Optimized summary-statistic-based single-cell eQTL meta-analysis

Abstract

The identification of expression quantitative trait loci (eQTLs) holds great potential to improve the interpretation of disease-associated genetic variation. As many such disease-associated variants act in a context-, tissue- or even cell-type-specific manner, single-cell RNA-sequencing (scRNA-seq) data is uniquely suitable for identifying the specific cell type or context in which these genetic variants act. However, due to the limited sample sizes in single-cell studies, discovery of cell-type-specific eQTLs is now limited. To improve power to detect such eQTLs, large-scale joint analyses are needed. These are however, complicated by privacy constraints due to sharing of genotype data and the measurement and technical variety across different scRNA-seq datasets as a result of differences in mRNA capture efficiency, experimental protocols, and sequencing strategies. A solution to these issues is a federated weighted meta-analysis (WMA) approach in which summary statistics are integrated using dataset-specific weights. Here, we compare different strategies and provide best practice recommendations for eQTL WMA across scRNA-seq datasets.

Keywords: Weighted meta-analysis; eQTL; scRNA-seq.

Fig. 1

Workflow and weighting strategies for meta-analysis of single-cell eQTL data. (a) The general workflow for weighted meta-analysis. (b) The cohort-, gene-, SNP-, and eQTL-specific weights used in the meta-analysis. (c, d) Description of the (c) peripheral blood mononuclear cell (PBMC) and (d) induced pluripotent stem cell (iPSC) datasets used in the analysis. The PBMC data consists of five single-cell datasets (n = 187 donors: 141 with 10X V2 and 46 with 10X V3 chemistry). The iPSC data consists of two single-cell datasets (n = 87 donors: 87 Smart-Seq2 and 25 donor-matched 10X V2). To analyze independent sample sets and test for the effects of combining smaller datasets, we split the 87 Smart-Seq2 samples into Set A (62 non-10X-matched samples) and Set C (25 10X-matched samples). Set A was further split into a smaller set of 25 non-10X-matched samples (Set D).

Fig. 2

Results of the improved sc-eQTL weighted meta-analysis in the PBMC pseudobulk datasets: (a–d) Comparison of the fraction of eGenes detected and accuracy (F1* score) in WMA over all five datasets (a), 10X V3 chemistry (b), 10X V2 and V3 chemistries (c), and 10X V2 chemistry combinations (d). e Three pairwise 10X V2 chemistry dataset combinations. The expansion at right enlarges the top-performing weights in these comparisons. (f, g) Comparison of the best-performing weights among 10 pairwise dataset combinations in terms of the (f) increase in the number of eGenes and (g) the weighted mean of the eGenes change in comparison to sample-size weighting.

Fig. 3

Meta-analysis of sample-size-like weights in monocyte datasets. (a–d) Combination of the number of eGenes detected and accuracy (F1* score) for (a) all five datasets and (b) sample-size-wise the largest pairwise comparisons of 10X V3 chemistry, (c) 10X V2 and V3 chemistries, and (d) 10X V2 chemistry combinations. (e) Three pairwise 10X V2 chemistry dataset combinations. The expansion at right enlarges the top-performing weights. (f, g) Comparison of the (f) main weights among 10 pairwise dataset combinations in terms of the increase in the number of eGenes and (g) the weighted mean of eGenes change.

Fig. 4

Weighted meta-analysis of sample-size-like characteristics in iPSC samples. a–d The number of eGenes detected and accuracy (F1* score) for all four iPSC datasets. e, f Results of the grid search in combinations of Smart-seq2 (Set B or C) and 10X (Set D) samples. Vertical lines indicate the ratio of weight of Set C versus weight of Set D (e) or of Set C versus Set D (f) for one of the weights of the WMA.

Fig. 5

Weighted meta-analysis with secondary weights on a monocyte and b iPSC datasets employing secondary weights derived from gene expression features and minor allele frequency (MAF) in combination with overall best-performing sample-size-derived weights. a For monocyte samples, we used the average number of cells per donor as a main weight. b For iPSC samples, we used sample size as a main weight. Numbers in the heatmap indicate the change in the number of eGenes detected. Green and red indicate an increase or decrease in the number of eGenes compared to the main weight, respectively. Color saturation indicates the fold-change of the number of eGenes detected.