Count-based modeling sharpens single-cell eQTL discovery at cohort scale
Large-scale single-cell RNA sequencing has changed what can be measured in human genetics. Instead of averaging gene expression across a tissue, investigators can now ask how genetic variation alters transcription in specific cellular contexts, where regulatory effects often disappear in bulk analysis. The challenge has been computational as much as biological: many current single-cell eQTL pipelines were not built for sparse count data, and some become impractical when applied to very large cohorts. This study addresses both problems with jaxQTL, a pseudobulk single-cell eQTL mapping framework built around efficient count-based models.
The central idea is straightforward. Single-cell RNA sequencing does not behave like a smooth continuous assay; it produces sparse counts with zero inflation and strong cell-to-cell variability, and those characteristics matter when testing genetic effects. Linear models can work, but they do not always respect the distribution of the data. In this analysis, a negative binomial framework delivered calibrated inference in simulations and outperformed linear models, a result that fits the underlying biology of transcriptional bursts, sampling noise, and overdispersed read counts. For large sc-eQTL studies, the statistical model is not a technical afterthought. It determines whether weak regulatory signals are recovered or lost.
Negative binomial inference outperformed linear models in simulations
Simulations showed that jaxQTL with a negative binomial model maintained calibrated inference while improving power relative to linear models. That matters because the detection problem in single-cell genetics is usually dominated by modest effect sizes rather than dramatic shifts in expression. When the distributional assumptions match the data, the test statistics behave more reliably, and borderline associations are less likely to be dismissed or inflated. The result is a more trustworthy map of genetic regulation.
The computational gain was just as important. jaxQTL ran substantially faster than single-cell Poisson mixed effect models, which often impose heavy processing demands when applied to population-scale datasets. Speed is not merely a convenience in this setting; it determines whether a study can move from proof of concept to genome-wide discovery. A method that cannot scale to thousands of samples or millions of variant-expression tests will miss the practical threshold required for modern regulatory genetics.
In OneK1K data, jaxQTL identified 16% more eGenes than linear models. A single percentage point can be misleading in genomics, but this is not a marginal change. More eGenes means more candidate regulatory loci, more putative target genes, and more opportunities to connect GWAS signals to transcriptional mechanisms. The increase also suggests that the linear framework was leaving measurable biology on the table, particularly for signals that were too subtle to survive dilution in less suitable models.
Distal regulatory effects emerged more clearly than in bulk tissue analyses
The most biologically revealing result came from fine-mapped signals. Fine-mapped sc-eQTLs were more distal from the transcription start site than those identified in all cells, with P = 1 × 10−4, and they were more enriched in cell-type-specific enhancers, with P = 3 × 10−10. Those two findings point in the same direction: single-cell resolution is surfacing regulatory variation that bulk tissue measurements tend to blur. Enhancer activity is often cell-type restricted and physically separated from the genes it regulates, so a bulk assay can dilute the relevant signal across many irrelevant cells.
This is the biological reason single-cell mapping matters. Distal enhancers act through chromatin looping and context-dependent transcription factor occupancy, and those mechanisms are frequently invisible when all cells in a tissue are averaged together. The enhancer enrichment seen here suggests that sc-eQTLs capture a more precise view of gene regulation, especially for loci whose effects depend on cell state or lineage identity. The distance to the transcription start site also carried information about specificity: sc-eQTLs were largely shared across cell types overall, but cell-type specificity increased with distance from the transcription start site. Proximal signals appear more broadly shared, while distal effects are more likely to depend on local chromatin context.
One sentence captures the core finding plainly: distal regulatory variants were easier to see once the analysis stopped averaging away the cell type in which they acted.
Blood and immune traits gained more heritability coverage from single-cell signals
The downstream impact was measured against human disease genetics. Across 16 blood and immune-related traits, sc-eQTLs explained a greater proportion of single nucleotide polymorphism-heritability than bulk-eQTLs, at 9.90 ± 0.88% versus 6.10 ± 0.76%. This comparison is important because it links method performance to trait interpretation. A model that captures more heritable signal is doing more than improving statistics; it is improving the fraction of genetic architecture that can be assigned to transcriptional regulation.
The study also prioritized disease-associated genes missed by the bulk approach, including IL6ST. That example is especially relevant in immune biology, where the same gene can have different regulatory controls depending on cell type, activation state, and local cytokine signaling. Missing such a gene in bulk analysis is not a minor omission. It can redirect downstream functional follow-up, candidate target selection, and pathway interpretation. Yet the result did not close the gap entirely. The missing link between GWAS and eQTLs has narrowed, but it has not disappeared.
What jaxQTL changes for regulatory genetics
jaxQTL is best understood as a methodological correction to a growing biological problem. As single-cell cohorts expand, the limiting factor is no longer just how many cells can be sequenced, but how well the data are modeled once they arrive. Count-based frameworks acknowledge the structure of transcriptomic noise instead of smoothing it away, and that choice improved both calibration and discovery in this work. The advantage was visible in simulations, in OneK1K data, and in trait-level heritability analyses.
The broader implication is that cell-type context is not optional when interpreting genetic risk. Bulk eQTL maps remain useful, but they can miss enhancer-linked, distal, and cell-type-dependent regulatory effects. The present analysis shows that single-cell data can recover some of those signals, especially when paired with an efficient model that respects sparse counts. Still, the study leaves several practical questions unresolved: how performance will shift across tissues beyond blood, how rare cell populations will affect power, and how far this approach can be generalized across sequencing platforms and study designs. Those are the next pressure points for the field, and they will determine how much of complex trait biology can finally be assigned to specific cells rather than averaged tissues.