When a gene carries a premature stop codon, cells usually destroy the faulty messenger RNA before much protein is made. This quality-control step is called nonsense-mediated decay (NMD). When decay fails or is bypassed ("escape"), the cell instead makes a truncated protein. That truncated protein can be harmless, partly functional, or toxic, and which of these happens decides whether a mutation acts like a missing gene copy or like a poison. The largest maps of decay efficiency come from bulk RNA sequencing of tumors and tissues. They score each variant by how far the mutant allele's share of RNA reads falls below its share of DNA reads. Bulk tissue is a mixture of cell types, though. In a tumor, infiltrating immune and stromal cells express only the normal allele. A gene that is strongly expressed in those cells therefore looks "decayed" even when the tumor cells fully escape NMD. Separately, sequencing centers differ in how efficiently reads carrying the variant align and amplify, and these technical artifacts shift the same ratio.
We propose to estimate NMD efficiency per cell type from bulk allele counts. The model treats each sample's mutant-allele reads as a mixture whose weights come from cell-type composition and from how much of the gene's RNA each cell type contributes. Each sequencing batch gets its own allelic bias term, calibrated on heterozygous germline sites in the same samples, which should sit at a 50/50 allele ratio. The core claim is narrow and testable. The quantity "how strongly a gene is expressed in contaminating cells relative to the mutant cells" is a confounder of published NMD-efficiency estimates. Removing it, along with batch-specific allelic bias, should make decay track the known exon-junction rules more tightly and should expose cell-type differences in decay that bulk averages hide.
What Prior Work Leaves Open
The motivating evidence comes from two bulk studies. Kim et al. define NMD efficiency as $-\log(\mathrm{VAF}_{\mathrm{RNA}}/\mathrm{VAF}_{\mathrm{DNA}})$, where VAF (variant allele frequency) is the mutant allele's share of reads. They fit a random forest on somatic nonsense variants in more than 9,000 tumors from The Cancer Genome Atlas (TCGA) and on germline variants in 54 tissues from the Genotype-Tissue Expression project (GTEx). Their main findings are that counting downstream exon-junction complexes in a 55–500 nucleotide window beats the binary 50-nucleotide rule, and that common germline variants sit in a "selection shadow" of near-zero decay. They also report that roughly 40% of the variance in decay efficiency remains unexplained.
Their within-sample ratio cancels library size and transcription rate, but it does not cancel dilution. For a somatic heterozygous mutation, DNA VAF scales with tumor purity. RNA VAF scales with the tumor's share of that gene's transcripts, which is a different number whenever stroma and tumor express the gene at different levels. Sun and Chen take a different route: they split GTEx expression variants into those that change the fraction of transcripts targeted for decay and those that change decay efficiency itself. Their decay variants are tissue-specific and concentrated in brain, but they explicitly cannot say which cell types drive the effect. Both papers motivate this proposal, and neither supplies a mixture model.
The method ingredients come from deconvolution and batch-correction work, which our model inherits only in part. Count Bridges deconvolves bulk integer counts into single-cell profiles with a count-valued generative process whose cells sum exactly to the bulk. POPPY maps bulk profiles into a single-cell foundation-model space using ontology-aware contrastive alignment. Both work on total gene counts and never on reference-versus-alternate reads at a locus.
The batch methods constrain the design rather than supply it. BERMUDA aligns only biologically matched clusters across batches using a distribution-matching loss, and scAdapt uses adversarial alignment of a latent space. Both replace counts with continuous embeddings. The allelic ratio we need is exactly what such embeddings throw away. The lesson we take from them is BERMUDA's: correct only what is shared, so that real batch-specific biology is not erased. We implement that lesson on counts, using internal controls rather than learned alignment.
Isoform ambiguity also matters, because whether a stop codon has an exon junction downstream depends on which isoform carries it. The long-read single-cell colorectal cancer atlas shows that transcript structures differ by cell type. We use it to assign cell-type-specific downstream-junction counts.
Core Claim and Model
For variant $v$ in bulk sample $s$, let $a_{vs}$ be alternate reads and $n_{vs}$ total reads at the locus. The model is
$$a_{vs} \sim \mathrm{BetaBinomial}\!\left(n_{vs},\ \mathrm{logit}^{-1}\!\Big(\mathrm{logit}\big(\textstyle\sum_k w_{gks}\, g_{vks}\, e_{vk}\big) + \beta_{b(s)} + \gamma_{b(s)} x_v\Big),\ \phi_{b(s)}\right).$$
The terms are defined as follows. $w_{gks} = \pi_{ks}\mu_{gk}/\sum_j \pi_{js}\mu_{gj}$ is the share of gene $g$'s transcripts coming from cell type $k$. It is built from the sample's cell-type proportions $\pi_{ks}$, which are estimated by deconvolution of total counts, and from reference per-type expression $\mu_{gk}$. $g_{vks}$ is the mutant allele's genomic dosage in type $k$: 0.5 for germline heterozygotes, 0.5 in tumor cells and 0 elsewhere for clonal somatic mutations. $e_{vk} \in [0,1]$ is the fraction of mutant transcripts that survive in type $k$, so cell-type decay efficiency is $-\log e_{vk}$.
The batch terms, indexed by sequencing batch $b(s)$, are a reference-mapping offset $\beta_b$, a slope $\gamma_b$ on locus features $x_v$ such as mismatch context, GC content and indel status, and a batch-specific overdispersion $\phi_b$. All batch terms are fit on heterozygous synonymous and intronic germline SNPs, where $e=1$ by construction, and then frozen.
Kim's estimator is the special case with one cell type, $w$ equal to purity, and no batch terms. The key new ingredient is $\mu_{gk}$, the expression weighting. The model makes a sharp prediction: naive decay efficiency should correlate positively with the stromal share $1-w_{g,\mathrm{tumor},s}$ across genes at fixed purity, and the corrected estimate should not.
Identifiability of separate $e_{vk}$ for germline variants comes from samples with different compositions sharing a variant, as in GTEx. For somatic variants we estimate only the tumor-compartment $e$, which is what a mixture of clonal tumor and wild-type stroma can actually pin down.
Research Program
The program has three stages. First, we fit the batch calibration across TCGA sequencing centers and GTEx collection sites. Within each batch we check that calibrated germline controls center at 0.5, and that their dispersion no longer differs by center.
Second, we build reference expression $\mu_{gk}$ and composition $\pi_{ks}$ from single-cell atlases. Composition is estimated with Count Bridges deconvolution, chosen because it respects count geometry. We then fit the hierarchical model with partial pooling of $e_{vk}$ toward a regression on junction-count and constraint features, the same features Kim et al. use. This makes the pooled model directly comparable to their random forest.
Third, we ask the biological question bulk cannot answer. In GTEx tissues, do germline nonsense variants show cell-type-dependent survival? For example, does survival differ between neurons and glia in brain? And do Sun and Chen's decay-efficiency loci act in one cell type or everywhere?
Baselines
The naive reference point is Kim et al.'s raw estimator, $-\log(\mathrm{VAF}_{\mathrm{RNA}}/\mathrm{VAF}_{\mathrm{DNA}})$, computed on identical variant sets.
The strongest existing predictive method is the Kim et al. random forest trained on those raw targets. The file gives no exact R² or other test metric, only the statement that about 40% of variance remains unexplained. We will retrain it on our splits with the same features and the same hyperparameter search budget as our pooling regression. Beating it means that, on held-out sequencing centers, decay efficiency explains more variance as a function of downstream-junction count. It also means the prediction error for corrected targets is lower than for raw targets, with no added sequence features.
The strongest deconvolution competitor runs Count Bridges separately on the reference-read and alternate-read channels and reads off per-type allele fractions. It gets the same atlas references and compute. No reported result for allele-level deconvolution with Count Bridges or POPPY was found on Lacuna.
We also run two ablations. The first, purity-only correction, sets $\mu_{gk}$ equal across types and isolates the expression-weighting ingredient. The second, no-batch-calibration, drops $\beta,\gamma,\phi$ and isolates the allelic batch model. As a control showing why embedding-based correction is unsuitable, we apply Harmony-style or BERMUDA correction to expression and then recompute VAF from the corrected values.
Why not just…?
Why not just divide by tumor purity? This is the most serious alternative. It will work exactly when a gene is expressed equally in tumor and stroma. Our purity-only ablation is that estimator. The decisive test is the predicted correlation between naive or purity-corrected efficiency and the stromal expression share. If that correlation is negligible in TCGA, expression weighting adds nothing. In that case we would report the batch calibration and the germline cell-type analysis as the contribution.
Why not just measure allele ratios directly in single-cell data? Droplet 3′ or 5′ sequencing covers few nonsense loci, and per-cell allele counts are dominated by dropout and allelic bursting. Single-cell cohorts are also far smaller than TCGA. We use single-cell data where it is strong, as ground truth for pseudobulk recovery and for references, and use bulk where it has the samples.
Why not just batch-correct expression with existing tools first? BERMUDA and scAdapt output continuous embeddings or normalized expression. They do not preserve paired reference and alternate counts at a locus, so there is no allele ratio left to model. The embedding-correction control will show this concretely as distorted germline control ratios.
Why not just deconvolve each allele channel with Count Bridges? This is plausible. It ignores the coupling that reference and alternate reads share one locus depth and one composition, and it has no batch-specific mapping bias. Whether that coupling matters at typical locus depths of tens of reads is an empirical question, and the pseudobulk benchmark below answers it directly.
Proposed Experiments
The first experiment is a ground-truth recovery test. We use Onek1k, the roughly $10^6$-cell genotyped blood cohort that Count Bridges validated on. We pick heterozygous germline sites covered by its reads. Because the cohort is genotyped, synonymous sites give known $e=1$. At nonsense sites, we treat per-cell-type observed allele fractions as targets. We then build pseudobulks with randomized compositions and simulated center-specific mapping offsets. The metric is error in recovered $e_{vk}$ against the pseudobulk truth, compared across all baselines.
The second experiment tests isoform-aware recovery. In the colorectal long-read atlas, we assign cell-type-specific downstream-junction counts and test whether cell-type differences in recovered $e_{vk}$ track those counts.
The third experiment is a held-out-center evaluation on TCGA. Batch terms are fit on training centers' germline controls and transferred via locus features. We report the stromal-share confound test, the variance in decay explained by junction features, and agreement of escape calls with the final-exon and start-proximal escape rules.
The decisive outcome has two parts. First, purity-only correction leaves a significant stromal-share correlation that our model removes. Second, the remaining variance shrinks on held-out centers. If either fails, the hypothesis that dilution and batch bias explain part of Kim's missing 40% is wrong, and we will say so.
References
- A spectrum of nonsense-mediated mRNA decay efficiency along the degree of mutational constraint
- Mapping genetic variants for nonsense-mediated mRNA decay regulation across human tissues
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- scAdapt: virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species
- An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing