Population-scale single-cell RNA sequencing allows researchers to trace how inherited genetic differences influence gene regulation within specific immune cell types, uncovering single-cell expression quantitative trait loci that govern cellular functions. Conventional computational methods for mapping these regulatory loci often face substantial statistical and numerical hurdles because single-cell datasets feature sparse count matrices with pervasive zeros. Many established analytical workflows either rely on linear models that normalize counts through continuous transformations or implement single-cell mixed-effects models that demand prohibitive computational run times when applied to thousands of human participants. To resolve these limitations, computational biologists developed jaxQTL, an open-source statistical framework that conducts expression quantitative trait locus mapping on pseudobulk count data using generalized linear models. By directly modeling count distributions rather than relying on rank-based normalizations, the platform preserves biological variation across varying cellular contexts while harnessing just-in-time compilation to dramatically accelerate large-scale population genomic discoveries.
At its mathematical foundation, jaxQTL aggregates single-cell read counts into cell-type-specific pseudobulk totals for each donor and evaluates the conditional expectation of gene expression using generalized linear models. The statistical formulation accommodates both Poisson and negative binomial likelihoods, incorporating donor-level covariates such as age, biological sex, and principal components derived from genotyping arrays alongside library size offset terms that adjust for sequencing depth differences. In this framework, allelic effect sizes directly correspond to proportional shifts in transcription rates, contrasting sharply with linear models applied to rank-inverse-normal-transformed counts where effect estimates reflect changes in relative rank. To fit the underlying null models without individual genetic variants, the software executes iteratively reweighted least squares utilizing Cholesky matrix factorizations. The platform subsequently tests cis-genetic variants across a one-megabase window centered on transcription start sites by applying score test statistics computed across genotype blocks, with gene-level significance determined through beta-approximated permutation routines or saddle point approximation-corrected score tests.
To evaluate statistical validity and sensitivity under controlled conditions, the investigators conducted extensive synthetic simulations benchmarking jaxQTL against existing tools including SAIGE-QTL and tensorQTL under single-cell Poisson mixed-effects models parameterized from empirical data. The simulations tested three distinct immune cell populations representing high, medium, and low abundances, specifically CD4 naive and central memory T cells, immature and naive B cells, and plasma cells across varying sample coverage rates. Under simulated null scenarios where cis-heritability equaled zero, jaxQTL using a negative binomial distribution maintained well-calibrated type I error rates across varying minor allele frequencies and participant sample sizes. In contrast, pseudobulk Poisson models exhibited pronounced test statistic inflation caused by underestimated standard errors. When evaluating statistical power, the negative binomial formulation demonstrated substantial advantages over linear regression models and tensorQTL, showing the most pronounced performance gains for genes with lower expression coverage and in rarer immune cell types where sparsity is most severe.
When deployed across empirical single-cell RNA sequencing data from the OneK1K cohort, which comprises 1,267,768 peripheral blood mononuclear cells gathered across fourteen distinct cell types from 982 donors of European ancestry, jaxQTL confirmed its enhanced detection power. Across all fourteen immune cell populations, jaxQTL with negative binomial modeling identified 18,907 eGenes at a false discovery rate below 0.05, representing a 14 percent increase over linear models implemented within the same framework and a 21 percent gain over Poisson models. Evaluating genes exhibiting sample coverage greater than 10 percent revealed that jaxQTL detected 19,115 eGenes, outperforming the widely used linear association tool tensorQTL by 16 percent while delivering comparable performance to the single-cell mixed-model package SAIGE-QTL. Genetic loci uniquely discovered by the negative binomial model displayed significantly lower expression coverage than shared discoveries, verifying that count-based modeling captures authentic regulatory signals for lowly expressed genes that standard linear transformations systematically miss during pseudobulk processing.
The computational architecture of jaxQTL provides dramatic efficiency gains by executing optimized operations via JAX just-in-time compilation across central processing units, graphics processing units, and tensor processing units. In benchmarking experiments evaluating fifty randomly selected genes on chromosome 1 across the 982 OneK1K donors, permutation-based jaxQTL completed analyses in twenty minutes on a single graphics processing unit and five minutes on a tensor processing unit, operating substantially faster than SAIGE-QTL running on identical cohorts. To model future biobank-scale studies, the authors upsampled the OneK1K dataset to 10,000 individuals, reflecting anticipated dimensions of upcoming population initiatives such as TenK10K. In abundant cellular populations like CD4 naive T cells, jaxQTL ran 646 times faster on graphics processing units and 1,416 times faster on tensor processing units than SAIGE-QTL. When replacing permutation routines with the ACAT-V combination test coupled with saddle point approximations, run times rivaled ultra-fast linear tools while retaining full count distribution accuracy.
To verify that discoveries from jaxQTL reflect genuine biological mechanisms rather than cohort-specific technical artifacts, the authors tested replication across independent experimental cohorts spanning multiple ancestral backgrounds. In the CLUES peripheral blood mononuclear cell study, evaluating 88 European-ancestry and 88 Asian-ancestry participants, lead single nucleotide polymorphism and eGene pairs replicated at rates of 57 percent and 40 percent, respectively, demonstrating consistent directional effects across matched immune populations. Furthermore, 81 percent of single-cell regulatory variants replicated in European whole-blood data from the Genotype-Tissue Expression project, while 57 percent replicated within fluorescence-activated cell-sorted immune subsets from the DICE study. The identified regulatory variants also mirrored established evolutionary selection constraints observed in bulk tissues: genes exhibiting depletion of loss-of-function variants harbored significantly smaller regulatory effect sizes, whereas lead regulatory variants with lower minor allele frequencies displayed larger overall effect magnitudes, confirming the biological fidelity of the mapped associations across independent platforms.
Applying statistical fine-mapping through the SuSiE algorithm to 18,281 eGenes outside complex genomic regions yielded 95 percent credible sets for 12,978 eGenes, with 88 percent explained by a single putative causal variant. Comparing these fine-mapped single-cell variants against baseline bulk tissue eQTLs identified substantial architectural divergence between resolution levels. Single-cell regulatory variants were situated significantly farther from transcription start sites and showed lower localization in immediate core promoter regions than bulk-derived counterparts. Stratified linkage disequilibrium score regression and functional genomics analysis demonstrated that fine-mapped single-cell variants were significantly enriched within cell-type-matched open chromatin peaks identified by single-cell assay for transposase-accessible chromatin sequencing, as well as cell-type-specific enhancers and active promoters cataloged in reference epigenomic databases. Furthermore, single-cell discoveries exhibited pronounced enrichment in cell-type-matched enhancer-gene connections previously predicted by single-cell chromatin accessibility modeling, underscoring their capacity to prioritize functional distal regulatory elements that remain obscured in aggregated bulk tissues.
Quantifying regulatory sharing across diverse immune cell lineages revealed key patterns governing transcriptional specificity. A multivariate joint analysis utilizing mashr, which directly models effect size magnitudes and residual sharing while mitigating differential statistical power, demonstrated that 67 percent of fine-mapped regulatory loci maintained consistent effect directions across all fourteen evaluated immune populations. However, quantitative effect sizes proved far more divergent: only 15 percent of variants showed universal sharing by magnitude, while 9 percent exhibited strict cell-type specificity, with monocyte populations harboring the highest proportion of unique regulatory signals. Crucially, cell-type-specific regulatory variants were situated substantially farther from transcription start sites, averaging over 102,000 base pairs away compared to approximately 53,700 base pairs for shared regulatory variants. These empirical distributions confirm that while proximal promoter regulation operates largely ubiquitously across related immune cells, transcriptional variation driven by distal enhancer elements exhibits heightened cell-type specificity and context-dependent regulatory activity.
Integrating single-cell regulatory loci with summary statistics from genome-wide association studies across sixteen complex immune disorders and hematological measurements provided deep insight into the cellular origins of disease risk. Fine-mapped single-cell regulatory variants across all fourteen cell types accounted for 9.90 percent of total single nucleotide polymorphism heritability across the tested traits, significantly exceeding the 6.10 percent captured by conventional bulk-derived regulatory variants. Conditioning trait heritability on cell-type-specific annotations correctly pointed to biologically established effector cell populations across diverse conditions. For example, naive and central memory CD8 T cells captured enrichment for inflammatory bowel disease and primary biliary cirrhosis, effector CD8 T cells were highlighted for rheumatoid arthritis and hypothyroidism, and naive CD4 T cells were implicated in celiac disease risk. Similarly, natural killer cell regulatory loci were specifically linked to eczema susceptibility, illustrating that cell-resolved regulatory mapping links polygenic trait heritability directly to distinct physiological cell compartments.
Beyond broad polygenic architecture, jaxQTL successfully prioritized specific causal candidate genes that escaped detection in bulk sequencing analyses, particularly across rheumatoid arthritis loci. At the ANKRD55-IL6ST risk locus, the primary genome-wide association study variant rs7731626 was nominated as the top causal single-cell regulatory variant for IL6ST in naive CD4 and CD8 T cells with high posterior inclusion probabilities, colocalizing strongly with rheumatoid arthritis risk. In contrast, bulk tissue analysis failed to identify any association with IL6ST due to cellular dilution across heterogeneous blood samples, instead only detecting the shared ANKRD55 association. Functional genomic datasets including promoter capture Hi-C contacts and H3K27ac histone modification peaks validated the presence of a T-cell-specific distal enhancer linking rs7731626 to the IL6ST promoter. Similar single-cell-specific colocalizations were identified for IRF5 in conventional monocytes and CD40 in plasma cells, demonstrating that count-based single-cell mapping successfully uncovers elusive molecular pathways underlying human complex autoimmune diseases.
Subject of Research: Statistical genetics and computational framework for single-cell expression quantitative trait locus mapping
Article Title: Efficient count-based models improve power and robustness for large-scale single-cell eQTL mapping
Article References: Zhang, Z. E., Kim, A., Suboc, N., Mancuso, N., & Gazal, S. (2026). Efficient count-based models improve power and robustness for large-scale single-cell eQTL mapping. Nature Genetics. https://doi.org/10.1038/s41588-026-02830-5
Image Credits: AI Generated
DOI: 10.1038/s41588-026-02830-5
Keywords: jaxQTL, sc-eQTL mapping, single-cell RNA sequencing, pseudobulk analysis, negative binomial model, generalized linear models, OneK1K cohort, computational genetics, JAX acceleration, fine-mapping, cis-regulatory elements, distal enhancers, mashr analysis, SNP heritability, stratified LD score regression, rheumatoid arthritis, IL6ST, SAIGE-QTL, tensorQTL
News Source: Juliet Wilcox. (November 12, 2026). Count-Based Pseudobulk Framework Accelerates Single-Cell Genetic Mapping and Uncovers Distal Regulators. Scienmag.
