Transcriptome-wide association studies at cell-state level using single-cell eQTL data

Summary

Transcriptome-wide association studies (TWASs) are widely used to prioritize genes for diseases. Current methods test gene-disease associations at the bulk tissue or cell-type-specific pseudobulk level, which do not account for the heterogeneity within cell types. We present TWiST, a statistical method for TWAS at cell-state resolution using single-cell expression quantitative trait locus (eQTL) data. Our method uses pseudotime to represent cell states and models the effect of gene expression on the trait as a continuous pseudotemporal curve. Therefore, it allows flexible hypothesis testing of global, dynamic, and nonlinear associations. Through simulation studies and real data analysis, we demonstrated that TWiST leads to significantly improved power compared to pseudobulk methods. Application to the OneK1K study identified hundreds of genes with dynamic effects on autoimmune diseases along the trajectory of immune cell differentiation. TWiST presents great promise to understand disease genetics using single-cell studies.

Graphical abstract

Highlights

• •TWiST captures cell states in single-cell TWAS analysis
• •Outperforms pseudobulk methods in statistical power
• •Flexible tests of global, dynamic, and nonlinear effects
• •Dynamic gene-autoimmune disease association along cell differentiation trajectory

Qi et al. developed a statistical method for transcriptome-wide association studies (TWASs) using single-cell eQTL data. By capturing continuous cell states, the method improves power and provides flexible tests of dynamic patterns. Application to the OneK1K cohort identified dynamic gene-disease associations along the differentiation trajectory of immune cells.

Introduction

Genome-wide association studies (GWASs) have been highly successful in identifying genetic variants associated with complex traits. However, variant-level associations are often difficult to interpret due to linkage disequilibrium, unclear target genes, and a lack of tissue information. Transcriptome-wide association studies (TWASs) are an approach that leverages expression quantitative trait loci (eQTLs) to reveal the molecular mechanisms underlying GWAS associations.^1,2 The statistical procedure of TWAS comprises two steps. First, models are trained to predict genetically regulated gene expression (GReX) from genotype data, using large eQTL studies such as the GTEx Consortium,³ the eQTLGen Consortium,⁴ or Depression Genes and Networks (DGN).⁵ Second, association testing is conducted between GReX and complex traits in a separate GWAS sample. Since its introduction, TWAS has been widely used to prioritize relevant genes and tissues.^6,7,8,9,10

The original methods, PrediXcan¹ and FUSION,² use sparse linear regression to train prediction models. Over the years, other TWAS methods have emerged, introducing innovation in several directions: improved prediction models,^{11,12,13,14,15,16} combining multiple tissues,^17,18,19,20 and multi-ethnic analysis.^21,22 However, almost all existing TWAS methods are based on eQTL studies in bulk tissues. Bulk tissues are composed of a mixture of cell types, and cell-type-specific effects cannot be accurately identified. Recently, the rapid expansion of single-cell eQTL studies^23,24 has enabled the identification of cell-type-specific eQTLs across a wide range of biological contexts, such as peripheral blood mononuclear cells (PBMCs),^25,26 endoderm cells,²⁷ dopaminergic neurons,²⁸ cardiomyocytes,²⁹ and lung tissues.³⁰ Current data analysis pipelines typically adopt pseudobulk aggregation,^{25,26,27,28,29,30} i.e., clustering single cells into cell types and computing the mean expression for each cell type in each individual. Methods for bulk eQTL data can then be used for eQTL mapping and TWAS analysis.³¹ Two recently proposed methods, scPrediXcan³² and Abe et al.,³³ aim at conducting TWAS using single-cell eQTL data. The former leverages deep learning to predict gene expression, and the latter uses time series models to assess Granger causality. However, both methods still rely on pseudobulk aggregation by cell type or experimental time point.

Despite being more informative than bulk eQTL, pseudobulk-based methods do not utilize the full potential of single-cell eQTL data. This is because pseudobulk treats each discrete cell type as a homogeneous group, while many studies have found significant heterogeneity within cell types. For example, in differentiation experiments, induced pluripotent stem cells (iPSCs) undergo continuous transition into terminal cell types.^27,29,34 Among immune cells, there are many cell states representing a spectrum between naive and memory cells.^26,35 Such cell-state heterogeneity can have important roles in disease etiology. For example, a previous study suggested that variants associated with immune diseases are functional in early rather than late activation of memory CD4⁺ T cells³⁶; variants associated with Alzheimer’s disease are enriched in different macrophage cell states.³⁶ Other studies indicated that some genes affect disease risk exclusively at specific stages of T cell stimulation.^37,38 Therefore, characterizing gene effects at specific cell states is important for understanding disease mechanisms and identifying high-resolution drug targets.³⁹

In this paper, we present TWiST (TWAS in pseudotime), a statistical method for TWAS at the resolution of cell states captured by pseudotime.^40,41,42 Using functional regression methods and smoothing splines, TWiST models the effect of gene expression on the trait at each cell state and enables flexible testing of global, dynamic, and nonlinear effects. Applying TWiST to single-cell eQTL data from the OneK1K study,²⁶ we identified novel susceptibility genes and cell states for seven autoimmune diseases, offering new insights into disease mechanisms. As single-cell eQTL datasets continue to expand, TWiST stands as a powerful tool to gain biological insights into the genetic basis of complex traits.

Design

TWiST is a method for single-cell TWAS analysis of heterogeneous cell types, where gene expression and eQTL effects can vary along a continuous cell state (Figure 1A). Within each cell type, the cell state is defined by pseudotime,^40,41,42 which positions cells along a continuous developmental or differentiation trajectory. As in standard TWAS, stage 1 of TWiST is to train a model to predict single-cell gene expression using single-nucleotide polymorphisms (SNPs) in the cis-region of the gene (Figure 1B) for each cell type separately. There are two main challenges unique to single-cell data. (1) The data are highly sparse, such that they cannot be modeled using a normal distribution (as in bulk TWAS) even after transformation. (2) There are many cells per individual that all have the same genotype; hence, the models of existing TWAS methods cannot be applied. To resolve these challenges, we model gene expression counts directly (denoted by x_ij) using a Poisson distribution with mean μ_ij (for cell j of individual i, Figure 1B). Poisson regression has been shown to be a suitable model for sparse single-cell RNA sequencing (scRNA-seq) data.⁴³ Next, we model mean expression μ_ij as a continuous function along pseudotime:

$$\log \left({\mu }_{ij}\right)=\log \left({\mu }_i\left(t_{ij}\right)\right)=c_0+\log \left({\alpha }_{ij}\right)+v_i\left(t_{ij}\right)+{\gamma }^T{\mathit{z}}_{ij}\text{.}$$

Figure 1.

Overview of TWiST

(A) Gene expression and expression quantitative trait locus (eQTL) effect in single-cell eQTL data can be viewed as a continuous function over pseudotime.

(B) In bulk eQTL data, gene expression is a scalar and can be predicted from SNPs using Lasso, elastic net (enet), etc. For single-cell eQTL data, TWiST constructs models that predict continuous curves of gene expression along pseudotime using a function-on-scalar regression model. A scalar-on-function regression model can then be used to estimate the effects of genetically regulated expression (GReX v_i(t)) on complex traits. The estimated effect β(t) is also a continuous curve and can be used for downstream hypothesis tests.

Here, v_i(t_ij) is the GReX, α_ij is the library size, and z_ij is a vector of covariates. GReX v_i(t_ij) is modeled as a linear function of genotypes where the coefficients w_k(t) are smooth B-spline functions: $v_i\left(t_{ij}\right)=\sum _{k=1}^K{w}_k\left(t_{ij}\right)g_{ik}$ (see STAR Methods for details).

In stage 2, we model the trait (y_i) by the aggregated effect of GReX across pseudotime t: y_i = ∫β(t)v_i(t)dt + e_i. The effect of GReX on the trait, β(t), is modeled as a spline function along pseudotime. By default, we use a spline function with a large number of knots to capture flexible curves and apply a penalty $\int {\left|{\beta }^{″}\left(t\right)\right|}^{2}dt$ to encourage smoothness (see STAR Methods). This model can be fitted using individual-level data or GWAS summary statistics. Coefficient β(t) can be used for downstream hypothesis testing of global (the existence of any effect on the trait at any pseudotime point), dynamic (the effect on the trait varies over pseudotime), and nonlinear effects. Details are described in STAR Methods.

Results

Simulation studies

To evaluate the performance of TWiST, we simulated scRNA-seq data and GWAS summary statistics using real genotypes from OneK1K. A time variable that had a broad impact on gene expression was simulated to represent cell states (see STAR Methods for simulation settings). For benchmarking, we aggregated cells into individual-specific pseudobulk samples and used FUSION² and summary data-based Mendelian randomization (SMR)⁴⁴ to conduct TWAS (referred to as “pseudobulk FUSION” and “pseudobulk SMR”). For global tests, cells across all pseudotime values were aggregated into one pseudobulk sample. For dynamic tests, cells were aggregated into two pseudobulk samples: one for the early stage (pseudotime < 0.5) and one for the late stage (pseudotime ≥ 0.5). In the scenarios where the gene effect on the trait (true β(t)) was null (Figure 2A), TWiST had a well-controlled type I error for global, dynamic, and nonlinear tests (Figure 2B). In the scenario of constant effects, where there was a global effect but no dynamic or nonlinear variation, TWiST and pseudobulk FUSION had similar power for the global test and a well-controlled type I error for the dynamic test (Figure 2C). TWiST was the only method that could conduct the nonlinear test and had a well-controlled type I error. Pseudobulk SMR had slightly lower power for the global test and could not be used for other tests (Figure 2C). In the presence of varying effects along pseudotime (unimodal and reverse), TWiST improved statistical power compared to pseudobulk methods (Figures 2D and 2E). In the unimodal scenario, both TWiST and pseudobulk had near-perfect power to detect global effects. For the dynamic test, the power of TWiST was nearly double that of pseudobulk methods when there was a large number (10 or 20) of causal SNPs in the cis-region. Gene effects were approximately the same between the early and late stages on aggregate, leading to low power for dynamic tests for pseudobulk FUSION. In contrast, TWiST was able to capture more complex pseudotemporal patterns using flexible spline functions. The advantage of TWiST was less pronounced when there were only 5 causal SNPs. This is likely due to a lack of degrees of freedom in the estimated v_i(t), which prevents TWiST from identifying complex patterns in β(t). For the nonlinear test, there was virtually no power when there were 10 or fewer causal SNPs. When there were 20 causal SNPs and the true gene effect curve had 3 or 5 knots, TWiST had around 20% power (Figure 2D). The pattern in the reverse scenario was similar, except that the power gain of TWiST was for the global, instead of the dynamic, test (Figure 2E). This could be due to pseudobulk aggregation of the early and late stages already capturing most pseudotemporal variation.

Figure 2.

Type I error and power observed in simulation studies

(A) True gene effect on trait along pseudotime for four scenarios of simulations: null, constant, unimodal, and reverse.

(B) Type I error for global, dynamic, and nonlinear tests in null simulations.

(C) Power of global test and type I error of dynamic and nonlinear tests in the constant scenario.

(D and E) Power in unimodal and reverse scenarios (non-null simulations, type I error cannot be evaluated). Three methods were compared: pseudobulk FUSION (elastic net), pseudobulk FUSION (Lasso), and TWiST. Pseudobulk methods cannot be used for nonlinear tests. The number of causal SNPs in the cis-region of the gene varies among 5, 10, and 20. The true gene effect on trait β(t) is a B-spline function between 0 and 1 with varying numbers of equidistant internal knots: 1, 3, and 5. Blue dashed lines represent the significance threshold, p = 0.05.

TWiST also provided an accurate estimation of the gene expression effect on the trait along pseudotime (Figure S1). In the null and constant scenarios, the average estimated curve across simulations closely tracked the true curve across all combinations of the number of causal SNPs and the true number of knots (Figure S1). The unimodal and reverse scenarios appeared to be more challenging (Figure S2). The curves estimated by TWiST averaged across simulations were smoother than the true curves. However, there was a wide range of variation across simulations, with the estimated curve in some simulations almost a straight line but in other simulations close to the true curve. This is likely due to the uncertainty of estimating the variance parameter (small σ² leads to smooth curves; see STAR Methods for details). However, when restricted to genes with a significant nonlinear test (p < 0.05), the average estimated curve tracked the true effect curve well (Figures S2C and S2D).

When genetic regulatory effects of SNPs on gene expression were constant over pseudotime (see STAR Methods for simulation settings), the type I error of TWiST remained well controlled in the null scenario and for the dynamic and nonlinear tests in the constant scenario (Figure S3). In addition, the power for the global test in the constant and unimodal scenarios remained high. However, TWiST was no longer able to detect dynamic or nonlinear effects (unimodal and reverse). The global test for the reverse scenario also had no power (Figure S3). The estimated curve was almost always a straight line, even in the unimodal and reverse settings (Figure S4). This observation indicates that pseudotemporal variation in genetic regulatory effects is crucial for identifying dynamic gene effects on the trait across cell states.

Training prediction models on the OneK1K data

The OneK1K cohort consists of predominantly European ancestry individuals in Hobart, Australia. Genotype and scRNA-seq data for PBMCs were available from 981 individuals.²⁶ When training models to predict GReX from SNPs, we focused on three major immune cell types (Figure 3A): CD4⁺ T cells (naive and central memory), CD8⁺ T cells (naive and central memory), and B cells (naive, transitional, and memory). We aggregated 5 cells on average into metacells to reduce the computational burden (STAR Methods). Cell-type-specific pseudotime was learned from the metacells using Slingshot,⁴¹ which captures a continuum of cell states (Figure 3B). After gene filtering (see STAR Methods for details), we trained a prediction model for each gene and cell type using the metacells, retaining genes with non-zero coefficients for at least 3 SNPs for downstream association analysis (3,311 genes for CD4⁺ T cells, 1,399 genes for CD8⁺ T cells, and 2,471 genes for B cells). Model training was computationally intensive in both memory usage and runtime (Table S1). Averaged across genes, training a prediction model for CD4⁺ T cells required 19 GB of memory and 1.2 h per gene (Table S1). The equivalent memory usage was 2.4 GB for CD8⁺ T cells and 4.9 GB for B cells, and the runtime was 0.48 h for CD8⁺ T cells and 0.55 h for B cells, which was less demanding than CD4⁺ T cells due to lower cell abundance. However, these models only need to be trained once on the single-cell eQTL data and can be applied to many diseases downstream.

Figure 3.

Prioritizing autoimmune disease genes in three cell types

(A) Uniform manifold approximation and projection (UMAP) of original scRNA-seq UMI count data from OneK1K with CD4⁺ T cells, CD8⁺ T cells, and B cells highlighted. Other cell types are colored in gray.

(B) Metacell principal-component analysis (PCA) plots for three cell types colored by pseudotime (pseudo).

(C) QQ plot for TWiST global test versus pseudobulk FUSION (referred to as pseudobulk in figure) and scPrediXcan in naive and memory cells for seven autoimmune diseases: rheumatoid arthritis (RA), systemic lupus erythematosus (SLE), Crohn’s disease (CD), inflammatory bowel disease (IBD), multiple sclerosis (MS), type 1 diabetes mellitus (T1DM), and ankylosing spondylitis (AS).

(D) Number of identified genes at a false discovery rate (FDR) < 0.05 by three tests.

We evaluated the prediction accuracy of the TWiST stage 1 model by computing the prediction R² in subsamples with varying numbers of individuals: n = 100, 300, 500, or 900. Although our model is Poisson based, a correlation can be computed between the log-transformed normalized gene expression and the GReX (see STAR Methods for details). This correlation can be defined at the cell level or the individual level by averaging the GReX and log-transformed normalized gene expression across cells within each individual. We restricted the analysis to CD4⁺ T cells due to the intensive computation of re-training many prediction models. Although the cell-level prediction R² of TWiST was low, the individual-level R² was comparable with that of pseudobulk FUSION (Figure S5). It is possible that the noisiness of scRNA-seq data makes it difficult to predict at the cell level, but noise is reduced by pooling information across cells. The average R² increased with the sample size of the training data, with a mean of 0.031 for a subsample of n = 100, 0.041 for n = 300, 0.039 for n = 500, and 0.045 for n = 900. This indicates that TWiST can be used on smaller single-cell eQTL studies of 100–300 individuals. These values are lower than the cross-validation R² of pseudobulk FUSION (mean = 0.077), possibly due to R² being a more favorable metric to the linear model adopted by FUSION, while TWiST minimizes a different objective function defined by the Poisson model.

Association analysis for autoimmune diseases

We conducted association analyses using GWAS summary statistics for seven autoimmune diseases that were also analyzed by the OneK1K paper²⁶: rheumatoid arthritis (RA),⁴⁵ systemic lupus erythematosus (SLE),⁴⁶ Crohn’s disease (CD),⁴⁷ inflammatory bowel disease (IBD),⁴⁷ multiple sclerosis (MS),⁴⁸ type 1 diabetes mellitus (T1DM),⁴⁹ and ankylosing spondylitis (AS).⁴⁹ We applied TWiST to each disease and the three cell types discussed above. Association analysis was computationally efficient, requiring <0.5 s per gene (Table S1). To benchmark the power of TWiST, we compared it with a standard TWAS (FUSION²) applied to pseudobulk samples of naive or memory cells and the recently developed scPrediXcan. We did not compare SMR in the real data due to its similarity to pseudobulk FUSION in the simulation studies. The QQ plot of TWiST revealed greater signal enrichment than that of pseudobulk TWAS or scPrediXcan across nearly all diseases and cell types, indicating higher statistical power (Figure 3C). The gain was more pronounced for RA, SLE, MS, and T1DM but was also notable for other diseases. Using TWiST, we identified hundreds of genes for which the expression in immune cells was associated with the risk of autoimmune diseases (Figure 3D). Among the three cell types, the largest number of associations was detected for CD4⁺ T cells, with 145 genes for RA (22,350 cases and 74,823 controls), 160 genes for MS (47,429 cases and 68,374 controls), and 106 genes for IBD (6,968 cases and 21,770 controls) showing global association (Figure 3D). In addition, dozens of genes were detected for other diseases with smaller sample sizes (see STAR Methods for sample size information).

TWiST identified a large number of genes with dynamically varying effects on autoimmune diseases along the trajectory from naive to memory CD4⁺ T cells, with the most associations detected from RA (82), MS (78), and SLE (35) (Figure 3D; see Tables S2, S3, and S4 for full results). Similar patterns were also observed for CD8⁺ T cells and B cells, though fewer genes were detected, which was likely due to a smaller number of cells limiting our ability to train accurate prediction models. To obtain stronger evidence for the dynamic genes, we conducted colocalization analysis in pseudobulk samples created by dividing cells into two groups: early (pseudotime < 0.5) and late (pseudotime ≥ 0.5) (see STAR Methods for details). Among the 165 gene-disease-cell type trios that were colocalized (posterior probability of H4, or PPH4 > 0.8) in at least one cell group (early/late), colocalization occurred mostly at the late stage, confirming dynamic effects that were only present in specific cell states (Figure S6). Similar results were observed in colocalization analysis based on naive versus memory cell labels provided by OneK1K (Figure S6). In addition, TWiST detected genes with nonlinearly varying effects from naive to memory cells. The smaller number of genes with nonlinear effects indicated increasing challenges to detect higher-order effects. Many of the strongest signals were found among the HLA genes (Figures 4, S7, and S8). Many other genes near HLA genes (defined as 26–34 Mb on chromosome 6⁵⁰) also had strong dynamic association with autoimmune diseases, which were often stronger than the HLA genes themselves. For RA, MS, T1DM, and AS, the signal in HLA and nearby genes had much stronger levels of significance than genes found on other chromosomes. This was also true for SLE in T cells (Figures 4 and S7). However, in B cells, PYCARD had the strongest dynamic association with SLE (p = 6.12 × 10⁻⁴³; Figure S8; Table S4). For CD and IBD, the genes with the strongest association were outside the HLA region. For CD4⁺ T cells, BRD7 (p = 2.59 × 10⁻²⁰) had the strongest dynamic association with CD and PTGER4 (p = 3.02 × 10⁻¹¹) with IBD (Figure 4). Such patterns were also observed for CD8⁺ T cells and B cells (Figures S7 and S8).

Figure 4.

Manhattan plot for dynamic tests in CD4⁺ T cells for seven autoimmune diseases

Top significant genes inside and outside the HLA region (26–34 Mb on chromosome 6) are annotated by gene name and p value for the TWiST dynamic test. HLA genes are colored in light green. The red dashed line represents the raw p value threshold corresponding to an FDR < 0.05. See Figures S7 and S8 for Manhattan plots for CD8⁺ T cells and B cells.

To evaluate whether the genes identified by TWiST represent novel findings, we compared them with previously reported TWAS genes cataloged by the TWAS Atlas,⁵¹ as well as genes identified by pseudobulk FUSION for corresponding cell types in our dataset (see STAR Methods for details). Among the six autoimmune diseases (excluding AS due to the small number of findings), the percentage of novel findings among genes identified by TWiST ranged from 11% to 61% (Table 1). The largest number of novel genes was found for MS, with 61–69 novel genes with global effects and 33–37 novel genes with dynamic effects on the disease. Fewer novel genes were found for T1DM and AS, with <10 across all cell types and tests.

Table 1.

Number of novel genes discovered by TWiST across three cell types and seven autoimmune diseases

Cell type	Test	RA	SLE	CD	IBD	MS	T1DM	AS
T_CD4	global	54 (145)	23 (86)	21 (78)	33 (106)	61 (160)	9 (43)	1 (16)
T_CD4	dynamic	30 (82)	8 (35)	7 (18)	9 (27)	34 (78)	4 (26)	0 (8)
T_CD8	global	29 (75)	28 (59)	14 (51)	28 (72)	63 (104)	6 (27)	3 (13)
T_CD8	dynamic	13 (35)	10 (26)	8 (21)	18 (31)	33 (55)	2 (18)	3 (12)
B	global	32 (101)	29 (86)	22 (71)	46 (114)	69 (142)	4 (36)	2 (10)
B	dynamic	19 (68)	15 (48)	9 (19)	17 (38)	37 (78)	4 (24)	1 (6)

Each cell is the number of novel genes (with the total number of significant genes in parentheses) found by TWiST.

Gene set enrichment for dynamic genes

For each disease, we combined the dynamic genes identified across the three cell types (Figure 5A) and conducted gene set enrichment analysis (GSEA) using Gene Ontology (GO).^52,53,54 The number of GO biological process pathways enriched in dynamic-disease-associated gene sets ranged from 30 to 203, with the largest number of pathways identified for RA, SLE, and MS (Figure 5B). Most of the enriched pathways were related to immune system processes (descendants of GO: 0002376). While the number of enriched pathways was correlated with the number of genes, CD and IBD appeared to be enriched in fewer pathways than other diseases with similar numbers of dynamically associated genes (Figure 5B). The 52 dynamic genes for CD were enriched in 30 pathways, substantially lower than the 117 pathways enriched in the 48 dynamic genes for T1DM. Similarly, the 82 dynamic genes for IBD were enriched in 81 pathways, fewer than the 140 pathways for the 82 dynamic genes for SLE. The proportion of immune pathways was also lower for CD and IBD than for other genes. This could be related to the different genetic architecture of CD and IBD, for which many of the strongest signals were found outside of HLA. For all six diseases (AS was not enriched in any pathway due to the small number of genes), the fold enrichment in immune pathways was substantially larger than that in non-immune pathways (Figure 5C, see the figure legend for a definition of fold enrichment).

Figure 5.

Gene set enrichment analysis of dynamic genes

(A) Number of dynamic genes across three cell types at FDR < 0.05 for TWiST.

(B) Gene Ontology (GO) biological process pathways enriched in the dynamic genes for seven autoimmune diseases at FDR < 0.05. The number of immune and non-immune pathways is annotated on each bar.

(C) Boxplot of log fold enrichment for immune (descendants of GO: 0002376) versus non-immune pathways. Fold enrichment is defined as the percentage of dynamic genes for each trait belonging to a pathway, divided by the corresponding percentage in the background.

(D) Top 30 enriched pathways defined by smallest minimum enrichment p value across seven diseases. See Table S5 for the remaining enriched pathways.

The top pathways appeared to show similar patterns for RA, SLE, MS, and T1DM (Figure 5D). The strongest enrichment was found in pathways related to antigen processing and presentation and to major histocompatibility complex (MHC) protein complex assembly. Enrichment was also found in pathways related to immune cell activation, cytotoxicity, immune effector process, etc. Although CD dynamic genes were not enriched in many of the top pathways, which was likely due to the small number of genes, IBD genes were enriched in many of the pathways (Figure 5D). In addition to the top pathways, dynamic genes were also enriched in many other pathways, such as T cell activation, differentiation, and proliferation (Table S5). Many of the pathways remained significantly enriched after replacing the background gene set with the top 1,000 highly expressed genes in PBMCs, demonstrating the robustness of the results (Figure S9). For five of the seven diseases (RA, SLE, IBD, MS, and T1DM), our analysis identified novel pathways that cannot be identified by GSEA using previously known genes (Table S5).

We further investigated the functions of dynamic disease genes that showed different pseudotemporal patterns. We chose to focus on MS for this analysis because it had the largest number of dynamic genes among the diseases we analyzed. The curve of the gene expression effect on MS was estimated using TWiST and clustered into six groups (Figure 6A). Clustering was performed by combining all cell types (see STAR Methods for details). Most of the genes fell into two clusters (Figure 6B): monotonically increasing (cluster 3) and monotonically decreasing (cluster 5). Other genes fell into clusters with more complex patterns, especially cluster 2, which appeared to suggest that there are intermediate cell states where higher expression leads to higher MS risk, while early and late cell states either had no effects or effects in the opposite direction. Sensitivity analysis showed that setting the number of clusters to 6 led to a higher silhouette score than fewer clusters and only slightly lower than 7 or 8 clusters (Table S6). We chose 6 clusters in favor of a more parsimonious model. GSEA identified differential enrichment patterns between clusters 3 and 5 (Figure 6C). Dynamic genes for CD4⁺ T cells in cluster 5 were enriched in pathways related to T cell differentiation and activation, cell-cell adhesion, the immune effector process, and the adaptive immune response. Genes in cluster 3 were not enriched in any pathway despite a larger number of genes. Combining clusters 3 and 5 led to 33 enriched pathways, which was several-fold more than the number of pathways enriched in cluster 5 alone (Table S7). Many of the pathways were related to immune functions. These results indicated that cluster 3 was likely to include genes in relevant pathways but lacked statistical power.

Figure 6.

Dynamic patterns of gene-disease effects for dynamic genes of multiple sclerosis

(A) Six clusters of genes based on k means clustering of estimated effect curve β(t). Thin lines represent individual genes, and thick lines represent the average across genes in the cluster.

(B) Number of genes in each cluster for three cell types.

(C) Enrichment of genes in clusters 3 and 5 in Gene Ontology (GO) biological process pathways. Fold enrichment is defined as the percentage of dynamic genes for each trait belonging to a pathway, divided by the corresponding percentage in the background. Only pathways with FDR-adjusted enrichment p < 0.05 are included.

For CD8⁺ T cells, genes in cluster 3 were enriched in 19 pathways (Figure 6), while genes in cluster 5 were not enriched in any pathway. Similarly, combining clusters 3 and 5 increased the number of enriched pathways to 52 (Table S8), indicating that cluster 5 harbored relevant genes but lacked statistical power for enrichment, possibly due to the small number of genes (n = 10). For B cells, dynamic genes in both clusters were enriched in pathways related to antigen processing and presentation, with cluster 5 enriched in those for exogenous antigen and cluster 3 enriched in those for endogenous antigen or via MHC class I (Figure 6). In addition, cluster 3 was enriched in genes related to cytotoxicity, while cluster 5 was enriched in genes related to immune cell activation. These observations indicated that genes with different pseudotemporal patterns could affect the disease through different biological mechanisms.

Discussion

We presented TWiST, a statistical method for single-cell TWAS along pseudotime. TWiST uses functional regression to model the effect of genes on diseases and produces hypothesis tests for multiple pseudotemporal patterns. Using TWiST, we identified hundreds of genes whose expression in immune cells is associated with autoimmune diseases, and the association strength could vary as the immune cells transitioned from naive to memory states. These genes were enriched in well-characterized immune pathways from GO. Finally, we identified clusters of genes exhibiting distinct dynamic patterns with MS, each associated with unique gene set enrichment profiles.

TWiST has several advantages over existing methods. Unlike pseudobulk methods that treat each cell type as homogeneous, TWiST captures the heterogeneous cell states within a cell type. Capturing such heterogeneity led to power gain for identifying disease-associated genes (Figure 3). In addition, TWiST provides dynamic and nonlinear tests that are not available in other TWAS methods. Dynamic genes could have important functions in the immune response, which is a dynamic process that involves many cell-state transitions. Finally, the effect of the gene on the trait estimated by TWiST at each cell state is adjusted for other cell states. The stage 2 model is effectively a continuous version of multiple regression with gene expression at all pseudotime points jointly as predictors. Therefore, the β(t) we estimate at each pseudotime point is free of contamination from other points and thus closer to the causal effect.

Other than immune cells, our method is particularly suitable for single-cell TWAS using cell-line differentiation data. Studies have demonstrated that stem cell differentiation is a continuous process^27,29,34 and have used pseudotime bin methods to conduct dynamic eQTL mapping.^29,34 Our method can advance the analysis to a finer resolution by modeling genetic regulatory effects and gene-trait effects using smooth B-spline functions. In addition, our method does not rely solely on pseudotime but can also be applied to real time, which is often available in cell-line differentiation studies. Applying TWiST to differentiation experiments can potentially identify novel genes and cell states associated with diseases. The main challenge is that TWAS analysis requires a large sample size to train prediction models. To date, the sample size for single-cell eQTL studies for differentiation experiments ranges from dozens to around 200,^{27,28,29,30,34} which is several times lower than that of OneK1K but could be sufficient for detecting strong signals. As single-cell eQTL studies continue to expand, the application of TWiST to differentiation experiments can be highly fruitful in identifying novel disease genes and cell states.

Limitations

Our method has a few limitations. First, TWiST can lead to no power gain or even slight power loss compared to pseudobulk in some scenarios, as evidenced in the simulation studies (Figure 2). In those scenarios, the gene-trait effect either had no pseudotemporal variation (Figure 2C) or the variation could be adequately captured by pseudobulk (Figures 2D, global, and 2E, dynamic). For the unimodal scenario, the gene effect on the trait was preserved when aggregated across the entire pseudobulk domain for global tests. However, when two pseudobulk samples were created by pseudotime <0.5 and pseudotime ≥0.5 for dynamic tests, the two samples carried the same amount of signal, and dynamic patterns could not be detected. Similarly, for the reverse scenarios, positive and negative gene-trait effects were canceled when aggregated across the entire pseudotime domain. However, when two pseudobulk samples were created by pseudotime <0.5 and pseudotime ≥0.5 for dynamic tests, the difference between the pseudobulk samples was preserved. The advantage of TWiST is its robust performance across all scenarios. The power loss was generally small and could be due to model misspecification or an increased number of parameters. When continuous cell states played an important role in the analysis (Figures 2D, dynamic, and 2E, global), TWiST led to a large power gain compared to pseudobulk. It allows an additional nonlinear test that cannot be achieved by pseudobulk. To capitalize on the power of TWiST, we recommend assessing whether heterogeneous cell states are likely to exist in a cell type and carefully constructing pseudotime to reflect actual biological cell states. Second, training the stage 1 model to predict gene expression from SNPs is computationally intensive (Table S1). This is due to the large number of predictors, which is the number of cis-SNPs multiplied by the number of spline bases. This issue is made worse by the large number of cells and the need to fit the model for each gene in each cell type. In the analyses conducted in this paper, we created metacells, which were composed of 5 cells on average. This reduced the number of cells to around 20% of the original number. In addition, we used a small number of knots (0, 0.25, 0.5, 0.75, and 1) to model SNP effects on gene expression in the stage 1 model. This approach reduced computational cost but restricted the ability to model more complex curves. It is possible to use more knots if more computational resources are available. Third, our current method assumes that each cell type only has one lineage. This is a reasonable assumption for our application in immune cells but may not be generalizable to other datasets. For example, cells often have bifurcating or trifurcating trajectories in a differentiation experiment, creating multiple lineages that begin with the same early-stage cells but end with different terminal cells. Therefore, pseudotime will be lineage specific, such that the same value does not necessarily represent the same cell state across lineages. The current version of TWiST can only analyze each lineage separately and does not allow information sharing across lineages, which may lead to a loss of efficiency. Lastly, the curve we estimate can be sensitive to the variance parameter in the stage 2 model. Therefore, we recommend using TWiST primarily for hypothesis testing. Although the estimated curve is provided, we recommend only using it to conduct gene clustering and advise against interpreting the curve at each pseudotime point. Future research will be dedicated to addressing these limitations, i.e., accelerating computation, accounting for multiple lineages, and generating more accurate estimates of the pseudotemporal curve. With accelerated computation, it is possible to use more flexible curves to model genetic effects on gene expression, potentially improving statistical power.

In summary, TWiST is a powerful method for conducting single-cell TWAS at cell-state resolution. TWiST identified hundreds of genes that are dynamically associated with autoimmune diseases along the trajectory of immune cell differentiation. TWiST will be a promising tool for gaining biological insights from the rapidly growing single-cell eQTL data.

Resource availability

Lead contact

Requests for further information and resources should be directed to the lead contact, Guanghao Qi (gqi@uw.edu).

Materials availability

This study did not generate new material.

Data and code availability

• •The TWiST R package and pre-trained prediction models are publicly available on GitHub (https://github.com/gqi/TWiST) and Zenodo (https://doi.org/10.5281/zenodo.17228167).
• •Secondary datasets analyzed in this study are available from the following sources. scRNA-seq and genotype data of OneK1K are available via Gene Expression Omnibus (GEO: GSE196830). Single-cell gene expression data are also available on the Human Cell Atlas: https://cellxgene.cziscience.com/collections/dde06e0f-ab3b-46be-96a2-a8082383c4a1. Processed genotype data were provided by the authors of the OneK1K eQTL paper.²⁶ GWAS summary statistics are publicly available via the links provided by the OneK1k eQTL paper.

Acknowledgments

G.Q. was supported by NIH/NHGRI award K01HG013983. A.B. was supported by NIH/NIGMS award R35GM139580.

Author contributions

G.Q. conceived the project and conducted the method development, simulation studies, and data analysis. E.L. refined the statistical model. Z.J. and W.S. provided advice on single-cell data analysis. E.L., A.S., and W.S. provided input on the statistical methods. A.B. provided access to processed genotype data. G.Q. drafted the paper. E.L., A.S., A.B., and W.S. edited the paper. All authors have read and approved the manuscript submission.

Declaration of interests

A.B. is a co-founder and equity holder of CellCipher, Inc., a stockholder in Alphabet, Inc.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Deposited data
Single-cell RNA-seq and genotype data from the OneK1K study	Gene Expression Omnibus	GEO: GSE196830
Processed OneK1K single-cell RNA-seq data	Chan Zuckerberg CELL by GENE Discover	https://cellxgene.cziscience.com/collections/dde06e0f-ab3b-46be-96a2-a8082383c4a1
GWAS summary statistics of 7 autoimmune diseases	Yazar et al.26	see resource availability
Software and algorithms
TWiST	This manuscript	https://github.com/gqi/TWiST
FUSION	Gusev et al.2	http://gusevlab.org/projects/fusion/
scPrediXcan	Zhou et al.32	https://predictdb.org/
FUMA GWAS	Watanabe et al.55	https://fuma.ctglab.nl/
SMR	Zhu et al.44	https://yanglab.westlake.edu.cn/software/smr/#Overview
Coloc	Giambartolomei et al.56	https://cran.r-project.org/web/packages/coloc/index.html
Gene Ontology	The Gene Ontology Consortium52,53,54	https://geneontology.org/

Method details

Stage 1 of TWiST: Predicting gene expression from SNPs

The following model aims to analyze a group of cells on a continuous trajectory, e.g., immune cells that transition from naive to memory cells during an immune response, and induced pluripotent stem cells (iPSCs) differentiating into terminal cell types. We assume that pseudotime has been constructed using methods such as Slingshot,⁴¹ TSCAN,⁴⁰ etc. We only consider pseudotime of one lineage and different lineages will be analyzed separately. To ease interpretation, we convert the pseudotime to its rank, and divide it by the number of cells, such that pseudotime is approximately uniformly distributed between 0 and 1. For a gene of interest, denote by x_ij the unique molecular identifier (UMI) count of the gene in individual i and cell j (i = 1, …,I;j = 1, …,J_i). Denote by z_ij cell covariates that typically include age, sex, expression principal components (PCs), and genotype PCs. Denote by t_ij the pseudotime of this cell and α_ij the library size. Denote the genotype of individual i and cis-SNP k by g_ik. Assume that the genotypes are standardized to have mean 0 and variance 1.

Analogous to standard TWAS using bulk eQTL data, the first step of our method is to train statistical models to predict gene expression from genotypes. However, in contrast to the bulk eQTL setting, single-cell eQTL data contain multiple cells per individual. We model gene expression as a function of pseudotime and link it to SNPs using a function-on-scalar Poisson regression model:

$$\begin{gathered} x_{ij}\sim Poisson\left({\mu }_{ij}\right), \\ \log \left({\mu }_{ij}\right)=\log \left({\mu }_i\left(t_{ij}\right)\right)=c_0+\log \left({\alpha }_{ij}\right)+v_i\left(t_{ij}\right)+{\gamma }^T{\mathit{z}}_{ij}\mathbf{,} \\ {v}_i\left(t_{ij}\right)=\sum _{k=1}^K{w}_k\left(t_{ij}\right)g_{ik}, \\ {w}_k\left(t\right)=\sum _{m=1}^M{w}_{km}{\varphi }_m\left(t\right). \end{gathered}$$ (Equation 1)

Here v_i(t_ij) is the genetically regulated gene expression (GReX), γ is the vector of regression coefficient for covariates, and c₀ is the intercept. We model the functional coefficients w_k(t_ij) using B-spline functions, where ϕ_m(t),m = 1, …,M are B-spline bases and w_km’s are the “weights.” Define ${\mathit{w}}_k={\left(w_{k1},\dots ,w_{kM}\right)}^T$. We use a group lasso regression to fit the model, with the loss function being

$$l\left({\mathit{w}}_1,\dots ,{\mathit{w}}_K,c_0,\gamma |\mathit{x}\right)+\lambda \sum _{k=1}^K{\sqrt{M}\left|\left|{\mathit{w}}_k\right|\right|}_2\text{.}$$

Here l(w₁, …,w_K,c₀,γ|x) is the likelihood of the Poisson regression described in model (1). The group lasso penalty on w_k ensures that only a small number of cis-SNPs have non-zero effects on gene expression, consistent with bulk TWAS models.^1,2 We assemble the weight coefficients into a K×M matrix

$$W=\left[\begin{array}{ccc}{w}_{11}& \dots & w_{1M}\\ \dots & \dots & \dots \\ w_{K1}& \dots & w_{KM}\end{array}\right]$$

for the description of the stage-2 model.

Although some studies used the negative binomial distribution to model scRNA-seq data,^57,58 recent studies found that Poisson distribution is adequate for modeling many scRNA-seq datasets.^43,59,60 Modeling overdispersion by negative binomial is more important for hypothesis testing since ignoring it can lead to false positives. However, the main objective of our stage-1 model is to predict gene expression, for which overdispersion is less important. Therefore, we chose the Poisson model for simplicity and computational efficiency.

Stage 2 of TWiST: Estimating effect of gene expression on traits

In a second sample for which both genotype and phenotype data are available, we model the trait using a scalar-on-function regression

$$y_i=\int \beta \left(t\right)v_i\left(t\right)dt+e_i\text{.}$$ (Equation 2)

Both the trait y_i and the genotypes g_i are standardized to have mean 0 and variance 1, hence an intercept is not included in the regression. Here e_i captures observational noise. The GReX v_i(t) is not directly observed but predicted using the model trained in stage 1. We model gene effect β(t) on trait as the sum of a linear function and a B-spline function $\beta \left(t\right)={\beta }_{l0}+{\beta }_{l1}t+\sum _{m=1}^{M^{\prime }}{\beta }_m{\psi }_m\left(t\right)$, where {ψ_m(t)} can be a different set of basis functions than those used in stage 1. Since β(t) is the main parameter of interest, we use a large number of B-spline bases (M′>M) to allow greater flexibility. We separate the linear function from B-spline to allow hypothesis testing of nonlinear effects. Specifically, we define {ψ_m(t)} as the output of bs( …, intercept = TRUE) in R package splines, excluding the first and last B-spline basis functions. The remaining B-spline bases, along with the intercept and linear term t, form a linearly independent set of bases. Plugging in model (1), we can express model (2) as

$$y_i=\int \left({\beta }_{l0}+{\beta }_{l1}t+\sum _{m=1}^{M^{\prime }}{\beta }_m{\psi }_m\left(t\right)\right)\left(\sum _{k=1}^K\sum _{m=1}^M{w}_{km}{\varphi }_m\left(t\right)g_{ik}\right)dt+e_i={\mathit{\beta }}_l^T{\mathrm{\Omega }}_{\mathrm{l}}^T{W}^T{\mathit{g}}_i+{\mathit{\beta }}^T{\mathrm{\Omega }}^T{W}^T{\mathit{g}}_i+e_i\text{.}$$ (Equation 3)

Here ${\mathit{\beta }}_l={\left({\beta }_{l0},{\beta }_{l1}\right)}^T$, $\mathit{\beta }={\left({\beta }_1,\dots ,{\beta }_{M^{\prime }}\right)}^T$, and ${\mathit{g}}_i={\left(g_{i1},g_{i2},\dots ,g_{iK}\right)}^T$ represents the vector of genotypes for cis-SNPs. In addition, Ω_l is a M × 2 matrix of inner products of basis functions, where entry (m,1) is ∫ϕ_m(t)dt and element (m,2) is ∫tϕ_m(t)dt. Similarly, Ω is a M×M′ matrix of inner products of B-spline bases, where entry (m,m′) is ${\mathrm{\Omega }}_{m{m}^{\prime }}=\int {\varphi }_m\left(t\right){\psi }_{m^{\prime }}\left(t\right)dt$.

In many scenarios, individual-level genotypes in GWAS are not available. However, model (3) implies the following model for GWAS summary statistics (see supplemental methods for details):

$$\hat{\mathit{b}}\sim N\left(RW{\mathrm{\Omega }}_{\mathrm{l}}{\mathit{\beta }}_l+RW\mathrm{\Omega }\mathit{\beta },R/N\right)\text{,}$$ (Equation 4)

where R is the pairwise linkage disequilibrium (LD) matrix for cis-SNPs (element r_pq = cor(g_ip,g_iq)) and N is the GWAS sample size. Again, both the trait y_i and the genotypes g_i are standardized to have mean 0 and variance 1. An underlying assumption is that only a small proportion of trait variance is explained by the GReX of this gene.

We assume that the effect of gene expression β(t) varies smoothly over pseudotime t. To enforce this, we apply the smoothness penalty $\int {\left|{\beta }^{″}\left(t\right)\right|}^{2}dt={\mathit{\beta }}^{\mathit{T}}{\mathrm{\Omega }}_2\mathit{\beta }$, where the entries of Ω₂ are inner products of second derivatives of B-spline basis functions: ${\mathrm{\Omega }}_{2,m{m}^{\prime }}=\int {\psi }_m^{″}\left(t\right){\psi }_{m^{\prime }}^{″}\left(t\right)dt$. This is equivalent to a random-effects model where

$$\mathit{\beta }\sim N\left(0,{\sigma }^2{\mathrm{\Omega }}_2^{-1}\right)\text{.}$$

We use maximum likelihood to estimate the parameters β_l and σ². To address singularity caused by perfect LD among SNPs, we conduct eigendecomposition of R and decorrelate summary statistics $\hat{\mathit{b}}$. The likelihood based on decorrelated summary statistics is then maximized to estimate the parameters (see supplemental methods for details).

• We test three null hypotheses to identify genes with pseudotemporal patterns:

• •Global test (no effect at any pseudotime point): H₀₁: β(t) = 0 for all t (i.e., β_l0 = 0,β_l1 = 0, σ² = 0)
• •Dynamic test (constant effect along pseudotime): H₀₂: β(t) = β_l0 for all t (i.e., β_l1 = 0, σ² = 0)
• •Nonlinear test (nonlinear effect along pseudotime): H₀₃: β(t) = β_l0+β_l1t (i.e., σ² = 0)

Since the null is at the boundary of the parameter space, it has been shown that the null distributions for the three tests are a mixture of chi-squared distributions⁶¹: (1) global test: $0.95{\chi }_2^2+0.05{\chi }_3^2$; (2) dynamic test: $0.95{\chi }_1^2+0.05{\chi }_2^2$; (3) nonlinear test: $0.95{\delta }_0+0.05{\chi }_1^2$ (δ₀ is point mass at 0).

In addition, we estimate gene effect on trait β(t) using best linear unbiased predictors (BLUP) and obtain confidence bands. See supplemental methods for details.

Simulation studies

We simulated gene expression from model (1) using real genotype data of chromosome 1 from the OneK1K study (n = 1,033, see Processing of genotype data for description of the data). We defined the cis-SNPs of a gene as those within 500kb from the transcription start site (TSS). We randomly selected 200 genes with at least 300 cis-SNPs. For each gene, we randomly chose a small number (5, 10, or 20) of cis-SNPs as causal SNPs for gene expression. The remaining cis-SNPs were assumed to have no effect on gene expression. This setting is motivated by the number of SNPs in existing TWAS models. For example, in the FUSION² elastic net model trained on GTEx v8 whole blood samples (http://gusevlab.org/projects/fusion/), the number of SNPs with non-zero weight averages 20 per gene, with a standard deviation (SD) of 13. In the FUSION lasso model, the average is 6.3 with an SD of 3.5. Next, we simulated the effect of causal SNPs on gene expression that vary over time t. We first simulated a time variable t∼Uniform[0,1] for 50,000 cells. The number of cells was in the range of the number of metacells in our OneK1K analysis (see Preprocessing of gene expression data). We then generated cubic B-spline basis functions with 0 and 1 as boundary knots and a varying number of equidistant internal knots (1, 3, or 5). For one internal knot, the only knot was 0.5; for 3 internal knots, the knots were 0.25, 0.5, and 0.75; for 5 internal knots, the knots were 0.16, 0.33, 0.5, 0.67, and 0.83. This step was implemented using the bs() function from splines package in R, with degree = 3 and intercept = TRUE. This procedure was repeated for each causal variant. Denote the m-th B-spline basis by ϕ_m(t). For causal variant k, the weight of the m-th B-spline basis was generated as $w_{km}={\tilde{w}}_k+{\tilde{w}}_{km}$. Here ${\tilde{w}}_k\sim N\left(0,\text{var}=0.01\right)$ is the SNP-specific effect, and ${\tilde{w}}_{km}\sim N\left(0,\text{var}=0.01\right)$ is the effect specific to the SNP and the m-th B-spline basis. The genetic regulatory effect of SNP k was then calculated as $w_k\left(t\right)={\sum }_m{w}_{km}{\varphi }_m\left(t\right)$, and the GReX calculated as $v_i\left(t_{ij}\right)={\sum }_k{w}_k\left(t_{ij}\right)g_{ik}$ (see model 1 for notations). All cells were assumed to have the same library size, which was hence not included in the model. Finally, we simulated gene expression from a Poisson distribution with mean exp(v_i(t_ij)+t_ij), where t_ij was added to reflect the effect of time on gene expression that is independent of genotypes.

The next step was to generate GWAS summary statistics for complex traits. Since GWAS sample size (typically 10ks or 100ks) are usually many times larger than an eQTL study, simulating single-cell gene expression data for a GWAS cohort is extremely time consuming. Therefore, we directly simulated GWAS summary statistics from a simplified version of model (4):

$$\hat{\mathit{b}}\sim N\left(RW\mathrm{\Omega }\mathit{\beta },R/N\right)\text{,}$$

The LD matrix R was calculated from the genotype data. We used the same set of B-spline for bases for β(t) as that of w_k(t), hence Ω was a squared matrix of inner product of B-spline bases that was calculated using the fda package in R. Here we did not separate the intercept and linear terms of β(t), but instead used the original set of bases generated by bs(). We set the GWAS sample size to n = 100k. Next, we generated β=(β₁, …,β_mid-1,β_mid,β_mid+1, …,β_M) in four scenarios:

• •Null: β₁ = … = β_M = 0
• •Constant: β₁ = … = β_M = 0.05
• •Unimodal: β_mid = 0.2, β_mid-1 = β_mid+1 = 0.1, and β_m = 0 for other m.
• •Switch (effect changes direction): β_mid-1 = 0.3, β_mid+1 = −0.3, and β_m = 0 for other m

These parameters were chosen such that power was the range of 10–75% for many scenarios, allowing comparison across methods. Here we only simulated one gene and its effect on the complex trait in each experiment, since TWAS analyzes one gene at a time. Although other genes may have effects on the trait, they can be considered part of the random error and do not affect the analysis of the current gene of interest. For each combination of the number of causal SNPs and the true number of internal knots (3 × 3 = 9 combinations), we repeated the simulation three times such that there were in total 200 × 3 = 600 genes for each scenario.

To analyze the simulated data, an empirical pseudotime was learned as the first principal component (PC) of log-transformed gene expression. It was then converted to its rank, and divided by the number of cells, such that it is approximately uniformly distributed between 0 and 1. To train stage-1 prediction models using TWiST, we used cubic B-splines with 3 internal knots (0.25, 0.5, 0.75) across all genes. For estimating β(t), we used cubic B-splines with 19 internal knots (0.05, 0.1, …, 0.9, 0.95) to allow more flexibility.

We compared our method with pseudobulk-based analysis. For global tests, we aggregated all the cells of an individual into a single pseudobulk sample. Pseudobulk expression was computed as the average count across cells, and log-transformed. GReX prediction models were trained using lasso and elastic net (enet) implemented by glmnet.⁶² Association testing was conducted using the same method as described in FUSION.² We also compared with the summary data-based Mendelian randomization (SMR)⁴⁴ method for global test. For dynamic tests, two pseudobulk samples were constructed by aggregating cells with pseudotime <0.5 (sample 1) and those with pseudotime ≥0.5 (sample 2), respectively. GReX prediction models were trained for each pseudobulk sample separately. Since TWAS methods do not provide dynamic tests, we implemented a customized extension of the FUSION approach. Define W_pb=(w₁,w₂), where w₁,w₂ are column vectors of weights in the prediction models for sample 1 and sample 2, respectively. The effect of the gene in sample 1 and sample 2 can be estimated as

$$\hat{\mathit{\beta }}={\left(W^{T}RW\right)}^{-1}{W}^T\hat{\mathit{b}}\text{.}$$

Test of hypothesis β₁ = β₂ was used as dynamic test for the pseudobulk setting.

For simulation studies where there was no temporal variation in genetic regulatory effects of SNPs on gene expression, we set ${\tilde{w}}_{km}\sim N\left({0,10}^{-6}\right)$. We only considered the scenarios of 10 and 20 causal SNPs and 3 and 5 true knots to save computational cost.

Quantification and statistical analysis

OneK1K data

The OneK1K cohort²⁶ consists of 1,104 donors of Northern European ancestry (58% women, 42% men). The cell by gene count from single-cell RNA-seq are available Human Cell Atlas (HCA) (https://cellxgene.cziscience.com/collections/dde06e0f-ab3b-46be-96a2-a8082383c4a1). Processed genotype data were provided by the authors of the OneK1K paper. The donors reported no active infection at the time of sample collection. The study identified cell-type-specific eQTLs for 14 immune cell types, with the largest number of discoveries in CD4⁺ naive and central memory T (CD4_NC) cells, natural killer (NK) cells, CD8⁺ T cells with an effector memory phenotype (CD8_ET), and CD8 naive and central memory T cells (CD8_NC). The study also investigated dynamic eQTL effects across the B cell landscape, and causal effects of gene expression in immune cell types on autoimmune diseases using Mendelian randomization and colocalization.²⁶

Preprocessing of genotype data

Genotype data were available for 1,104 individuals and 759,993 variants. We excluded donors with >3% missing genotypes, and those with excessive heterozygosity (heterozygosity Z score > 3). We used GCTA⁶³ to construct a genetic relationship matrix, and further removed individuals with estimated relatedness >0.125. A total of 1,033 individuals were retained. Genotype data were then imputed using the Michigan Imputation Server with Haplotype Reference Consortium (HRC) r1.1 as the reference panel (European ancestry). The reference genome was hg19, and variants with imputation quality r² < 0.3 were removed. We kept only HapMap3 European ancestry SNPs with minor allele frequency (MAF) > 1%, leading to a total of 1,170,214 SNPs.

Preprocessing of gene expression data

The Seurat object of scRNA-seq comprised 36,571 genes and 1,248,980 cells from 981 individuals. PCA and UMAP embeddings were provided with the Seurat object. Within each individual, we aggregated cells with similar expression profiles into metacells using micropooling in the VISION package.⁶⁴ On average, each metacell was pooled from 5 neighboring cells in a latent space determined by top 50 expression PCs (provided with the Seurat object from OneK1K). Gene expression counts for each metacell were computed as sum of the counts of comprising single cells. The goal of this step was to reduce the sparsity of scRNA-seq data and ease computational burden. To compute principal components (PCs) and pseudotime, we further normalized the metacell expression by the metacell library size, and multiplied by a factor of 10,000. We used Seurat to identify highly variable genes, scaled gene expression such that all genes have equal variance, and computed PCs of metacell expression of top 2,000 highly variables genes using Seurat.⁶⁵

We focused on analyzing three cell types that have continuous transition of cell states: (1) CD4⁺ T cells, (2) CD8⁺ T cells, and (3) B cells. The original cell type classification provided by OneK1K was more refined, hence we defined each of the three cell types as the combination of two or three subtypes: (1) CD4⁺ T cells included naive thymus-derived CD4⁺, alpha-beta T cells (referred to as naive CD4⁺ T cells in this paper) and central memory CD4-positive, alpha-beta T cells (referred to as central memory CD4⁺ T cells); (2) CD8⁺ T cells included naive thymus-derived CD8⁺, alpha-beta T cells (referred to as naive CD8⁺ T cells) and central memory CD8⁺, alpha-beta T cells (referred to as central memory CD8⁺ T cells); (3) B cells include naive, transitional stage, and memory B cells; We defined a metacell as a B cell if it is comprised of >70% B cells. Similar definitions were used for CD4⁺ T cells CD8⁺ T cells. Metacells with ≤70% CD4⁺ T cells, CD8⁺ T cells, and B cells were discarded. We used Slingshot⁴¹ to infer pseudotime using top 5 PCs separately for each cell type. Although many pseudotime methods are available, a previous benchmarking study⁶⁶ demonstrated that Slingshot was one of the few methods with good performance across several evaluation criteria. When needed, we reversed the order of pseudotime such that larger pseudotime corresponds to memory cells. For each cell type, we removed genes that had zero count in >80% metacells of this cell type. Prediction models were only trained for the remaining genes. This led to 4,372 genes and 97,895 metacells for CD4⁺ T cells, 4,274 genes and 10,550 metacells for CD8⁺ T cells, 4,132 genes and 24,096 metacells for B cells.

Training expression prediction models

Stage-1 prediction models were trained using TWiST to predict single-cell gene expression from cis-SNPs (±500kb from transcription start site). We used cubic B-splines with 3 internal knots (0.25, 0.5, 0.75). Model fitting was conducted using grpreg,⁶⁷ an R package for fitting group lasso. We assigned equal weights to group L1 and L2 penalties, creating a group elastic net penalty. Tuning parameters were selected using cross-validation. We included age, sex, metacell library size, top 10 metacell gene expression PCs, and top 10 genotype PCs as covariates. We removed genes with non-zero weights in <3 SNPs and retained 3,311 genes for CD4⁺ T cells, 1,399 genes for CD8⁺ T cells, and 2,471 genes for B cells for downstream analysis.

To evaluate the prediction performance of TWiST, we randomly selected subsets of the OneK1K data for CD4⁺ T cells of a varying number of donors as training data: n = 100, 300, 500, 900. R² values were computed on the test data that were part of or all the remaining samples. For n = 100, 300, and 500, a randomly chosen 400 of the remaining individuals were used as test data. For n = 900, all the remaining samples (n = 81) were used as test data. Using the model trained on the training data, we computed the genetically regulated expression (${\hat{v}}_i\left(t_{ij}\right)$) in the test data. We used pseudotime value t_ij previously inferred from the entire dataset. We further computed the log-transformed, normalized expression as ${\tilde{x}}_{ij}=\log \left(1+10000x_{ij}/{\alpha }_{ij}\right)$ and regressed out covariates: age, sex, top 10 metacell gene expression PCs, and top 10 genotype PCs. The residual was used to compute a correlation with ${\hat{v}}_i\left(t_{ij}\right)$, and the squared correlation was defined as the R² at cell level. Here we removed covariates from gene expression in analogy with previous bulk TWAS models which were trained on residual gene expression with covariates removed. To compare prediction performance with bulk TWAS methods, we averaged the residuals across cells within each individual and obtained individual-level gene expression. Similarly, we averaged ${\hat{v}}_i\left(t_{ij}\right)$ across cells within each individual and obtained individual-level predicted expression. Finally, we computed the correlation between individual level predicted and observed expression and the squared correlation as R².

GWAS summary statistics and TWiST analysis

We collected GWAS summary statistics for seven autoimmune diseases: rheumatoid arthritis (RA, 22,350 cases and 74,823 controls),⁴⁵ systemic lupus erythematosus (SLE, 4,943 cases and 8,483 controls),⁴⁶ Crohn’s disease (CD, 5,956 cases and 21,770 controls),⁴⁷ inflammatory bowel disease (IBD, 6,968 cases and 21,770 controls),⁴⁷ multiple sclerosis (MS, 47,429 cases and 68,374 controls),⁴⁸ type 1 diabetes (T1DM, 3,545 cases and 409,155 controls)⁴⁹ and ankylosing spondylitis (AS, 495 cases and 371,238 controls).⁴⁹ GWAS Summary statistics were downloaded via the links provide in the Supplementary Files of Yazar et al.²⁶ We merged GWAS summary statistics with TWiST model weights, removed SNPs with zero weights, and computed the LD matrix of the remaining SNPs. We removed genes for which the rank of LD matrix of model SNPs was less than 5 due to insufficient rank to model complex pseudotemporal patterns. TWiST was applied to estimate and test the effect of gene expression in CD4⁺ T cells, CD8⁺ T cells, and B cells on the seven autoimmune diseases listed above. This analysis was performed using the stage-2 model of TWiST. Gene effect β(t) was modeled using cubic B-splines with 0 and 1 as boundary knots and 19 equidistant internal knots: 0, 0.05, 0.1, …, 0.95, 1.

Pseudobulk TWAS and scPrediXcan

Pseudobulk aggregation was performed on the original counts (instead of metacells). We identified cells belonging to one of the following subtypes based on cell type labels given by the OneK1K data: naive CD4⁺ T cells, central memory CD4⁺ T cells, naive CD8⁺ T cells, central memory CD8⁺ T cells, naive B cells, and memory B cells. For each individual, we summed the UMI counts across cells within each of the subtypes and created 6 pseudobulk samples. For each subtype, we aggregated pseudobulk UMI counts across all individuals and created a gene x (number of individuals) matrix. Pseudobulk gene expression matrix was then normalized using TMM normalization.⁶⁸ We further removed genes with zero expression in more than 50% of the cells. The remaining data were log-transformed and scaled such that all genes have unit variance. Expression PCs were computed from 500 highly variable genes. Covariates, including age, sex, 10 expression PCs, and 10 genotype PCs were regressed out of the log-transformed scaled expression. The residuals were inverse-normally transformed such that they were normally distributed, and then used to train models to predict gene expression using cis-SNPs (±500kb from TSS). FUSION² was used for model training and downstream association tests with autoimmune diseases. scPrediXcan³² is a recently developed single-cell TWAS method incorporating deep learning and epigenomic data. scPrediXcan also tests cell-type-specific associations at pseudobulk level. We downloaded pre-trained weights for the following cell types in OneK1K from PredictDB (https://predictdb.org/): naive_thymus-derived_CD4-positive_alpha-beta_T_cell, central_memory_CD4-positive_alpha-beta_T_cell, naive_thymus-derived_CD8-positive_alpha-beta_T_cell, central_memory_CD8-positive_alpha-beta_T_cell, naive_B_cell, and memory_B_cell. Association analysis was conducted using the pre-trained weights and GWAS summary statistics for seven autoimmune diseases.

Pseudobulk colocalization analysis

To evaluate whether the dynamic genes reflect biologically meaningful transition in cell states, we conducted colocalization analysis between eQTL and GWAS effects in pseudobulk samples. First, we defined pseudobulk samples for each cell type (CD4⁺ T cell, CD8⁺ T cell, B cell) based on pseudotime bins: early stage (pseudotime ≤0.5, closer to naive cells) and late stage (pseudotime>0.5, closer to memory cells). Pseudobulk gene expression was processed using the same pipeline as in the pseudobulk TWAS analysis, as described in the previous subsection. We conducted eQTL association testing for the early and late stage separately, using linear regression with expression residual as outcome and cis-SNPs as exposure. Colocalization analysis was conducted using eQTL and GWAS summary statistics by R package coloc.⁵⁶ We also created pseudobulk samples based on naive versus memory for each of the three cell types using subtype annotations provided by the OneK1K study. Colocalization was determined by threshold PPH4>0.8.

Novel genes identified by TWiST

To evaluate whether TWiST identified novel disease susceptibility genes, we compared the lists of genes identified by TWiST global and dynamic tests with previously reported genes in the TWAS Atlas⁵¹ (accessed on May 19, 2025) and genes identified by pseudobulk FUSION in our dataset (see previous subsection, Pseudobulk TWAS and scPrediXcan). For each combination of disease, cell type, and statistical test (e.g., RA, T_CD4, dynamic test), we defined a gene as novel if it had not been reported to be associated with the disease in the TWAS Atlas, or identified as significant (FDR<0.05) by pseudobulk FUSION.

Clustering pseudotemporal patterns for multiple sclerosis

First, we selected the genes that were dynamically associated with MS. Curves representing the effect of gene expression on diseases at varying cell states were estimated using the BLUP approach in TWiST. We then computed the value of the curve on a fine grid between 0 and 1: 0, 0.01, 0.02, …, 0.99, 1. We scaled the values for each gene such at the average effect across pseudotime is 0 and the variance is 1. The scaling accounted for the difference in magnitude across genes and focused on the shape of the curve. We then used K-means clustering to cluster the genes into 6 groups based on pseudotemporal patterns. Mean silhouette scores⁶⁹ computed across genes and 20 random initiations were used to measure clustering quality.

Gene set enrichment analysis

For both the set of dynamic disease genes and genes in clusters, we conducted gene set enrichment analysis (GSEA) using Gene Ontology.^52,53 We restricted the analysis to biological processes. Pathways overrepresented by the gene sets with FDR<0.05 were considered significant. Significant pathways were classified as immune and non-immune based on whether it was a descendant of GO:0002376 (immune system processes). To identify novel pathways, we conducted GSEA for the list of previously known genes (TWAS Atlas and pseudobulk FUSION, see Novel genes identified by TWiST) for each autoimmune disease. We reported the pathways that were only enriched for TWiST dynamic genes but not for previously known genes as novel.

As a sensitivity analysis, we used the top 1,000 highly expressed genes in immune cells as an alternative list of background genes. This list was generated by summing UMI counts across all cells in the OneK1K scRNA-seq data for each gene, and retrieving the 1000 genes with largest total count. Gene set enrichment analysis was conducted using FUMA GWAS⁵⁵ (https://fuma.ctglab.nl/). Significant enrichment was determined at FDR<0.05.

Published: November 3, 2025

Supplemental information

Table S7. Gene Ontology pathways enriched in combined gene set of MS cluster 3 and 5 dynamic genes in CD4⁺ T cells, related to Figure 6

Rows in bold face are pathways enriched in cluster 5 genes only.