Ribosome profiling (Ribo-seq) sequences the short stretches of messenger RNA that ribosomes physically shield from enzyme digestion. It therefore measures which genes are being translated into protein, rather than merely transcribed. Its headline statistic is translational efficiency (TE): ribosome footprints on a gene divided by that gene's mRNA abundance. Ribo-seq is almost always run on lysates of whole tissue. A whole tissue is a mixture of cell types that may translate the same gene at very different rates. When the mixture shifts, for example when immune cells infiltrate an injured kidney, the bulk TE of a gene can change even though no single cell changed how it translates that gene. Current tools for detecting differential translation treat each sample as one homogeneous entity, so they would report such a composition shift as translational regulation.
We propose a deconvolution model for paired bulk RNA-seq and Ribo-seq that does three things. It writes bulk footprints as a cell-type mixture of mRNA times cell-type-specific TE. It absorbs the gene-specific technical biases of each Ribo-seq library batch into a separate term. It estimates only the quantities that this structure makes identifiable: how TE differs between cell types, and how TE changes within a cell type across conditions. The practical deliverable is a test that splits every bulk "differential translation" call into a composition part and a within-cell-type part. The scientific claim to be tested is that, in heterogeneous tissue, a substantial share of published bulk TE changes are composition effects.
Core Claim
Bulk TE is a weighted average of cell-type TEs. The weights depend on cell-type proportions and on how much mRNA each type contributes for each gene. Changes in TE can therefore be produced by changes in the weights alone. Ribo-seq's own technical distortions come from the library chemistry: RNase digestion preferences, ligation bias and fragment size selection. These act per gene and per library batch, multiplying a gene's footprints by roughly the same factor no matter which cell type the footprints came from.
That structure has a useful consequence. Absolute per-cell-type TE cannot be recovered from bulk data without spike-in standards, because it is confounded with the batch bias. Ratios of TE between cell types are invariant to that bias, and so are within-cell-type TE changes between conditions profiled in the same batch, and both remain recoverable. We will build the estimator around these invariant quantities rather than chasing absolute cell-type translatomes. Gap 1 of the research file motivates that broader goal, but it is not identifiable from bulk data alone.
Model
For gene $g$ in sample $s$, let $R_{gs}$ be the RNA-seq count and $F_{gs}$ the footprint count. Let $p_{ks}$ be the proportion of cell type $k$ in sample $s$, and let $m_{gk}$ be the reference mRNA profile of type $k$ taken from a single-cell RNA atlas. Let $a_g$ be the gene-specific scaling between that reference platform and bulk RNA-seq, and let $b_{g,\beta(s)}$ be the footprint bias of gene $g$ in the Ribo-seq batch $\beta(s)$ that sample $s$ was prepared in. Finally, let $\tau_{gkc}$ be the TE of gene $g$ in cell type $k$ under condition $c(s)$. The model is
$$ R_{gs} \sim \mathrm{NB}\Big(a_g \sum_k p_{ks} m_{gk},\ \kappa^{\mathrm{RNA}}_g\Big), \qquad F_{gs} \sim \mathrm{NB}\Big(b_{g,\beta(s)}\, a_g \sum_k p_{ks} m_{gk}\, \tau_{g k c(s)},\ \kappa^{\mathrm{Ribo}}_g\Big), $$
where NB is a negative binomial with mean as the first argument and gene-level dispersion $\kappa$ as the second.
The proportions $p$ are shared between the two assays. Each assay has its own dispersion, following the finding in RiboDiff that footprint libraries are markedly more overdispersed than RNA libraries; RiboDiff fits separate dispersions for the two assays, and we reuse that technique. We also shrink TE toward a gene-level shared value, using empirical Bayes in the RiboDiff style. The RNA half of the model is essentially the measurement-error generalized linear model of MEAD, which supplies both the method and its identifiability argument for $p$ under gene-specific platform scaling.
MEAD assumes that scaling ratios are unrelated to true expression. That assumption would be badly violated if one fed footprints into MEAD directly, because TE is gene-class specific (Gap 4 of the file). Our model avoids the violation by giving TE its own parameter instead of hiding it inside a scaling ratio.
Because $b_{g,\beta}$ and the overall level of $\tau_{g\cdot\cdot}$ trade off exactly, we report only $\tau_{gkc}/\tau_{gk'c}$ (cell-type contrasts) and $\tau_{gkc}/\tau_{gkc'}$ (within-type condition changes). Both are free of $b$. Identifying cell-type TE requires proportions to vary across samples. The design limit is therefore roughly that the number of samples per batch must exceed the number of cell types modeled. We will make this limit explicit and fall back to coarser cell-type groupings when it is not met.
Research Program
We will implement the model with alternating estimation. First, $p$ and $a$ are fit from RNA-seq alone, as in MEAD. Then $\tau$ and $b$ are fit from the footprints with $p$ held fixed. A final joint refinement lets the footprints sharpen $p$, which matters when cell types differ in how much ribosome they carry. A parametric bootstrap gives uncertainty for the reported ratios.
On top of the fitted model sits a decomposition test. For any gene whose bulk TE changes between conditions, we compute the bulk TE change predicted from the fitted proportion shift with $\tau$ held fixed (the composition part). We then test whether a residual within-type change in $\tau$ is still required (the translational part). This turns the output of Xtail or RiboDiff into an explained list rather than a raw one.
The engineering is modest. The real work is building benchmarks with known answers, because the file reports no existing ground truth for cell-type TE in mixtures. We will build three:
- Simulations. We will simulate data from the model at RiboDiff-like dispersions, with batch biases drawn from empirical footprint bias profiles.
- In silico mixtures. We will create pseudo-bulk mixtures by subsampling and combining reads from Ribo-seq libraries of distinct pure populations. Examples are the PC3 prostate cancer mTOR-inhibition libraries (GEO GSE56887, used by Xtail and RiboDiff) and the interferon-gamma-treated macrophage libraries analysed by Xtail. Pure libraries prepared in different batches will be mixed so that batch bias is real rather than simulated.
- A wet-lab mixture. Two cell lines will be physically mixed at five known ratios. Ribo-seq and RNA-seq will be prepared in two independent library batches, with each pure line profiled alone as truth.
Baselines
The naive reference point is to run Xtail and RiboDiff on the bulk mixtures and take their per-gene TE calls as if the tissue were homogeneous. Both tools are run exactly as published, on the same counts. On homogeneous simulated data Xtail reports an area under the curve of 0.96 with three replicates Xtail. On GSE56887, RiboDiff reports 1,142 genes with significant TE change, nearly double the original study RiboDiff. Our first measure of success is how many of these calls on mixtures are composition artifacts, assessed against the pure-population truth.
The strongest existing competitor for the specific question "is translation changing within cell types?" is a covariate-adjusted RiboDiff. In this baseline, proportions estimated by MEAD from the paired RNA-seq are added as covariates to RiboDiff's generalized linear model on the log scale. This is the obvious patch, it uses the same inputs, and it can absorb some composition effects. It will receive the same proportion estimates, the same batch labels as covariates, and the same samples.
A second competitor is a two-step ratio approach. Here one runs TAPE, whose deep autoencoder learns cell-type-specific expression from bulk data, separately on the RNA and footprint matrices, then divides the per-type outputs. TAPE gets the same reference and a matched tuning budget: the same number of fine-tuning epochs, and hyperparameters selected on one held-out mixture set for both methods. TAPE reports lower error than CIBERSORTx, Scaden and MuSiC on transcriptomic proportions, but the file gives no numeric value. No reported result was found on Lacuna for any method recovering cell-type-specific TE from bulk Ribo-seq.
Two ablations isolate the key ingredients. The first removes the batch bias term $b$, collapsing it to a gene-level constant. The second replaces the separate footprint dispersion with the RNA dispersion.
Beating the field means three things on held-out mixtures:
- higher precision at matched recall for within-type TE changes than covariate-adjusted RiboDiff;
- lower error on cell-type TE ratios than two-step TAPE;
- for the batch-term ablation, a measurable loss of ratio accuracy when mixtures span batches, with no loss when they come from a single batch.
Why not just…?
Why not just deconvolve the RNA-seq and apply those proportions to the footprints? This assumes every cell type has the same TE for every gene, and the same total ribosome output. That is exactly the quantity under study. Chen & Dickman found that mRNA and translation rates are largely decoupled in fly muscle ($r^2 = 0.100$). Our model contains this alternative as the special case where TE is constant across cell types, so the likelihood-ratio comparison on the mixtures directly tests whether it suffices.
Why not just add proportions as covariates to RiboDiff? This is the real threat. Mixing is additive on the count scale, but a log-scale covariate assumes multiplicative effects. The patch is therefore misspecified whenever cell types differ strongly in TE or proportions move far. The misspecification may still be harmless in practice when proportions change only slightly. The wet-lab mixture series spans large proportion shifts specifically to settle this. If the covariate patch matches our model across that range, we will report that the contribution is the decomposition diagnostic and the identifiability analysis, not the estimator.
Why not just use TRAP, the transgenic tagging of ribosomes in one cell type? TRAP needs engineered driver lines, so it cannot be applied to human biopsies. The fly study also reports that tissue-specific Ribo-seq had 138% higher sensitivity than TRAP for low-TE genes Chen & Dickman. Where tagged data exist, we will treat them as validation rather than competition.
Why not just batch-correct the footprints with a single-cell integration tool? scBatchProx aligns precomputed embeddings with per-batch adapters, and BERMUDA aligns matched clusters with an autoencoder. Both operate on single cells in a learned latent space. Neither gives a parameter that cancels in TE ratios, and with few bulk samples per batch such alignment risks erasing real condition differences. The batch-term ablation measures whether an explicit gene-by-batch factor is needed at all.
Decisive Checks
The method succeeds if, on two-batch in silico and physical mixtures, it recovers known between-type TE ratios and correctly declares no within-type change for genes whose bulk TE moved only through composition. It must do this better than covariate-adjusted RiboDiff, with properly calibrated false-discovery control.
The scientific claim is then tested on real tissue. We will reanalyse the whole-kidney injury Ribo-seq of Heruye et al., whose authors note that their signal averages epithelial, endothelial and immune cells. The question is what fraction of injury-associated TE changes, including those near AZIN1, survive once composition is modeled.
A negative result would also be informative: few calls being reattributed would show that bulk Ribo-seq conclusions are robust to tissue mixing. The identifiability limit on sample number will be reported alongside, so users know when the analysis is impossible.
References
- RiboDiff: Detecting Changes of Translation Efficiency from Ribosome Footprints
- Statistical inference for cell type deconvolution
- Genome-wide assessment of differential translations with ribosome profiling data
- Deep autoencoder for interpretable tissue-adaptive deconvolution and cell-type-specific gene analysis
- Development of a tissue-specific ribosome profiling approach in Drosophila enables genome-wide evaluation of translational adaptations
- Federated-inspired Single-cell Batch Integration in Latent Space
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- Inflammation primes the murine kidney for recovery by activating AZIN1 adenosine-to-inosine editing