, and on the narrow question "is pathway X spuriously enriched?" it may match our method. We run it head to head. If it matches on tissue-level enrichment, we will report that as the recommended cheap fix. What it cannot produce is attribution: which cell type carries the binding, or in which cell type a disease-associated change occurs. We therefore evaluate attribution separately, and the proposal's claim rests there.
Why not just sort cells, or use single-cell binding assays? Sorting before RIP multiplies the 10^6–10^7-cell requirement by the number of populations. That is infeasible for most clinical samples and for rare subsets, and sorting stress may itself alter binding. Where sorting is feasible, for example with abundant blood populations, we use it as ground truth rather than as a competitor.
Proposed Experiments
In silico mixtures with ground truth. The first benchmark mixes reads from RIP or eCLIP libraries of the same binding protein profiled in different cell lines, such as K562 and HepG2 as named in the file. Mixing is done at known proportions, varied across 6–12 pseudo-samples, and applies to both IP and input libraries. Reads are subsampled to tissue-like depth. Ground truth is each line's own occupancy and its own pathway enrichment. Metrics are:
- area under the precision-recall curve for recovering each line's target list;
- rank correlation between estimated and true occupancy;
- precision of the top-ranked pathways against pure-line enrichment.
Composition-only null. The decisive check keeps per-cell-type binding identical between two conditions and changes only the proportions. The expected outcome is that the naive pipeline reports significant differential binding and enriched marker pathways of the expanding cell type, while our estimator reports essentially none. The second decisive check plants a binding change in only one cell line. It asks whether the method assigns that change to the right cell type, which covariate-adjusted enrichment cannot do by construction.
Physical cell-mixing experiment. A wet-lab version mixes two cell lines at known ratios before lysis. This adds real post-lysis reassociation and a realistic IP background, which the file flags as a limitation of native RIP. Reference robustness is tested by swapping single-cell references across protocols and measuring degradation, with and without SCDC-style weighting.
Published tissue studies. Finally, we apply the diagnostic to published tissue RIP enrichment results such as the Prdx1 brain study. For each reported pathway, we ask how much is predicted by marker overlap with the cell types that express the binding protein. We report which claims survive per-cell-type re-analysis, framing the results as testable re-analyses rather than refutations.
References
- COVID-19 tissue atlases reveal SARS-CoV-2 pathology and cellular targets
- Global immune characterization of HBV/HCV-related hepatocellular carcinoma identifies macrophage and T-cell subsets associated with disease progression
- Prdx1 Reduces Intracerebral Hemorrhage-Induced Brain Injury via Targeting Inflammation- and Apoptosis-Related mRNA Stability
- TDP-43 loss induces cryptic polyadenylation in ALS/FTD
- Disrupting of IGF2BP3-stabilized CLDN11 mRNA by TNF-α increases intestinal permeability in obesity-related severe acute pancreatitis
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- omniCLIP: probabilistic identification of protein-RNA interactions from CLIP-seq data
- PaRPI predicts RNA-Protein interactions from cross-protocol and cross-batch RNA-binding protein datasets
- Towards in silico CLIP-seq: predicting protein-RNA interaction via sequence-to-signal learning