Reading Cell-to-Cell Differences in RNA Editing from the Molecules of a Bulk Sample

Plant mitochondria and chloroplasts cannot make working proteins from their genes as written in DNA. Hundreds of individual letters in their transcripts must be chemically changed from C to U. This change, called C-to-U RNA editing, creates missing start codons, creates stop codons, and restores conserved amino acids. When a bulk RNA sample, meaning RNA pooled from millions of cells, reports that a start-codon site in nad4L is 30% edited, that number has two very different readings. Every cell might carry a modest share of edited transcripts, or a minority of cells might be fully edited while the rest cannot make the protein at all. Bulk sequencing cannot tell these apart. Single-cell sequencing cannot either, in practice: resolving the difference needs at least 20 reads per site per cell, and droplet single-cell methods largely miss organellar transcripts, which lack standard poly(A) tails.

We propose to recover this hidden structure from bulk data by reading each RNA molecule as a small record of its own history. A long sequencing read that spans many editing sites records which of them were edited together on that one molecule. If editing is spread evenly across cells, edits at different sites on the same molecule should be close to statistically independent. If editing is concentrated in a subpopulation, edits cluster: molecules tend to be edited everywhere or nowhere. That clustering is visible in a bulk long-read sample without isolating any cell. We then assign the clusters of molecules to cell types by tracking how their abundance rises and falls across samples whose cell-type makeup differs.

Core Claim

Co-occurrence of edits within single molecules, combined with variation in cell-type composition across samples, makes the split between distributed and concentrated editing identifiable from bulk long-read RNA sequencing. No existing method estimates this split without deeply sequenced single cells. The idea is concrete. Model each bulk sample's molecules as a mixture of a few latent molecular classes, each with its own editing profile across sites. Test whether more than one class is needed after accounting for known technical and processing sources of correlation. Then link class weights to cell-type proportions estimated from nuclear gene expression.

The closest prior work is RNA editing generates cellular subsets with diverse sequence within populations. That paper defines the distinction between distributed and concentrated editing for mammalian APOBEC1 targets and fits a two-layer model: a binomial count of edited reads within each cell, with per-cell editing rates drawn from a population-level beta distribution. It showed that Cybb, 23% edited in bulk, was 61.5% edited in one macrophage and 0% in 11 others, while Anxa5 was uniform. It motivates our question and constrains our design. Its power analysis requires 20 or more reads per site per cell, and it needs single-cell data alongside bulk data. We replace "reads per cell" with "sites per molecule" as the source of statistical power. That removes the need for single cells at inference time, though not at validation.

Research Program

The model works at the level of reads. Read $r$ from sample $j$ covers a set of editing sites $S_r$ and shows edited-or-not calls $y_{rs}\in\{0,1\}$ at each covered site $s$. Each read belongs to a latent class $z_r\in\{1,\dots,K\}$ with sample-specific weight $\pi_{jk}$. Within a class, sites are edited independently with class-specific probability $\theta_{ks}$:

$$P(y_r)=\sum_{k=1}^{K}\pi_{jk}\prod_{s\in S_r}\theta_{ks}^{\,y_{rs}}(1-\theta_{ks})^{1-y_{rs}}.$$

With $K=1$, the model is the distributed-editing null: every molecule faces the same per-site probabilities, and bulk editing fractions tell the whole story. Latent class models of this kind are identifiable once each molecule covers three or more conditionally independent sites. Plant organellar genes supply many more; nad5 in Abies beshanzuensis carries 72 predicted sites (Liu et al.). This is why long reads, and not short reads, are the right measurement.

To assign classes to cell types, we write each sample's class weights as $\pi_{jk}=\sum_{t} w_{jt}A_{tk}$. Here $w_{jt}$ is the proportion of cell type $t$ in sample $j$, estimated from nuclear transcripts in the same library, and $A_{tk}$ is the fraction of cell type $t$'s organellar molecules that fall in class $k$. This is shared across samples. A cell type whose $A_{t\cdot}$ puts weight on both a fully-edited class and a poorly-edited class is internally heterogeneous. A type whose weight sits on one intermediate class edits in a distributed way.

The proportions $w_{jt}$ come from existing reference-based deconvolution. For this step we reuse SCDC (Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References), which combines several single-cell references to reduce batch effects. Nuclear transcripts are poly(A)-tailed, so standard plant single-cell atlases can serve as references. The model is fit by expectation-maximization, with $K$ chosen by held-out likelihood. Uncertainty in $w_{jt}$ is propagated by refitting under bootstrap draws.

The difficult part is not fitting the model but ruling out non-cellular reasons for molecules to be edited all-or-nothing. We name four and control for each in the design.

The first is contaminating genomic DNA. It produces reads that look completely unedited and would masquerade as an unedited subpopulation. We control for it with DNase treatment, a matched genomic DNA library, and a check that the unedited class does not carry DNA-only features.

The second is immature transcripts. Unspliced or unprocessed precursors are often less edited, a genuine molecular subpopulation that has nothing to do with cell identity. We control for this by stratifying reads on splice junctions and processed ends.

The third is DNA copies of organellar sequence, including chloroplast-derived fragments inside mitochondrial genomes, which are extensive in Melia azedarach (Hao et al.). These can produce misplaced reads with correlated "edits." We control for this through unique-mapping filters built against both organellar genomes.

The fourth cannot be removed. Heterogeneity among organelles within one cell gives the same molecular signature as heterogeneity among cells. We report the molecular-class result as "molecular subpopulations." The cross-sample link to $w_{jt}$ is the part that justifies a claim about cells, and single-cell validation settles the remainder.

Baselines

The naive reference point is the standard bulk editome pipeline: strand-aware alignment with HISAT2 and site calling with REDItools 2.0, as used for the Glycyrrhiza uralensis plastid (Ma et al.). It reports one editing fraction per site. Implicitly it assumes distributed editing, so its prediction for the fraction of fully-edited molecules or cells is fixed by the bulk rate. We run it on the same reads, and any structural claim must beat this implied prediction against single-cell ground truth.

The strongest competitor that uses bulk data only is composition regression. First estimate cell-type proportions per sample with SCDC. Then regress each site's bulk editing fraction on those proportions across samples to obtain a mean editing rate per cell type. This is the natural extension of expression deconvolution to editing. It exploits the same variation in composition that our model does, so it isolates exactly what reading whole molecules adds. It recovers cell-type means but by construction cannot see variation within a cell type. It also degrades when one lineage dominates, as the breast-cancer benchmark of 38,000 simulated mixtures showed for closely related lineages (Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures).

To keep the comparison fair, both methods use the same samples, the same SCDC proportions, and the same site set. Hyperparameters are tuned on the same held-out simulated mixtures. For proportion estimation itself, SCDC reports Pearson $r=0.991$ against the true 6:3:1 ratio of a three-cell-line mixture. That is the accuracy ceiling our cell-type assignment inherits. For recovering per-cell-type or per-cell editing distributions from bulk mixtures, we found no reported result on Lacuna. The comparison is therefore defined on our own ground truth. Beating the baseline means two things. Where editing is homogeneous within types, our per-cell-type mean editing error must be equal or lower. Where single-cell ground truth shows within-type bimodality, we must correctly flag it, and the regression baseline cannot.

The key ablation removes the within-molecule information. We fit the same model after shuffling each site's calls across reads within a sample, which preserves every per-site bulk fraction but destroys linkage between sites. If performance survives the shuffle, the linkage idea is doing nothing. A second ablation truncates reads to cover one or two sites, which shows how much read length the method actually needs.

We also report the single-cell beta-binomial model of RNA editing generates cellular subsets with diverse sequence within populations fit to deep targeted single-cell data. It serves as the ground-truth estimator, not as a competitor.

Why not just…?

Why not just sequence single cells and count edited reads per cell? Standard droplet platforms capture transcripts through their poly(A) tails. Organellar transcripts map poorly under poly(A) selection (Ma et al.), and droplet coverage falls far below the 20 reads per site per cell that the macrophage study found necessary. Targeted plate-based amplification of a few organellar amplicons per cell can reach that depth, and it is a real alternative for a handful of sites. We use exactly that as our ground truth. Its costs are that it scales with cells times sites, needs protoplasting (which may itself perturb organelles), and cannot be applied to existing bulk collections across tissues and conditions. If our method matched targeted single-cell results only at very high long-read depth, making the bulk approach no cheaper, we would say so. The depth sweep in the experiments measures this directly.

Why not just deconvolve nuclear expression and regress editing on cell-type proportions? This is the composition-regression baseline above, and it is the serious threat. If plant organellar editing turns out to be homogeneous within cell types and differ only between them, regression will match us at lower cost. That would be a useful negative finding in its own right. Our claim holds only where the within-type flag is confirmed by single-cell amplicons. The shuffle ablation shows whether any gain comes from linkage between sites rather than extra modeling flexibility.

Why not just sort cell types physically and sequence each in bulk? Sorting into mesophyll, epidermal and vascular fractions answers the question between cell types directly, and we use it as validation. It still returns an average within each sorted fraction, so it is blind to concentrated editing inside a type, the case where Cybb showed its most striking behaviour. It also requires tissue dissociation for every condition studied.

Why not just predict editing from sequence with Deepred-mt? Sequence predictors, as used to call 827 sites in Abies beshanzuensis (Liu et al.), estimate whether a site can be edited from the static genome. Their output is the same for every cell, so they cannot speak to distribution across cells. We use such predictions only to choose which sites to model.

Proposed Experiments

The first experiment is simulation. We generate organellar molecules from known mixtures of distributed and concentrated editing across simulated cell types. We vary sites per read from 1 to 40, reads per sample, dominant-type purity from 5% to 95% (following the breast-cancer benchmark design), and the rates of genomic DNA contamination and precursor transcripts. Metrics are recovery of $A_{tk}$ (total variation distance), calibration of the test for $K=1$ versus $K>1$, and the minimum sites-per-read at which the concentrated case becomes detectable. The minimum sites-per-read is the molecular analogue of the macrophage study's 20-read threshold.

The second experiment uses real mammalian ground truth. We pool reads from the 24 single macrophages and cognate bulk data of RNA editing generates cellular subsets with diverse sequence within populations, hide cell labels, and ask whether molecular co-editing alone classifies Cybb as concentrated and Anxa5 as distributed. We also check whether the estimated fully-edited fraction matches the observed share of edited cells. This works only for sites close enough to share a read. Hyperediting was detected in that study, which suggests such reads exist, and we restrict the test to those loci.

The third experiment is the plant study. In Glycyrrhiza uralensis, whose 38 high-confidence plastid sites across 19 genes are already catalogued (Ma et al.), we generate rRNA-depleted, DNase-treated long-read RNA libraries with matched genomic DNA. Samples span leaf, stem and root and a developmental series, so that cell-type composition varies substantially. We focus on the multi-site ndh clusters and on mitochondrial start- and stop-creating sites of the kind documented in nad4L and atp9 (Hao et al.). For validation, we sort protoplasts into major leaf cell types for bulk long-read sequencing and run targeted plate-based single-cell amplicon sequencing at 20 or more reads per site per cell for a panel of 10 to 15 sites.

The decisive check is whether sites the model flags as concentrated within a type show bimodal per-cell editing in the amplicon data, and whether sites it calls distributed do not. We also require that the shuffle ablation erases the flag, which confirms the signal comes from linkage between sites. A clean pass establishes that bulk long reads can reveal where functional organellar proteins are actually being made. A failure has two readable outcomes. If the flag survives the shuffle, it reflects precursors or organelle-level mosaicism and not cell identity. If flagged sites are unimodal in the single-cell data, the within-type claim is false.