Genetic variants associated with cell-type-specific intra-individual gene expression variability reveal mechanisms of genome regulation

Abstract

Gene expression varies substantially across cells, even within seemingly homogeneous populations. Understanding how genetic variation influences this variability is critical for elucidating mechanisms of genome regulation, yet the genetic control of cell-to-cell expression variability remains poorly characterised. Here, we introduce MEOTIVE (Mapping genetic Effects On inTra-Individual Variability of gene Expression), a statistical framework that identifies genetic effects on gene expression variability (sc-veQTLs) while modelling mean-variance dependence. Applying MEOTIVE to single-cell RNA-seq data from 1.2 million peripheral blood mononuclear cells across 980 donors in the OneK1K cohort, we identify 14–3,488 genes with significant sc-veQTLs across blood cell types, including 2,103 shared across multiple cell types. We further detect 55 SNP-gene pairs (34 genes) associated with gene expression dispersion (sc-deQTLs), enriched for immune and viral response pathways; 32 of these associations remain significant after additional adjustment for mean expression. Notably, rs1131017 in the 5′UTR of RPS26 is associated with widespread expression dispersion across cell types and reduced autoimmune disease risk. We also identify a monocyte-specific deQTL in LYZ (rs1384), which replicates in an independent cohort. MEOTIVE provides a robust framework for mapping the genetic regulation of gene-expression variability at single-cell resolution and reveals pathways linking cellular heterogeneity to autoimmune disease risk.

Using over 1.2 million single immune cells, the authors identify genetic variants that influence cell-to-cell variation in gene expression, revealing a previously hidden layer of genome regulation associated with immune function and autoimmune disease.

Introduction

Dissecting the genetic control of gene expression is fundamental to improving our understanding of the biological mechanisms of genome regulation. With the development of single-cell sequencing technology, many studies have sought to identify how genetic effects that underlie gene expression operate at the level of individual cells. These single-cell expression Quantitative Trait Loci (sc-eQTL) have revealed that the genetic effects underlying mean differences in gene expression between individuals are frequently cell-type specific^1–7. The eQTL association analysis is based on the model assumption that allelic variants have an additive effect on the mean mRNA expression levels across genotype groups. It is typically assessed using linear regression of a SNP’s genotype on the mean gene expression across a population^8–12. The variance QTL (vQTL), which differs from the eQTL and influences phenotypic variation (Fig. 1A), has been studied in complex human traits such as body mass index (BMI)⁸, bone marrow density (BMD)¹³, vitamin D¹⁴, glycemic traits¹⁵, and serum cardiometabolic biomarkers¹⁶. However, only a few studies^17–20 have investigated the characteristics of variance eQTL (veQTL), i.e., eQTL that affect variation in gene expression within genotype groups^17,18. They have identified inter-individual veQTLs (genetic effects on variance in expression levels across individuals) and linked them to G x G effects, such as cis-epistatic interactions¹⁸ and interactions with a second gene’s expression level^19,20. It could also be due to gene-environment (G x E) effects, such as interaction with the cellular environment. Most of these analyses are based on variance tests that examine phenotypic variability among genotype groups, such as Levene’s test²¹ and the Brown-Forsythe test²².

Fig. 1. The illustration of variance eQTL and overview of the study design.

A Illustrative plot to explain the bulk eQTL/veQTL, mean-variance relationship in the single-cell RNA-seq data, and single-cell eQTL/veQTL/deQTL. In bulk-level analysis, each dot indicates the gene expression count for one individual. In single-cell level analysis, each dot indicates the pseudobulk mean ($\mu$), variance (${\sigma }^2$), and dispersion ($\theta$), for one gene from one individual in a specific cell type. The colour of the dots indicates the genotype group. B The analytical pipeline of the MEOTIVE framework identifies genetic variants that affect intra-individual gene expression variability. NB stands for negative binomial, and the colour of the histogram represents NB distribution with different parameters. The analytical pipeline figure was created in BioRender. Xue, A. (https://biorender.com/shortURL) is licensed under CC BY 4.0.

Beyond the inter-individual veQTL, intra-individual veQTL has been an under-investigated area that could expand our knowledge of the genetic control of gene expression variation at the single-cell level. This type of eQTL is defined as a genetic variant that affects gene expression variability across cells within an individual and between genotype groups (Fig. 1A). Such variability in a cell population (or within a predefined cell type) reflects cell-to-cell heterogeneity and may arise from genetic regulation, immune responses, the cell cycle, cell state, and/or stochasticity in cellular gene expression. One of the few empirical studies to tackle this question tested for intra-veQTLs (veQTLs hereafter for brevity unless otherwise specified) in sc-RNA-seq data from 5,447 induced Pluripotent Stem Cells (iPSCs) collected across 53 individuals. They identified five study-wide significant sc-veQTL, but demonstrated that the differences in variance were driven solely by statistical dependence rather than by true biological effect²³. As the authors noted, the sample size and the number of cells limited the discovery of sc-veQTL.

Intra-individual variation is particularly interesting because cell-to-cell heterogeneity in gene expression (especially in immune cells) within an individual might result from important functional consequences, i.e., genetic regulatory mechanisms on gene expression during development and homoeostasis, as well as a potential response to therapies, and pathological development (such as tumour growth). For example, a recent study²⁴, which collected peripheral blood mononuclear cells (PBMCs) from asthma patients, has linked sc-veQTL to the transcriptional response to immune stimuli and glucocorticoids. In addition, intra-individual variation could be a potential indicator for the G x E interaction with cellular traits such as the cell cycle and cell state at the single-cell level.

Several challenges are key to identifying sc-veQTLs and understanding their characteristics. First, we need to accurately estimate the intra-individual variance from the scRNA-seq data, where expression counts across cells are typically assumed to follow either Poisson, negative binomial (NB), or zero-inflated negative binomial (ZINB) distribution, especially when the number of cells per sample is low. Second, we need to account for the mean-variance relationship, as the aforementioned models are mathematically correlated. For example, assuming the gene expression level of a gene across cells follows a negative binomial distribution, the variance (${\sigma }^2$) can be estimated such that ${\sigma }^2=\mu +{\mu }^2*\theta$, where $\mu$ is the mean and $\theta$ is the dispersion of the distribution. In this scenario, sc-veQTLs could be primarily driven by differences in mean expression levels between individuals in different genotype groups. However, the dispersion ($\theta$) captures variability not attributable to the mean level. Third, the choice of appropriate statistical tests to detect sc-veQTLs remains uncertain. This may depend on the distributions of intra-individual variance/dispersion estimates across the cohort, and such distributions are likely to vary between genes and cell types.

Here, we introduce MEOTIVE (Mapping genetic Effects On inTra-Individual Variability of gene Expression), a statistical framework to identify the relationship between genetic variants and within-individual variability in gene expression at single-cell resolution. Leveraging data from the OneK1K cohort⁶, comprising genotype and scRNA-seq data from 1.27 million PBMCs collected across 980 individuals, we test the role of genetic variants in the variance of intra-individual gene expression at the population level while accounting for the mean-variance relationship (Fig. 1B). Considering the statistical challenges, we decided to use two methods in parallel to identify the sc-veQTLs that are not induced by mean effect (ME): (1) a model-free strategy to detect those significant sc-veQTLs that do not show significant results in the sc-eQTL association analysis and (2) a model-based strategy to first estimate the intra-individual dispersion parameter ($\theta$) of the expression and then perform the dispersion-eQTL (sc-deQTL) association analysis. Because mean and dispersion are intrinsically related in count data, MEOTIVE may not completely disentangle genetic effects on mean expression from effects on dispersion in all cases. We therefore also evaluate signals using mean-adjusted dispersion measures to identify associations robust to this dependence. We also investigate how sc-deQTLs vary across cell states and seek potential biological explanations. Overall, we provide a statistical framework to identify single-cell variance eQTL and reveal mechanisms by which genetic variants affect single-cell gene expression variability.

Results

To understand the relationship between genetic loci and intra-individual variation in single-cell gene expression, we began by exploring how to estimate data parameters and their relationships in sc-eQTL models. This is important due to: (1) previously identified relationships between estimates such as mean and variance^25–27, and (2) the potential impact of differences in the distribution of a gene’s expression levels, with genes varying between zero-inflated negative binomial through to Gaussian distributions.

The relationship between the mean and the intra-individual variance effects

Identifying genetic effects on variance is attractive because they are easy to interpret. However, as previously described, mean differences between individuals can correlate with intra-individual variance effects²³. Nevertheless, variance effects can exist independently, most likely as ‘true’ genetic effects arising from genotype-by-genotype (GxG) or genotype-by-environment (GxE) interactions. Using TensorQTL²⁸, we tested for intra-individual sc-veQTL per cell type, adjusting for sex, age, and latent variables (Methods). In total, we identified 4642 significant (q value < 0.05) vGenes across all 14 cell types (10,527 unique SNP-gene pairs), ranging from 14 to 3488 for each cell type. We define a vGene as a gene in which the top SNP is associated after a permutation test and multiple-testing correction (Methods). As expected, intra-individual mean estimates are highly correlated with variance estimates across all cell types (Fig. 2A and Supplementary Fig. 1A), and this correlation depends strongly on the mean estimates and the proportion of cells with zero expression (Supplementary Fig. 1B, C). Testing for overlap with mean sc-eQTL effects, we observe that 94.5% of vGenes also have mean effects (eGenes), and conversely, 60.4% of eGenes also have variance effects (Table 1 and Supplementary Fig. 1D). The results suggest that a small number of sc-veQTL may have variance-only effects and warrant further investigation. We evaluated the effect size, significance level, and TSS distance between vGenes with or without mean effects, and the comparison showed that the latter group is merely above the significance threshold (q value = 0.05), showing that there are not many genes whose intra-individual variability is not induced by mean difference (Supplementary Figs. 2–4 and Methods). The number of vGenes identified per cell type depends on sample size (${\widehat{\rho }}_s=0.93$) and the average number of cells per donor (${\widehat{\rho }}_s=0.98$) (Supplementary Fig. 1E, F), an observation also seen with eGenes (Table 1). Furthermore, sc-veQTL and sc-eQTL allelic effects are highly correlated (${\widehat{\rho }}_s=0.942~0.991$), confirming the relationship between mean and variance effects (Supplementary Fig. 5). Finally, for the vGenes that are also eGenes, all the top sc-veQTLs have the same direction of allelic effect, with RPS15A the only exception, but the mean effect of the top sc-veQTL (rs7193785) is not significant (nominal p value = 0.335). These results confirm the strong correlation between the intra-individual mean and variance (Supplementary Fig. 1A and Supplementary Fig. 6), verifying that most sc-veQTLs can be explained by mean effects.

Fig. 2. Effect size comparison among the mean, variance, and dispersion eQTLs.

A The x-axis and y-axis denote the beta estimate for significant sc-veQTL and their sc-eQTL effect. Each dot indicates an SNP-gene pair test, and the dot’s colour indicates the cell type. The grey dashed line denotes the diagonal line of the coordinate panel. B An Upset plot for the number of dGenes in each cell type and the dGenes shared across cell types. The number of each intersection was annotated as the number above the bar plot.

Table 1.

Summary of eGene and vGene identified in 14 cell types

Cell type	N	Median nr cells	Nr genes	eGene	vGene	overlap	vGene with ME	eGene with VE
BIN	975	69	12219	843	605	574	0.949	0.681
BMEM	970	40	11663	672	480	454	0.946	0.676
CD4NC	980	461	15433	4252	3488	3363	0.964	0.791
CD4ET	980	57	12040	986	692	655	0.947	0.664
CD4SOX4	295	8	8234	37	14	13	0.929	0.351
CD8NC	980	126	13508	1760	1263	1213	0.960	0.689
CD8ET	980	177	13798	1884	1369	1314	0.960	0.697
CD8S100B	959	29	11070	461	295	271	0.919	0.588
DC	726	9	10149	228	127	118	0.929	0.518
MonoC	851	29	11620	517	319	305	0.956	0.590
MonoNC	690	16	10544	497	297	284	0.956	0.571
NKR	750	10	9224	179	105	98	0.933	0.547
NK	980	143.5	13577	2099	1440	1360	0.944	0.648
Plasma	253	7	8983	70	33	31	0.939	0.443

N sample size; Median nr cells the median number of cells per individual; Nr genes number of genes tested in the QTL analysis; eGene number of significant eGene; vGene number of significant vGene; overlap the number of genes that are both eGene and vGene; vGene with ME the proportion of vGene that are also eGene; eGene with VE the proportion of eGene that are also vGene.

One explanation is that variance effects occur in addition to mean effects, i.e., the variance effects exceed those expected under the mean-variance relationship. To test this, we assessed genetic effects in the residuals from regressing variance on mean per gene (Methods). We identified 307 genes (q value < 0.05) with at least one significant residual-eQTL (Supplementary Data 1). However, only seven genes overlap with the vGenes and show no detectable mean effects. This limited overlap suggests that residual-based eQTL approaches may not reliably separate mean and variance effects in sparse count data, as 77.2% of genes exhibit a systematic relationship between mean expression and residuals (Supplementary Fig. 7), indicating potential model mis-specification. Previous studies have attempted other parameterisations to test for genetic effects on intra-individual gene expression variance, including the variance-to-mean ratio (VMR, also known as the Fano factor) and the coefficient of variation (CV)²³. However, these two metrics depend heavily on intra-individual mean differences (Methods and Supplementary Figs. 8, 9). On average, 70.0 ~ 94.1% cvGenes are also eGenes, and 57.9−90.0% vmrGenes are also eGenes, but we observed substantially lower power to detect effects with VMR (Supplementary Data 2). Another way to adjust the mean-variance relationship is to include the mean as a covariate when detecting sc-veQTLs per gene²⁹. However, this strategy is expected to be valid only when the mean and variance are linearly related; otherwise, the residuals may exhibit a spurious quadratic relationship with the mean. A recent study³⁰ proposed using a single polynomial regression to remove mean-variance dependence, but it might still ignore the genuine biological relationship between mean and dispersion because it forces the residuals to be independent of the mean. Morgan et al.³¹ used a local polynomial fit between the squared CV (i.e., CV²) and the mean across individuals to identify variability protein-QTLs. Another relevant study²⁷ estimates a latent gene-specific residual dispersion parameter by fitting a global trend between mean and over-dispersion estimates for all genes, but this was for differential variability testing between two groups of cells, and the method was based on the strong assumption that genes with similar expression levels are more likely to have similar dispersion levels.

MEOTIVE: direct estimation of dispersion identifies independent genetic effects in intra-individual single-cell expression variability

To address challenges arising from the mean–variance relationship, we developed MEOTIVE, a framework that tests for allelic effects on intra-individual expression dispersion while modelling the dependence between mean expression and dispersion. MEOTIVE first directly estimates the dispersion parameter (θ) for each gene’s intra-individual expression distribution using Cox–Reid adjusted maximum likelihood estimation (CR-MLE)³² (Methods and Supplementary Fig. 10). We then tested cis-SNP genotypes ( < 1 Mb) for association with intra-individual dispersion for each gene in each cell type, accounting for parameter-estimation accuracy (Methods). In total, we identified 55 significant (q value < 0.05) deQTL-dGene pairs across 34 unique dGenes (Table 2, Fig. 2B and Supplementary Fig. 11). We define a dGene as a gene with at least one SNP significantly associated with its dispersion level across samples. Most (78%) sc-deQTLs show an opposite allelic effect to their corresponding sc-veQTLs (Supplementary Data 3). This is because the distribution type of gene expression count per genotype group varies. For example, when the expression is low for a specific genotype group, the count distribution is more left-skewed, deflating variance estimates. For heterozygous individuals, the distribution approaches Gaussian; thus, the dispersion becomes much smaller, and the same holds for individuals with alternative homozygous alleles (see the example of RPS26 below, Fig. 3). This feature results in a negative relationship between intra-individual variance and intra-individual dispersion for many genes with significant sc-veQTLs and sc-deQTLs. In such cases, dispersion is a more reliable metric of variability because it explicitly accounts for mean-dependent changes in the underlying count distribution.

Table 2.

Summary of the 55 deQTL-dGene pairs identified in 14 cell types

Gene	Cell type	CHR:BP	rsID	A1	A2	AF	Dispersion				Mean			Variance
							b	se	pval	qval	b	se	pval	b	se	pval
S100A4	NK	1:153514241	rs7535476	C	A	0.050	0.565	0.093	1.95E-09	2.21E-03	-0.283	0.090	1.63E-03	0.091	0.093	3.30E-01
CDC42	CD4NC	1:22426187	rs1534949	C	A	0.419	-0.363	0.030	1.09E-31	1.12E-22	0.247	0.017	2.99E-44	0.079	0.019	2.47E-05
	NK					0.419	-0.353	0.030	8.92E-30	9.34E-21	0.192	0.026	2.05E-13	-0.052	0.025	3.79E-02
CD52	CD4NC	1:26616280	rs3924324	A	G	0.177	-0.380	0.051	1.27E-13	7.70E-07	0.259	0.030	1.10E-17	-0.012	0.042	7.75E-01
	CD8ET	1:26645806	rs11589222	C	T	0.253	0.407	0.047	1.86E-17	3.75E-10	-0.231	0.040	1.15E-08	-0.008	0.043	8.44E-01
GYPC	CD4NC	2:127516475	rs6732878	G	C	0.283	-0.182	0.029	4.61E-10	2.80E-04	0.666	0.030	1.92E-90	0.668	0.040	8.52E-55
GNLY	CD8ET	2:85934499	rs12151621	A	C	0.223	-0.708	0.037	1.16E-68	1.42E-55	0.863	0.038	1.04E-90	0.674	0.047	3.61E-43
	NKR	2:85916005	rs4832181	G	A	0.353	0.338	0.054	5.17E-10	2.02E-02	-0.635	0.047	2.00E-37	-0.142	0.054	8.80E-03
	NK	2:85920249	rs3755007	G	T	0.344	0.849	0.040	1.72E-81	9.68E-62	-0.942	0.036	5.92E-114	-0.588	0.040	3.51E-44
SNHG8	CD4NC	4:119204466	rs28517808	T	C	0.474	-0.139	0.025	2.02E-08	2.00E-02	0.918	0.023	3.85E-206	1.037	0.028	6.08E-188
FGFBP2	NK	4:15957763	rs4698429	G	A	0.151	-0.270	0.046	7.74E-09	6.01E-03	0.390	0.041	2.00E-20	0.302	0.044	1.92E-11
HNRNPH1	CD4NC	5:178986632	rs7703730	C	A	0.370	0.289	0.031	2.45E-19	7.95E-12	0.020	0.025	4.33E-01	0.103	0.024	1.51E-05
RNASET2	CD4NC	6:167370999	rs2769346	G	A	0.467	-0.153	0.026	5.70E-09	1.74E-03	0.876	0.021	1.91E-217	0.967	0.029	3.26E-160
HLA-A	CD4NC	6:29913266	rs1061156	T	G	0.159	-0.491	0.065	6.78E-14	4.44E-07	0.135	0.039	5.62E-04	-0.156	0.046	7.98E-04
HLA-C	CD4NC	6:31221914	rs9264219	C	T	0.316	0.302	0.050	2.03E-09	1.64E-03	0.014	0.037	7.10E-01	0.167	0.042	6.92E-05
	CD8NC	6:31321360	rs2844585	A	G	0.172	0.441	0.081	5.72E-08	3.86E-02	-0.333	0.055	1.46E-09	-0.005	0.064	9.39E-01
	CD8ET	6:31263051	rs2853926	G	A	0.277	-0.370	0.056	4.69E-11	1.26E-03	-0.013	0.040	7.48E-01	-0.268	0.049	4.48E-08
HLA-B	CD4NC	6:31366295	rs9394070	A	C	0.066	1.077	0.084	7.64E-35	7.75E-15	-1.483	0.061	4.43E-103	-1.222	0.066	1.82E-65
	CD4ET	6:31324955	rs34437781	T	C	0.065	0.748	0.092	1.81E-15	9.61E-06	-1.460	0.066	4.40E-88	-1.074	0.079	3.03E-38
	CD8NC	6:31327701	rs9378249	G	T	0.067	0.896	0.087	7.06E-24	6.40E-13	-1.432	0.060	1.20E-98	-1.098	0.069	1.02E-50
HLA-DQA1	BMEM	6:32589326	rs9271503	A	C	0.336	-0.321	0.052	7.47E-10	2.92E-03	0.047	0.037	1.97E-01	-0.165	0.047	4.28E-04
RPS18	BIN	6:33239869	rs17215231	T	C	0.070	0.494	0.077	2.40E-10	1.40E-03	-2.134	0.053	4.82E-208	-1.077	0.076	5.40E-42
	CD4NC					0.070	0.394	0.070	2.16E-08	9.57E-03	-2.274	0.041	2.20E-301	-1.650	0.057	2.60E-132
RPS10	CD8NC	6:34372804	rs7775635	G	A	0.078	0.382	0.071	1.01E-07	3.81E-02	0.997	0.042	3.43E-99	1.004	0.048	3.92E-81
LY6E	NK	8:144075281	rs4424237	T	C	0.474	0.232	0.030	1.97E-14	5.27E-08	-0.277	0.036	6.14E-14	-0.123	0.041	2.77E-03
PTGDS	NK	9:139848273	rs2271869	G	A	0.438	-0.419	0.039	2.18E-25	1.07E-17	0.486	0.036	1.05E-37	0.452	0.037	1.07E-32
ANXA1	CD4NC	9:75769950	rs2795112	G	C	0.113	-0.260	0.051	3.19E-07	4.54E-02	0.183	0.036	4.47E-07	0.100	0.042	1.64E-02
IFITM2	CD4NC	11:303271	rs6598046	A	T	0.248	-0.200	0.035	1.06E-08	2.25E-03	0.473	0.036	8.08E-37	0.379	0.037	2.71E-23
	CD8ET	11:307539	rs111412325	A	G	0.072	-0.460	0.065	2.25E-12	5.36E-06	0.463	0.070	4.59E-11	0.179	0.077	2.09E-02
	NK	11:349122	rs7117996	C	T	0.294	-0.383	0.042	6.73E-19	1.12E-11	0.589	0.038	2.26E-48	0.370	0.043	2.62E-17
CTSW	CD8ET	11:65645354	rs596002	A	G	0.196	0.671	0.038	1.68E-60	4.62E-45	-1.017	0.030	7.63E-167	-0.888	0.041	3.43E-86
	NK	11:65644027	rs583887	T	C	0.194	0.552	0.043	2.22E-35	6.10E-25	-0.816	0.032	3.06E-110	-0.631	0.044	6.60E-43
RPLP2	CD4NC	11:802902	rs28360884	T	G	0.321	0.216	0.035	7.20E-10	4.08E-04	0.601	0.026	1.22E-92	0.523	0.032	4.58E-54
RPS26	BIN	12:56435929	rs1131017	C	G	0.421	-0.661	0.041	9.09E-52	9.27E-38	1.333	0.016	<1E-323	1.150	0.026	1.41E-234
	CD4NC					0.420	-0.930	0.034	1.70E-124	9.22E-101	1.346	0.016	<1E-323	1.202	0.024	3.61E-274
	CD4ET					0.420	-0.860	0.036	3.10E-97	4.62E-78	1.340	0.016	<1E-323	1.146	0.026	3.07E-228
	CD8NC					0.420	-0.933	0.034	6.40E-122	6.46E-100	1.345	0.016	<1E-323	1.182	0.025	6.56E-254
	CD8ET					0.420	-0.680	0.040	1.48E-56	1.56E-42	1.355	0.015	<1E-323	1.250	0.022	1.25E-310
LYZ	MonoC	12:69747834	rs1384	T	C	0.495	0.280	0.041	1.48E-11	1.66E-04	-0.852	0.034	1.25E-101	-0.329	0.046	2.52E-12
KLRB1	CD4ET	12:9623841	rs10743738	T	G	0.413	0.236	0.041	1.28E-08	5.61E-03	-0.153	0.041	1.72E-04	0.065	0.041	1.12E-01
GZMH	NK	14:25083383	rs11158812	G	A	0.494	0.351	0.037	2.16E-20	8.43E-13	-0.251	0.032	9.77E-15	-0.112	0.034	9.82E-04
IL32	NK	16:3115272	rs45499297	C	T	0.088	-0.619	0.068	3.07E-19	3.03E-12	0.538	0.070	4.32E−14	0.340	0.068	8.39E-07
	CD4NC	16:3115628	rs1554999	A	C	0.370	0.847	0.039	4.44E-87	7.03E-71	-0.536	0.027	4.27E-73	-0.170	0.030	1.31E-08
	CD4ET					0.370	0.533	0.042	5.57E-34	8.43E-24	-0.406	0.033	3.03E-33	-0.047	0.039	2.29E-01
	CD8NC					0.370	0.555	0.042	4.88E-37	4.30E-27	-0.312	0.025	3.95E-34	-0.089	0.032	4.88E-03
	CD8ET					0.370	0.375	0.042	3.83E-18	3.82E-10	-0.291	0.033	5.00E-18	-0.043	0.041	2.85E-01
CCL3	NK	17:34397258	rs854471	G	A	0.317	0.279	0.042	5.19E-11	9.79E-05	-0.185	0.042	9.54E-06	-0.113	0.042	7.16E-03
CCL4	CD8ET	17:34411105	rs1634490	A	G	0.220	-0.336	0.041	4.45E-16	3.62E-09	0.286	0.048	3.30E-09	0.099	0.051	5.23E-02
EIF5A	CD4NC	17:7207964	rs7503161	A	C	0.412	-0.619	0.025	4.33E-104	3.32E-85	1.153	0.021	7.81E-304	0.990	0.027	3.63E-189
FXYD5	CD4NC	19:35658380	rs12461097	G	T	0.276	-0.180	0.033	6.79E-08	4.25E-02	0.158	0.026	1.44E-09	0.045	0.041	2.71E-01
RPS9	CD4NC	19:54700668	rs34172242	C	T	0.395	0.247	0.035	3.40E-12	2.10E-05	-0.841	0.028	9.33E-140	-0.530	0.037	9.05E-42
PPDPF	NK	20:62152519	rs72629024	G	C	0.145	0.315	0.052	1.79E-09	2.69E-03	-1.088	0.051	1.51E-83	-0.820	0.051	1.70E-51
ITGB2	CD8ET	21:46328099	rs760462	T	C	0.164	-0.257	0.040	2.45E-10	7.81E-05	0.906	0.029	1.00E-145	0.907	0.043	4.71E-80
	NK					0.164	-0.436	0.044	5.69E-22	5.20E-15	1.083	0.027	5.00E-203	1.071	0.041	4.36E-112
LGALS1	NK	22:38069305	rs62236671	A	G	0.340	0.264	0.039	2.91E-11	1.07E-05	-0.111	0.042	7.95E-03	0.033	0.044	4.52E-01

The columns indicate cell type, gene name, SNP rsID, minor allele, major allele, minor allele frequency, beta/se/nominal p-value/q-value estimates for deQTL, beta/se/nominal p-value for eQTL, and beta/se/nominal p-value for veQTL. A p-value of 1.95E-09 means 1.95 ×10−9. The statistical tests are two-sided and the p-values are exact values from nominal linear regression. The q-value is the gene-level multiple correction based on a beta-approximation permutation of each top SNP.

Fig. 3. The association between rs1131017 and RPS26 expression change in CD4 naïve cells.

A–D The violin plots of 980 individual genotypes of SNP rs1131017 correspond to the intra-individual mean, variance, residual, and dispersion of RPS26 expression in CD4 naïve cells. The x-axis indicates the genotype (coded as 0, 1, 2 indicating the number of G alleles carried). The number of donors in each genotype group is indicated in the bracket, i.e., CC (n = 169), CG (n = 485), GG (n = 326). Box plots show the median (centre line), interquartile range (box, 25th−75th percentiles), and 1.5× IQR (whiskers). E Scatter plot of intra-individual mean against intra-individual variance of expression. F Scatter plot of intra-individual mean against intra-individual dispersion of expression. The dispersion was estimated by CR-MLE method. G The distribution of SCT transformed count expression of RPS26 per cell. There are 463,496 cells, and each bar indicates the number of cells with the corresponding count expression. The table included the mean and dispersion estimates for the whole cohort and within each genotype group. The s.e. represents the standard error of the dispersion estimate. The genotype group of GG alleles have much higher intra-individual dispersion of RPS26 expression than the other two groups. H The distribution of SCT transformed count expression for all 980 individuals in CD4 naïve cells separated by three genotype groups. The colour of each square denotes the density of a certain count within a corresponding individual, and darker purple denotes higher density.

We subsequently evaluated the degree of cell-type specificity of the 34 dGenes. Of these, 67.6% (23) have evidence of an allelic effect in just one cell type. Of the remaining 11 dGenes, five have sc-deQTL in two cell types, four in three cell types, and two in five cell types (Fig. 2B and Table 2). Notably, two genes (IL32 and RPS26) had significant sc-deQTLs in five cell types (Table 2). For example, in both naïve and effector CD4 and CD8 cells, we identified rs1554999 as the top sc-deQTL for IL32. The A allele increases this gene’s dispersion level. On average, individuals carrying each copy of the A allele of rs1554999 have an increase of 0.25 in the dispersion level of mRNA transcript molecules across cells. The top sc-deQTL in the NK cell is a different SNP, rs45499297, but in LD with rs1554999 (R² = 0.056, D’ = 1). rs1554999 is located in the 5’ UTR region of IL32 and was previously reported to be strongly associated with methylation levels at three CpG sites in CD4 + T cells³³. While most dGenes were also vGenes, there was minimal overlap with residual genes (Supplementary Data 4), further supporting that intra-individual residuals capture different sources of variation than intra-individual dispersion. We additionally repeated the association analysis using dispersion residuals obtained after regressing dispersion on mean expression. Of the 55 deQTL–dGene pairs identified in the primary analysis, 32 remained significant in this mean-adjusted analysis. The remaining 23 associations therefore cannot be clearly distinguished from effects related to mean expression and should be interpreted with greater caution (Methods and Supplementary Data 5).

Functional annotations of sc-deQTLs and dGenes

To characterise sc-deQTLs and dGenes, we first used ANNOVAR³⁴ to annotate the top sc-deQTLs and then tested for dGenes overlap with GWAS loci (Methods). Most sc-deQTLs are in intergenic or intronic regions, suggesting that they or their tagged causal variants affect intra-individual variation via regulatory mechanisms. This observation aligns with those made for sc-eQTLs. Most dGenes (31 of 34) are reported to be associated with inflammation, immune responses, or viral infections. An example is GNLY, which has significant sc-deQTL in NK, NK_R, and CD8_ET cells (Table 2). GNLY encodes a protein called granulysin, present in natural killer cells and cytotoxic T lymphocytes, and has been shown to have antimicrobial activity against tuberculosis through changes in cell membrane integrity and reducing the viability of the bacillus^35,36. All of the sc-deQTLs for GNLY are located in intergenic regions. For example, the sc-deQTL for CD8_ET (rs12151621[A/C]) is located in a CTCF binding site (ENSR00000291834), ~8.1 kb downstream of GNLY. The A allele of this SNP has a decreasing effect on the dispersion of GNLY expression but an increasing effect on the mean expression, which means individuals with the CA or CC genotype will have lower mean expression but higher dispersion than individuals with the AA genotype (Supplementary Fig. 11). We hypothesise that the C allele confers a higher binding affinity for CTCF, which represses GNLY expression. The A allele has been reported to have an increasing effect on protein levels ³⁷and toxoplasma antibody IgG levels³⁸(p-value = 9.1 ×10⁻⁴; not genome-wide significant given the limited sample size = 557). This may suggest that individuals with an A allele tend to have higher IgG antibody levels against the Toxoplasma parasite. The body will produce more GNLY proteins in NK cells, with less variability across cells, ensuring sufficient functional proteins to kill microbes. A second example is seen in GZMH, which encodes a serine protease (granzyme H) and is constitutively expressed in CD8_ET and NK cells. We identified a sc-deQTL (rs11158812[G/A]) that has a significant effect only in NK cells (p value = 2.16 × 10⁻²⁰, q value = 8.43 × 10⁻¹³). The top sc-deQTL is in the intergenic region between GZMH and GZMB (no association between rs11158812 and GZMB in any cell type), a region previously reported to be associated with vitiligo³⁹. The top GWAS hit (rs8192917) is a missense variant (Arg55Gln) in GZMB, but it is not in LD with rs11158812 (R² = 1.2 × 10⁻³). Another example is IFITM2, which encodes interferon-induced transmembrane protein 2, a member of the interferon-induced antiviral protein family. This locus is not associated with human diseases but is highly associated with human blood cell traits⁴⁰ such as granulocyte, neutrophil, and eosinophil counts. The top sc-deQTL of IFITM2 differs in three cell types, but for naïve CD4 (CD4_NC) and CD8_ET cells, the top sc-deQTL is the same as the top sc-eQTL and in linkage disequilibrium (R² = 0.11 and D’ = 0.69). For NK cells, the top sc-deQTL (rs7117996) is located in the intergenic region and is independent of the other two top sc-deQTLs (R² < 0.01), but its top sc-eQTL (rs1059091, independent of rs7117996) is a missense variant in IFITM2.

There are also dGenes whose top sc-deQTL is located outside the intergenic region. For example, ITGB2 (integrin subunit beta 2) is a significant dGene in CD8_ET (q value = 7.81 × 10⁻⁵) and NK cells (q value = 5.20 × 10⁻¹⁵). In a previous study, the top sc-deQTL (rs760462[A1/A2]) located in intron 3 of ITGB2 was annotated as a splice-acceptor variant⁴¹. LYZ encodes lysozyme, which has antibacterial activity against several bacterial species and is highly expressed in monocytes (Mono) and dendritic cells (DC). The top sc-deQTL rs1384 is a 3’ UTR variant, and a recent study⁴² has suggested a monocyte-specific trans- action mediated by LYZ in this site. Fairfax et al.⁴³ have reported LYZ as a monocyte-specific master regulator, and its monocyte-specific cis-eQTL (rs10784774, in complete LD with rs1384) is also a trans-eQTL for 62 genes. The G allele of rs10784774, as well as the T allele of rs1384, is associated with lower expression but higher dispersion level of LYZ in classical monocytes (Table 2), and is also reported to be the increasing allele for neutrophil percentage of white cells⁴⁰. In addition, three dGenes are not eGenes: KLRB1, CCL3, and LGALS1. KLRB1 is the cell type marker for CD4 effector memory or central memory cells (CD4_ET). This gene encodes a killer-like receptor and was found to be upregulated in CD4 + T cells from patients with primary Sjögren’s syndrome⁴⁴. CCL3 is C-C motif chemokine ligand 3 and is rapidly upregulated during influenza infection⁴⁵. LGALS1 encodes galectin-1, which plays a key role in establishing and maintaining immune tolerance at the maternal–foetal interface by acting directly on NK cells⁴⁶.

Using the FUMA platform, we performed pathway enrichment analyses on the dGenes (Methods) and identified significant enrichment in three Hallmark gene sets: “Allograft Rejection” (P_adjusted = 1.86 × 10⁻⁶), “Interferon Gamma Response” (P_adjusted = 3.45 × 10⁻⁵), and “Interferon Alpha Response” (P_adjusted = 1.91 × 10⁻³). The KEGG pathway enrichment identified 12 significant pathways, with the top three “KEGG Viral Myocarditis”, “KEGG Ribosome”, and “KEGG Allograft Rejection”. However, after removing the five MHC genes, only “Allograft Rejection” remained significant in Hallmark gene sets and “KEGG Ribosome” in the KEGG pathway (Supplementary Data 6). We also performed a GO biological process enrichment analysis of dGenes. The most enriched process is “Interspecies Interaction Between Organisms” (P_adjusted = 2.44 × 10⁻⁷), “Cytokine Mediated Signalling Pathway” (P_adjusted = 3.97 × 10⁻⁷), and “Viral Gene Expression” (P_adjusted = 1.22 × 10⁻⁶). Our results suggest that dispersion effects on gene expression across cells within individuals are enriched for immune response and viral infection. This is consistent with prior knowledge that the G x E effect can induce phenotypic variability, and thus, the potential environment warrants further investigation.

Trans-regulation of dGenes partially explained the dispersion difference

Next, we explored potential genomic factors underlying sc-deQTLs. Previous vQTL GWAS, which test genetic effects on high-order trait variability between individuals, have often explored whether variance effects are caused by G x G or G x E interactions. Intuitively, we hypothesise that the dispersion difference in gene expression could be explained by (1) G x G cis-epistasis, where multiple independent cis-SNPs affect the expression variability of the target gene; (2) G x G trans-epistasis, where trans-QTLs are regarded as an interaction effect for the target gene and influence the variability level via trans-regulation.

We compared the number of independent cis-QTLs for the 34 candidate dGenes to assess whether multiple cis-regulatory mechanisms could drive the dispersion difference. On average, each dGene has 2.36 independent cis-eQTLs and 1.71 independent cis-veQTLs. The comparison also showed that no dGene has obviously more independent cis-veQTLs than cis-eQTLs. However, 14 dGenes have significant trans-deQTLs, among which there are also five dGenes (RPS18, SNHG8, GNLY, CCL3, and LGALS1) that do not have any trans- eQTL or veQTL (Supplementary Data 7). For example, rs78089025 [A1/A2] (9:73039725) showed a genome-wide significant (p-value = 4.04 × 10⁻⁸) association with the dispersion levels of GNLY in NK cells but not with the mean or variance levels. This SNP is an intronic variant in KLF9-DT, a divergent transcript of the transcription factor KLF9. When fitting GNLY’s trans-deQTL (rs78089025) and top cis-deQTL (rs3755007) in the same association model, the interaction term showed significant effects on dispersion level (P_interaction = 4.93 × 10⁻¹²) and a significant change in the main effect. Specifically, when only top cis-deQTL was fitted, beta = −0.171, s.e. = 0.009; when trans-deQTL was fitted in the interaction model, beta = −0.484, s.e. = 0.046. These results imply that sc-deQTLs are not induced by the G × G effect of independent cis-SNPs but can be partially explained by trans-SNP effects on the dGenes.

Genetic control of variance heterogeneity underlying different contexts

Since we show that trans-regulation could be a putative driver of sc-deQTLs, we further asked whether the G x E interaction between genetic effects and cellular state also affects intra-individual dispersion. To test this, we inferred the cell-state landscape for B cells and included the average cell state per individual as an interaction term in the sc-deQTL association model (Methods). We identified only one significant interaction between genotype and cell state for dGene RPS18 in B_MEM cells (adjusted p value = 4.66 × 10⁻³). One plausible explanation is that the pseudotime was calculated based on highly variable genes, so the gene expression variability of the dGenes might not be well captured by this approach.

We also tested whether sc-deQTL effects are associated with the interaction between genotype and sex or age. For genotype-by-age interaction, only HLA-B in naïve CD8 (CD8_NC) cells and RP11-1143 G9.4 in classical monocyte (Mono_C) cells showed significant associations (Supplementary Data 8). For genotype-by-sex interaction, only HLA-B showed significant interactions in effector CD4 (CD4_ET) cells (rs34437781) and CD8_NC cells (rs9394070). The rs34437781 was a significant sc-deQTL in CD4_ET cells (nominal p value = 1.81 × 10⁻¹⁵, q value = 9.61 × 10⁻⁶). rs9394070, which is in strong linkage disequilibrium (LD) with rs34437781 (R² = 0.886, D’ = 0.949), is a significant sc-deQTL for HLA-B in CD8 naïve cells and top sc-deQTL in CD4 naïve cells. However, when we fit the genotype-by-sex interaction term, the genotype itself became insignificant (genotype p value = 5.74 × 10⁻³, interaction p value = 5.43 × 10⁻⁸). The interaction is mainly driven by females with the TT genotype (4/980), whose intra-individual mean estimates are the lowest among 980 individuals, whereas dispersion estimates are relatively higher (Supplementary Fig. 11). After removing the four individuals with the TT genotype, the interaction term was no longer significant (interaction p value = 0.029). Similarly, the beta of the sc-deQTL in separate sex groups did not differ significantly (p value = 0.165). These results indicate that neither sex nor age is the main driver of genetic effects on dispersion levels across cells.

Association between dispersion eQTL and immune phenotypes

To understand the relationship between sc-deQTL and disease risk, we tested for overlap between sc-deQTL loci and public GWAS associations in the GWAS Catalogue (Methods). The most frequent traits include blood protein levels, asthma, eosinophil counts, type 1 diabetes, Crohn’s disease, height, and rheumatoid arthritis. Taken together with the FUMA enrichment results above, this further suggests that intra-individual dispersion effects are enriched among genetic associations with autoimmune and infectious diseases.

Highlighting RPS26 as an example, carrying copies of the G allele at rs1131017[G/C] increases intra-individual dispersion (Fig. 3). This SNP was tested against our association analysis’s dispersion level of 53 cis-genes but was only significant for RPS26, with shared allelic effects across five cell types (innate B cell, naïve/effector CD4, and naïve/effector CD8). This locus has previously been reported to be strongly associated with autoimmune diseases, including type 1 diabetes⁴⁷, asthma⁴⁸, vitiligo⁴⁹ and rheumatoid arthritis⁵⁰. The top SNPs are not the same one but all are in strong LD with each other (R² > 0.8). Furthermore, the C allele of rs1131017 was consistently associated with increased autoimmune disease risk, suggesting that lower intra-individual RPS26 expression dispersion is associated with higher autoimmune disease risk. When we directly estimate the dispersion of the unique molecular identifier (UMI) count distribution (after SCTranformation) across all cells in each genotype group (using CD4_NC cells as the example), the GG genotype group shows a much larger over-dispersion ($\widehat{\theta }=0.843$) than the CG or CC group ($\widehat{\theta }=0.046\mathrm{and}0.051$) (Fig. 3G). So, we further dissect the distribution of counts for each individual. We observe that the UMI count distribution for individuals in the GG genotype group mostly follows a negative binomial distribution, while the UMI count distributions for CG and CC individuals follow a Poisson-like distribution, which is statistically equivalent to a negative binomial distribution with a tiny dispersion parameter and can be modelled by a negative binomial model. This suggests that even for the same gene, genetic effects could affect the distribution of cell types within an individual. This observation would only be seen if we examined the count distribution by genotype group during sc-deQTL analysis.

The top sc-deQTL for RPS26 is located in the gene’s 5’UTR, the binding site for six transcription factors (RBM39, TCF7, LEF1, KLF6, CD74, and MAF)⁵¹. We hypothesised that if these transcription factors (TFs) regulate the expression level of RPS26 via binding to this site, rs1131017 should be detected as a co-expression eQTL between RPS26 and these TFs in our data. To evaluate this, we calculated co-expression between each TF and RPS26 for each cell type and performed a co-eQTL association analysis (Methods). For all 84 (6 × 14) cell-type tests, we identified 40 significant co-eQTLs (FDR < 0.05) (Fig. 4). The most frequent (11/14) co-eQTL is between RPS26 and CD74. Interestingly, the allelic effect in T cells is opposite to that in B cells and monocytes (Fig. 4A). These results indicate that promoter-mediated regulation of RPS26 by CD74 is cell-type specific. In CD4_NC cells, all six transcription factors have significant co-eQTL estimates, and TCF7 and LEF1 showed the opposite direction of co-expression to the other four TFs (Fig. 4B). Interestingly, co-eQTL effect sizes are generally larger in naïve than in effector cells, and this difference is not driven by cell-number differences between naïve and effector cell types. For CD4 cells, all six transcription factor genes showed significantly (p-adjust <0.05) larger co-eQTL effect size in the naïve cell type. This relationship is also observed in CD8 T cells for KLF6, CD74, and MAF. These results indicate that the effect of rs1131017 on co-expression between RPS26 and the transcription factors is influenced by the immune response.

Fig. 4. The putative regulatory mechanism of dispersion eQTL of RPS26.

A The beta effect estimates for co-expression eQTL between RPS26 and six transcription factors (TFs) across 980 donors in 14 cell types. Each point is an estimate coloured and grouped based on cell types, and the error bar denotes the standard error of the estimate. Some error bars are not obvious due to small values. Six TFs are annotated with different shapes. The significant estimates (BH corrected FDR < 0.05) are strengthened by transparency of the dot. B The violin plot for co-expression per genotype group in CD4NC cells. The y-axis indicates the standardised co-expression estimate (measured by Spearman’s correlation across cells per individual) between RPS26 and the TF gene. Box plots show the median (centre line), interquartile range (box, 25th–75th percentiles), 1.5× IQR (whiskers), and outliers (individual points). Four outlier dots were omitted for illustrative purpose.

Replication of sc-deQTL in an independent cohort

We attempted to replicate our identified sc-deQTL in an independent cohort from Perez et al.⁵² This dataset has three subcohorts: EAS CLUES of East Asian ancestry (N = 97) and EUR CLUES and the ImmVar Consortium of European ancestry (N = 89 and 46). Of the 34 dGenes we identified in OneK1K, we replicated 26 genes in the corresponding cell types at FDR < 0.05. For the most significant example of rs1131017-RPS26, we replicated the sc-deQTL association in 4 cell types (classical and nonclassical monocytes, CD4 T, and CD8 T cells). The association patterns were very consistent between OneK1K and Perez et al. and the example in classical monocyte (p_nominal = 4.97E-07, FDR = 2.83E-5) was demonstrated in Supplementary Fig. 12, and the correlation of the test statistics -log10(p_nominal) between the two datasets is very high (Pearson’s cor = 0.894). These results showed that we have obtained a reasonably great replication rate for dGenes.

Discussion

In this study, we present MEOTIVE, a framework for identifying genetic variants associated with intra-individual variability in gene expression at the single-cell level. MEOTIVE addresses the mean-variance relationship by modelling dispersion directly and applying additional filtering and sensitivity analyses. However, because mean expression and dispersion are intrinsically linked under count-based models such as the negative binomial, empirical data do not always allow complete separation of mean and variability effects. A key limitation is that MEOTIVE cannot fully disentangle dispersion effects from those associated with mean expression in all cases. When we repeated the association analysis using dispersion residuals obtained after regressing dispersion on mean expression, 23 of 55 deQTL–dGene associations (42%) no longer reached significance. These signals should therefore be interpreted with caution, as their apparent effects on dispersion may be at least partly related to effects on mean expression. Further methodological development is needed to derive measures of expression variability that are more completely separable from the mean. By applying MEOTIVE to data from the OneK1K cohort, we can successfully identify intra-individual deQTLs at the population scale. While most previous studies have focused on the mean effects of genetic variants on gene expression, the genetic effects on variability and dispersion across cells within individuals are poorly understood, and single-cell RNA-seq data provide a means to dissect the high-dimensional effects on genome regulation. In total, we identified 34 dGenes in our primary analysis, which accounted for the mean-variance dependency, and these were enriched for biological pathways relevant to interferon response, interspecies interactions, allograft rejection, and viral gene expression. These results suggest that single-cell transcriptional variability could arise from immune and/or external stimuli²⁴ and that this variability is under genetic control. These mechanisms are unlikely to reflect immune-cell analysis bias, as permutation-based enrichment analyses using 34 randomly selected eGenes (100 permutations) showed no enrichment for immune response or viral infection pathways (permuted P > 0.05). In our primary analysis, we identified 55 deQTL–dGene associations involving 34 unique dGenes. When the association analysis was repeated using dispersion residuals obtained after regressing dispersion on mean expression, 32 of the 55 deQTL–dGene pairs remained significant. Although we identified only a few pairs in the current study, with a larger sample size and more cells per individual, we would expect to discover thousands of genetic variants that affect the dispersion of intra-individual gene expression. Analyses with larger datasets are set to uncover a fundamentally new avenue of genetic effects on human genome regulation.

Herein, we also propose an explanation for the “zero-inflated” model in single-cell RNA-seq data. In the example of rs1131017-RPS26, we note that an observed zero-inflated negative binomial (ZINB) distribution across a population of cells results from a mixture of three genotype groups (CC, CG, GG), each showing a negative binomial distribution (Fig. 3G). In such cases, the so-called “structural” or “excessive” zeros are not generated from a separate biological process but from the genotype group with low-abundance expression. This finding challenges the conventional understanding of the underlying model for scRNA-seq count data and necessitates re-evaluating the previous zero-inflated negative binomial model.

This study has several limitations. First, although it is the largest single-cell eQTL cohort, with data from 980 individuals, we have limited statistical power to test only SNPs with MAF > 5%. Sarkar et al. ²³. predicted that 4000 individuals are needed to detect sc-deQTLs at 80% power. The number of cells per individual, per cell type, is also an important limitation because several cell types have only 100 to 200 effective samples. Second, dispersion-estimation accuracy depends mainly on the mean and the number of cells per individual. From our simulations, we observe that, with 500 cells, a mean >0.3 is required to obtain an accurate dispersion estimate. With 5000 cells per individual for a given cell type, we can lower the mean-expression filtering threshold to 0.1, which can rescue more genes for association testing (from 3% to 10%). Assuming CD4_NC cells are the largest group, the minimum number of cells required for single-cell RNA-seq data is ~10,000 per individual. Even in such cases, rare cell types, such as plasma or dendritic cells, would still have only 70 to 100 cells per individual. Third, the high sparsity in the 10X data is one reason we cannot better understand the underlying model of scRNA-seq data. A recent study⁵³ demonstrated that lower sequencing depth makes the observed data more consistent with a Poisson distribution, even if the true model is overdispersed. Since ~50% of intra-individual means are 0 in our data obtained using the 10x v2 kit, processing samples with a higher capture rate will improve the estimation of the true underlying distribution of scRNA-seq data. Finally, 23 of 55 deQTL signals did not reach significance when testing dispersion residuals obtained after regressing dispersion on mean expression. These signals should therefore be interpreted with caution, underscoring the need for further work to dissect the underlying count model and to develop more robust metrics for expression variability that are fully independent of the mean.

In summary, we identified genetic effects on within-individual variability in gene expression while accounting for mean-variance dependence in scRNA-seq data from human PBMCs. The MEOTIVE statistical framework we present here can be applied to any single-cell RNA-seq dataset with genotype information to identify genetic variants that influence intra-individual variability in gene expression. As cohort sample sizes increase (e.g., TenK10K), ongoing analyses will continue to reveal genetic mechanisms underlying inter-individual variability and cell-to-cell heterogeneity.

Methods

The OneK1K cohort

The OneK1K cohort comprises genotype and single-cell gene expression data for 982 individuals of Northern European ancestry. Ethics for this study were approved by the Tasmanian Health and Medical Human Research Ethics Committee (H0012902). We have obtained informed consent from all participants. Each individual was genotyped and imputed using 759,993 SNPs from the HRC panel⁵⁴. There are 1,267,758 peripheral blood mononuclear cells (PBMCs) with gene expression data after demultiplexing and doublet removal. Following the cell-type classification in Yazar et al.⁶ we predicted the OneK1K cohort into 14 cell types using the scPred (v1.9.2)⁵⁵ (Table 1). During a sensitivity test of latent variables⁵⁶, we identified two outlier samples (one due to a low number of cells and the other due to extremely imbalanced cell composition). We excluded them from all the analyses in this study. Thus, the final sample size we retained in this study is 980.

Strategy of estimating genetic effects on intra-individual expression variability

To accurately estimate the mean and variance of intra-individual gene expression, we applied several steps to exclude potential confounding factors in the single-cell RNA-seq data. First, the count matrix was pre-processed by Seurat using sctransform v0.3.2 algorithm⁵⁷ to remove the technical confounders such as sequencing depth and batch effects. Second, we classified all cells into 14 cell types using a semi-supervised method (scPred), and excluded individuals with fewer than 5 cells in each cell type to avoid biased estimation of intra-individual mean and variance driven by outliers. Third, genes expressed in fewer than 10% of individuals or with an intra-individual mean across the cohort less than 0.001 were also excluded (Methods). After quality control, the median number of cells per individual ranges from 7 (plasma cells) to 461 (CD4_NC cells), and the numbers of individuals and genes also vary across cell types (Table 1).

We first used the moment estimate of the intra-individual mean and variance per gene to generate two M * N matrices for each cell type (Methods), where M is the number of genes and N is the number of individuals for each cell type.

The moment estimates of mean and variance for gene $k$ of individual $i$ across cell $j$, are denoted as the following:

$${\mu }_{ik}=\frac{1}{n_i}{\sum }_{j=1}^{n_i}{x}_{ijk}$$ 1

$${\sigma }_{ik}^2=\frac{1}{n_i-1}{\sum }_{j=1}^{n_i}{\left(x_{ijk}-{\mu }_{ik}\right)}^2$$ 2

where $n_i$ indicates the number of cells for individual $i$. We also calculated the proportion of zero expression within an individual (${\pi }_0$) and detected a strong negative relationship between the ${\pi }_0$ and $\mu$ (Supplementary Fig. 13).

For single-cell RNA-seq data, the intra-individual mean and variance are correlated because a large proportion of them follow non-normal distributions such as Poisson, NB or ZINB distributions^23,53. In our OneK1K data set, Spearman’s correlation coefficients between the intra-individual mean and variance across individuals for each gene were high. For example, ~94.6% of genes showed ${\widehat{\rho }}_s>0.8$ and ~ 75.2% showed ${\widehat{\rho }}_s>0.99$ (Supplementary Fig. 1A and Supplementary Note 1). On the other hand, the correlation estimates were strongly dependent on the mean expression level in a negative trend (Supplementary Fig. 1B), and so did the proportion of non-expression individuals per gene (Supplementary Fig. 1C and Supplementary Fig. 13). Given such a strong mean-variance dependency, we predict that the significant sc-veQTLs could be primarily explained by the effects of the mean difference. To better understand the characteristics of vGenes and their relationship with eGenes, we compared vGenes with and without ME across several aspects. First, vGenes without ME showed significantly larger q values than those with ME across all cell types, with most q values clustering around 0.05 (Supplementary Fig. 2). Second, the effect sizes of the sc-veQTLs without ME were much smaller than those with ME in five cell types (CD4_NC, CD4_ET, CD8_NC, CD8_ET, and NK), with 1.54 ~ 3.94 fold smaller for the median effect size (Supplementary Fig. 3). Third, we tested whether the sc-veQTLs without ME were closer to the TSS location than those with ME. The results showed that the TSS distance of sc-eQTLs is not significantly different from sc-veQTLs (Supplementary Fig. 4, Welch Two Sample t-test, p value = 0.91), and sc-veQTLs without ME are uniformly distributed based on the TSS distance. Thus, alternative methods to rule out mean effects are needed when estimating genetic effects on the variability of intra-individual gene expression. We also estimated the relationship between the mean and other metrics, including the variance-mean ratio (VMR) and the coefficient of variation (CV). These two metrics are also highly correlated with the intra-individual mean (Supplementary Figs. 8–9). A previous study has used them as the dispersion indicator in the eQTL data set²³. However, since most of the mean of gene expression in single-cell RNA-seq data are close to 0, the VMR or CV could approach large numbers and are sensitive to even tiny changes when the mean expression is low. Thus, VMR and CV are not suitable as dispersion indicators for single-cell datasets (Supplementary Note 2).

The TensorQTL for SNP-gene association analysis

To understand how genetic variation between individuals affects the variance in intra-individual gene expression, we performed association analysis using TensorQTL (v1.0.7)²⁸ for each cell type. We first filtered out those genes with expression in less than 10% of individuals or extremely low inter-individual abundance ($\mu <0.001$). The intra-individual mean, variance, and dispersion were log(x + 1) transformed and then z-score normalised per gene to avoid extreme outliers. The residual expression matrix was just z-score normalised per gene. The sex, age, first 6 principal components (PCs), and first 10 PEER factors⁵⁸ were fitted as covariates in the model. The PCs are calculated by PLINK⁵⁹ based on the genotype information. The PEER factors are derived from the intra-individual mean of the gene expression matrix for each cell type to capture latent variables. We chose 10 PEER factors to be fitted into the association model using a sensitivity analysis and a local greedy method to balance discovery power and overfitting⁵⁶. The number of remaining individuals, genes, and the median number of cells per individual for each cell type are presented in Table 1. In the cis-QTL analysis, we retained only ~4.2 million SNPs located within a ± 1 Mb cis-region from the centre of the gene body and with a minor allele frequency (MAF) greater than 0.05. After obtaining the nominal p values for every SNP-gene pair, a beta-approximation permutation was applied to correct the p values and 10,000 times of permutations were conducted for each gene. The most significant SNP for each gene (top cis- eQTL or veQTL) was further corrected, and the permuted p value was converted to a q value to control the false positive per chromosome⁶⁰. An SNP-gene association with q value < 0.05 was deemed significant.

The estimation of dispersion in gene expression distribution

We adopted two methods to generate dispersion estimates for intra-individual gene expression.

First, we used a straightforward method: regressing out the intra-mean from the intra-variance, and using the residuals as the dispersion indicator.

Second, we assume all the intra-individual gene expression follows a negative binomial (NB) distribution. This assumption is based on the comprehensive evaluation that thousands of genes are in favour of the NB model rather than the Poisson model in at least 10% individuals (Supplementary Note 3 and Supplementary Figs. 14−15). Let:

• $x_{ijk}$ be the number of molecules for individual $i$, cell $j$, gene $k$ after accounting for confounders and size factor

• ${\mu }_{ik}$ be the mean of expression of gene $k$ in individual $i$

• ${\theta }_{ik}$ be the dispersion of expression of gene $k$ in individual $i$

Then we assume,

$$x_{ijk}~NB\left(\cdot ;{\mu }_{ik},{\theta }_{ik}\right)$$ 3

The likelihood function for the intra-individual distribution of each gene is

$$L\left(r,p;x_1,x_2,\dots ,x_n\right)={\prod }_{j=1}^n\frac{\mathrm{\Gamma }\left(x_j+r\right)}{x_j!\mathrm{\Gamma }\left(\mathrm{r}\right)}{\left(1-p\right)}^{x_j}{\left(p\right)}^r$$ 4

We need to estimate the r and p, where $\mu =\frac{pr}{1-p}$ and ${\sigma }^2=\frac{pr}{{\left(1-p\right)}^2}$. Alternative parameterisation is to use theta ($\theta =1/r$) as the dispersion parameter. We used the “glmGamPoi” R package (v1.8.0) which implements the Cox-Reid adjusted MLE⁶¹ method to estimate the dispersion parameter based on the SCTranformed count data. Then we generated an intra-individual dispersion matrix for each cell type.

When checking the preliminary results (326 sc-deQTLs with q value < 0.05), however, we found that, for many genes, the CR-MLE estimates were highly inflated, especially for genes with low abundance or for individuals with a small number of cells (Supplementary Fig. 16). This is because when the mean of an NB distribution is low, the likelihood curve is very flat, making it extremely difficult for the optimisation algorithm to locate the maximum. This scenario became even worse in the single-cell RNA-seq data, since fewer than 10% of genes have an intra-individual mean greater than 0.1, and around half of the intra-individual expression is 0 (Supplementary Fig. 2). Although the CR-MLE method partially mitigates the problem by using a penalised log-likelihood⁶², we still observed inflated dispersion estimates in many genes from the real data (Supplementary Fig. 16). To mitigate spurious discovery of sc-deQTLs arising from instability in dispersion estimation, particularly in the regime where the true dispersion lies near the boundary of the parameter space (θ ≈ 0), we evaluated the behaviour of Cox–Reid adjusted maximum likelihood (CR-MLE) dispersion estimates across a range of mean expression levels, dispersion parameters, and sample sizes (Supplementary Note 4, Supplementary Figs. 17–18). In sparse count regimes typical of single-cell data, we observe an upward bias in dispersion estimates for lowly expressed genes, which can lead to apparent overdispersion even when the underlying count process is effectively Poisson. Because our framework aims to identify genetic effects on intra-individual variability, such estimation artefacts can generate false-positive dispersion QTL signals if not accounted for.

To quantify this effect, we simulated count data under a negative binomial model across a grid of mean expression values (μ), dispersion parameters (θ), and cell numbers (M), and estimated dispersion using the CR-MLE approach. These simulations allowed us to determine the conditions under which dispersion estimates reliably recover the true underlying parameter. Based on these results, we derived data-driven thresholds on intra-individual mean expression for each cell type, defined such that >90% of simulated estimates fall within the true dispersion parameter (Supplementary Data 9).

For example, in CD4_NC cells, genes were required to have a mean expression >0.3 to ensure reliable dispersion estimation under this criterion. Applying these thresholds reduced the influence of mean-dependent dispersion inflation and yielded 64 candidate deQTL–dGene pairs. Nevertheless, we observed residual inflation of dispersion estimates in some genotype groups (examples shown in Supplementary Fig. 19), consistent with remaining mean–variance coupling. To further minimise this effect, we excluded genes for which any genotype group had mean expression below the cell-type-specific threshold. After this additional filtering step, 55 significant deQTL–dGene pairs (q value < 0.05), corresponding to 34 unique genes, were retained.

Because dispersion and mean expression can remain partially coupled in negative binomial models, and many deQTLs coincide with eQTLs, we further assessed whether the detected signals could be explained by mean effects. We therefore repeated the QTL analysis using dispersion residuals obtained after regressing dispersion on mean expression. Of the 55 deQTL pairs (34 unique genes), 32 pairs (24 unique genes) remained significant (Supplementary Data 5), indicating that the majority of detected deQTL signals cannot be explained solely by genetic effects on mean expression.

We also tried to estimate the dispersion for gene $k$ of individual $i$ based on the methods of moments, such that

$${\sigma }_{ik}^2={\mu }_{ik}+{\theta }_{ik}*{\mu }_{ik}^2$$ 5

so, the dispersion can be estimated as,

$${\theta }_{ik}=\frac{{\sigma }_{ik}^2-{\mu }_{ik}}{{\mu }_{ik}^2}$$ 6

From the equation, it is obvious that (i) when the variance and mean are very close to each other and the mean is not so small, the moment estimator will be close to 0; (ii) when the mean is 0, the moment estimator does not exist but for such case, we manually assign the dispersion level as 0; (iii) in real data, the moment estimator could be a negative number, but the scale would not be large (CR-MLE estimate will always be non-negative).

To assess empirically how much the dependence between CV/VMR metrics and mean expression levels affects QTL mapping, we forced these metrics to be 1 when the mean is 0 and ran the association test. More details on different dispersion indices and their special cases when the mean is small are discussed in Supplementary Note 5 and Supplementary Data 10.

We also test whether the estimation of the dispersion parameter is affected by the ‘sctransform’ function. We performed an analysis comparing dispersion estimates before and after ‘sctransform’ in CD4_NC cells (Supplementary Fig. 20) and found negligible differences between the two groups, suggesting that sctransform would not result in artificial deQTL signals.

Functional annotation and gene set enrichment analysis

We used FUMA⁶³ (v1.5.3) to perform functional annotation of the sc-deQTLs and perform gene enrichment for the dGenes. SNP functional annotation is performed by the built-in ANNOVAR³⁴ on the platform. To test whether the list of dGenes is overrepresented in certain biological functions, the list is compared against gene sets from the Molecular Signatures Database (MSigDB). The Hallmark gene sets are from the MSigDB h collection (50 gene sets), the KEGG are from MSigDB c2 collection (186 gene sets) and GO biological processes from MSigDB c5 collection (7751 gene sets). The significant enrichment is defined at an adjusted p value ≤  0.05. We utilised the FUMA platform to overlap the sc-deQTL with public GWAS associations from the GWAS catalogue (database update by 27/4/2023). We identified 553 matched associations across 233 studies.

The G x G epistasis analysis to identify driving factors for sc-deQTL

We performed two complementary analyses to investigate whether G x G and G x E effects could be the potential driving factors for sc-deQTLs. To identify multiple independent cis-signals for the same gene, we sought to map conditionally independent cis-eQTLs and veQTLs using a stepwise regression procedure²⁸. For the trans-QTL analysis (both for the mean and variance), we tested all SNPs located > 1 Mb away from the gene body centre and matched the results to 64 candidate deQTL-dGenes pairs in each cell type. Significant trans-QTL is defined at a nominal p-value < 5 × 10^-8. We further fit the genotype of those trans-veQTLs into the cis-veQTL or cis-deQTL association model to see whether the estimates change significantly.

Pseudotime trajectory of intra-individual variance and interaction tests

To understand the context-dependent effects of the prioritised sc-deQTLs on cell state, we estimated pseudotime for each cell in the inferred B cells (B_IN + B_MEM). We used SCTransform⁵⁷ to calculate Pearson residuals for scaled expression using the top 500 highly variable genes and fitted the percentage of mitochondrial expression and experimental pools as covariates. After transformation, we calculate the principal components (PCs) of the expression matrix and construct a UMAP using the first 30 PCs via the RunUMAP() function built into Seurat. We then used PHATE (v1.0.7)⁶⁴ to estimate the quantitative indicator of cell state (i.e., pseudotime) in a two-dimensional space for each cell. For each individual, we computed the mean pseudotime across all cells and created a mean pseudotime trait. Then the mean pseudotime is tested as an interaction term (G x E) in the QTL association model (for mean, variance, and dispersion separately) in TensorQTL software. Sex or age was also tested for interaction effects. The nominal p value is first corrected by multiple testing based on the effective number of independent variants in the cis-window, and then converted to a Benjamini-Hochberg adjusted p value for each chromosome.

Identifying co-eQTLs for RPS26 and its transcription factors

We subset the SCTransformed count matrix for RPS26 and six transcription factors per cell type, and calculate Spearman’s correlation between RPS26 and each TF gene across cells within each individual. For each TF gene, we have a co-expression estimate for each individual as the phenotype, and we run a linear regression of the co-expression phenotype on the genotype at rs1131017. The nominal p values are then converted to Benjamini-Hochberg FDR, and the test with FDR < 0.05 is deemed significant.

Replication in an independent cohort of non-European ancestry

To replicate our findings of sc-deQTLs in OneK1K of European (EUR) ancestry, we utilised another single-cell cohort from Perez et al.⁵² and conducted sc-deQTL mapping for the individuals of East Asian (EAS) ancestry (97 individuals, including 75 healthy controls and 22 lupus patients). The other two subsets of European ancestry (i.e., EUR CLUES cohort and ImmVar Consortium) were also included (N = 89 and 46). The gene expression matrix was processed using the same protocols as for OneK1K. Intra-individual dispersion of gene expression was also estimated for each gene and cell type. For the sc-deQTL mapping, covariates were adjusted for in the association model, including sex, age, batch, the first 6 PCs, the first 2 PEER factors, and lupus disease status. Given the differences in SNP panels between the two datasets, we investigate only the 34 genes with significant sc-deQTLs in the OneK1K cohort in the replication cohort. For the meta-analysis of eQTL summary statistics from three subsets, we adopted the inverse-variance-weighted approach and analysed only SNP-gene pairs shared across the three subsets.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

J.E.P. conceived the idea and the project. A.X. performed the computational analysis with the assistance of S.Y., J.A.H., A.C., and A.S. G.G., P.K., and C.J.Y. provided technical assistance for replication analysis. A.W.H. contributes to the data collection and provides constructive comments. A.X. and J.E.P. wrote the manuscript with the participation of all authors. All authors read and approved the final manuscript.

Peer review

Peer review information

Nature Communications thanks Daniel Jordan, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

J.E.P. is supported by a National Health and Medical Research Council Investigator Fellowship (1175781). This work was also supported by the National Health and Medical Research Council Project Grant (1143163) and the Australian Research Council Discovery Project (190100825).

Data availability

All data analysed in the study is available on the Gene Expression Omnibus (GSE196830) and also the HCA data science platform (https://cellxgene.cziscience.com/collections/dde06e0f-ab3b-46be-96a2-a8082383c4a1). The genotype, covariates, pseudobulk matrices eQTL/veQTL/deQTL summary statistics, and updated Seurat object of the raw and normalised single-cell gene expression matrix used in this study are available in the Zenodo platform (https://doi.org/10.5281/zenodo.18910121).

Code availability

The analysis code is available on (https://github.com/powellgenomicslab/sc-veQTL) and (https://doi.org/10.5281/zenodo.20746908)⁶⁵.

Competing interests

No competing interests to declare.

Supplementary information

The online version contains supplementary material available at (https://doi.org/10.1038/s41467-026-77453-9).