Cells from the same individual share common genetic and environmental backgrounds and are not statistically independent; therefore, they are subsamples or pseudoreplicates. Thus, single-cell data have a hierarchical structure that many current single-cell methods do not address, leading to biased inference, highly inflated type 1 error rates, and reduced robustness and reproducibility. This includes methods that use a batch effect correction for individual as a means of accounting for within-sample correlation. Here, we document this dependence across a range of cell types and show that pseudo-bulk aggregation methods are conservative and underpowered relative to mixed models. To compute differential expression within a specific cell type across treatment groups, we propose applying generalized linear mixed models with a random effect for individual, to properly account for both zero inflation and the correlation structure among measures from cells within an individual. Finally, we provide power estimates across a range of experimental conditions to assist researchers in designing appropriately powered studies.

Background: Study design is a critical aspect of any experiment, and sample size calculations for statistical power that are consistent with that study design are central to robust and reproducible results. However, the existing power calculators for tests of differential expression in single-cell RNA-seq data focus on the total number of cells and not the number of independent experimental units, the true unit of interest for power. Thus, current methods grossly overestimate the power. Results: Hierarchicell is the first single-cell power calculator to explicitly simulate and account for the hierarchical correlation structure (i.e., within sample correlation) that exists in single-cell RNA-seq data. Hierarchicell, an R-package available on GitHub, estimates the within sample correlation structure from real data to simulate hierarchical single-cell RNA-seq data and estimate power for tests of differential expression. This multi-stage approach models gene dropout rates, intra-individual dispersion, inter-individual variation, variable or fixed number of cells per individual, and the correlation among cells within an individual. Without modeling the within sample correlation structure and without properly accounting for the correlation in downstream analysis, we demonstrate that estimates of power are falsely inflated. Hierarchicell can be used to estimate power for binary and continuous phenotypes based on user-specified number of independent experimental units (e.g., individuals) and cells within the experimental unit. Conclusions: Hierarchicell is a user-friendly R-package that provides accurate estimates of power for testing hypotheses of differential expression in single-cell RNA-seq data. This R-package represents an important addition to single-cell RNA analytic tools and will help researchers design experiments with appropriate and accurate power, increasing discovery and improving robustness and reproducibility.

Gas chromatography–coupled mass spectrometry (GC–MS) has been used in biomedical research to analyze volatile, non-polar, and polar metabolites in a wide array of sample types. Despite advances in technology, missing values are still common in metabolomics datasets and must be properly handled. We evaluated the performance of ten commonly used missing value imputation methods with metabolites analyzed on an HR GC–MS instrument. By introducing missing values into the complete (i.e., data without any missing values) National Institute of Standards and Technology (NIST) plasma dataset, we demonstrate that random forest (RF), glmnet ridge regression (GRR), and Bayesian principal component analysis (BPCA) shared the lowest root mean squared error (RMSE) in technical replicate data. Further examination of these three methods in data from baboon plasma and liver samples demonstrated they all maintained high accuracy. Overall, our analysis suggests that any of the three imputation methods can be applied effectively to untargeted metabolomics datasets with high accuracy. However, it is important to note that imputation will alter the correlation structure of the dataset and bias downstream regression coefficients and p-values.

Peer review information Nature Communications thanks the anonymous reviewer(s) for their contribution to the peer review of this work.

Systemic lupus erythematosus (SLE) is a chronic, multisystem, autoimmune inflammatory disease with genomic and non-genomic contributions to risk. We hypothesize that epigenetic factors are a significant contributor to SLE risk and may be informative for identifying pathogenic mechanisms and therapeutic targets. To test this hypothesis while controlling for genetic background, we performed an epigenome-wide analysis of DNA methylation in genomic DNA from whole blood in three pairs of female monozygotic (MZ) twins of European ancestry, discordant for SLE. Results were replicated on the same array in four cell types from a set of four Danish female MZ twin pairs discordant for SLE. Genes implicated by the epigenetic analyses were then evaluated in 10 independent SLE gene expression datasets from the Gene Expression Omnibus (GEO). There were 59 differentially methylated loci between unaffected and affected MZ twins in whole blood, including 11 novel loci. All but two of these loci were hypomethylated in the SLE twins relative to the unaffected twins. The genes harboring these hypomethylated loci exhibited increased expression in multiple independent datasets of SLE patients. This pattern was largely consistent regardless of disease activity, cell type, or renal tissue type. The genes proximal to CpGs exhibiting differential methylation (DM) in the SLE-discordant MZ twins and exhibiting differential expression (DE) in independent SLE GEO cohorts (DM-DE genes) clustered into two pathways: the nucleic acid-sensing pathway and the type I interferon pathway. The DM-DE genes were also informatically queried for potential gene–drug interactions, yielding a list of 41 drugs including a known SLE therapy. The DM-DE genes delineate two important biologic pathways that are not only reflective of the heterogeneity of SLE but may also correlate with distinct IFN responses that depend on the source, type, and location of nucleic acid molecules and the activated receptors in individual patients. Cell- and tissue-specific analyses will be critical to the understanding of genetic factors dysregulating the nucleic acid-sensing and IFN pathways and whether these factors could be appropriate targets for therapeutic intervention.

Gas chromatography-coupled mass spectrometry (GC-MS) has been used in biomedical research to analyze volatile, non-polar, and polar metabolites in a wide array of sample types. Despite advances in technology, missing values are still common in metabolomics datasets and must be properly handled. We evaluated the performance of ten commonly used missing value imputation methods with metabolites analyzed on an HR GC-MS instrument. By introducing missing values into the complete (i.e., data without any missing values) NIST plasma dataset we demonstrate that Random Forest (RF), Glmnet Ridge Regression (GRR), and Bayesian Principal Component Analysis (BPCA) shared the lowest Root Mean Squared Error (RMSE) in technical replicate data. Further examination of these three methods in data from baboon plasma and liver samples demonstrated they all maintained high accuracy. Overall, our analysis suggests that any of the three imputation methods can be applied effectively to untargeted metabolomics datasets with high accuracy. However, it is important to note that imputation will alter the correlation structure of the dataset, and bias downstream regression coefficients and p-values.