Literature

A reading list grouped by the questions that studies of many samples ask (Overview). Each line says what the paper shows, in its own terms, and numbers are those the papers report. The methods behind CytePool’s functions are cited on the pages of How the statistics work; every work is in the Bibliography.

Cells or samples¶

• Squair et al. (2021) compared differential expression methods on 18 datasets with matched bulk RNA-seq. Methods that ignore the variation between biological replicates were biased and prone to false discoveries, finding hundreds of differentially expressed genes where there were no biological differences.

• Zimmerman et al. (2021) showed that cells of one individual are not independent, and that ignoring it inflates the type 1 error. They proposed mixed models with a random effect for the individual, and found pseudobulk methods conservative in comparison.

• Murphy & Skene (2022) reran that benchmark with a measure that balances false positives and false negatives, and found that pseudobulk methods performed best.

• Junttila et al. (2022) compared 18 methods for differential states in multi-subject data: pseudobulk methods performed generally best, and methods that treat cells as independent gave many false positives.

• Crowell et al. (2020) surveyed cell-level mixed models and pseudobulk methods for differential state analysis across samples and conditions, with a simulation of multi-sample single-cell data (muscat).

• Lun & Marioni (2017) showed that ignoring plate effects loses control of the type I error, and that summing the counts of the cells of each plate restores it, without losing power in their simulations.

• Hafner et al. (2025) found that methods built for single cells offered no advantage over pseudobulk DESeq2 on individual datasets in nested designs, and suggested dream as a compromise at atlas scale.

• Nguyen et al. (2023) benchmarked 46 workflows for data from several batches: batch-corrected data rarely improved the analysis of sparse data, while modelling the batch as a covariate helped where batch effects were large.

• Heumos et al. (2023) summarized independent benchmarks into best-practice workflows for single-cell analysis.

Which genes change, and do they replicate¶

• Kang et al. (2018) pooled cells of 8 lupus patients, split into interferon-beta stimulated and untreated cultures, and assigned cells to donors by their genotypes; the documentation uses these data (Datasets).

• Perez et al. (2022) profiled more than 1.2 million blood cells of 162 lupus cases and 99 controls, and found interferon-stimulated genes raised in monocytes and fewer naive CD4 T cells.

• Murphy et al. (2023) re-analysed the first single-nucleus study of Alzheimer’s disease: testing donors rather than cells gave 26 differentially expressed genes at a false discovery rate of 0.05, against 14,274, 549 times fewer.

• Nakatsuka et al. (2025) found that over 85% of the Alzheimer’s disease genes called in one dataset were not reproduced in any of 16 others, and that a meta-analysis of ranks across datasets found genes that predicted disease better.

• Garg et al. (2021) re-analysed COVID-19 single-cell studies in a standard way and confirmed 8 of 20 published results across datasets.

Which cell types expand or contract¶

• Phipson et al. (2022) introduced propeller, which tests transformed cell-type proportions with linear models and empirical Bayes moderation across replicates.

• Büttner et al. (2021) introduced scCODA, a Bayesian model of compositional single-cell data.

• Mangiola et al. (2023) introduced sccomp, which models composition and its variability together, and found that the variability of proportions depends on their mean in every technology examined.

• Hoffman & Roussos (2026) introduced crumblr, precision-weighted mixed models of cell-type composition that test several levels of the cell lineage at once.

• Dann et al. (2022) introduced Milo, which tests differential abundance in neighbourhoods of a k-nearest neighbour graph instead of clusters.

• Yi et al. (2024) benchmarked six differential abundance methods for condition-specific cell states, and gave recommendations by dataset.

• Stephenson et al. (2021) profiled over 780,000 blood cells of 130 COVID-19 patients and related the expansion of cell populations to disease severity.

Which pathways and regulators respond¶

• Wu & Smyth (2012) showed that gene set tests that permute genes are very sensitive to the correlation between genes, and introduced camera, which allows for it.

• Wu et al. (2010) introduced ROAST, which rotates the residuals of a linear model instead of permuting samples, so that it applies to almost any design.

• Subramanian et al. (2005) introduced gene set enrichment analysis.

• Holland et al. (2020) found that tools for pathway and transcription-factor activity built for bulk data, with curated gene sets, work on single-cell data, and that their results depend more on the gene sets than on the statistic.

• Mompel et al. (2022) collected activity methods in one framework, decoupleR, and found simple linear models among the best at predicting perturbed regulators.

What drives variation between donors¶

• Hoffman & Schadt (2016) introduced variancePartition, which uses linear mixed models to attribute the variation of each gene to disease, sex, tissue, genetics or technical factors.

• Hoffman & Roussos (2021) introduced dream, mixed models for repeated measures designs, and showed that other methods can give reproducible false positives driven by the genetic regulation of expression.

• Hoffman et al. (2026) introduced dreamlet, which runs precision-weighted mixed models on the pseudobulk profiles of every cell type, and applied it to 1.4 million nuclei from 150 Alzheimer’s disease cases and 149 controls.

• Yazar et al. (2022) profiled 1,267,758 blood cells of 982 donors and mapped 26,597 independent cis-expression quantitative trait loci across 14 cell types.

Which interactions change¶

• Dimitrov et al. (2022) compared 16 resources and 7 methods for cell-cell communication, and found that the choice of both strongly changes the predicted interactions.

• Dimitrov et al. (2024) introduced LIANA+, a framework for cell-cell communication in single-cell and spatial data, from one or several conditions.

How cells respond to drugs¶

• Srivatsan et al. (2020) introduced sci-Plex and profiled about 650,000 single cells across about 5,000 samples of three cancer cell lines exposed to 188 compounds.

Combining studies¶

• CZI Cell Science Program et al. (2025) described CZ CELLxGENE Discover, a curated corpus of over 93 million unique cells with standardized metadata.

• Sikkema et al. (2023) integrated 49 datasets into the Human Lung Cell Atlas, over 2.4 million cells from 486 individuals.

• Urbut et al. (2019) introduced mash, which estimates many effects across many conditions while allowing for correlations between them.

• Viechtbauer (2010) described metafor, the R package for meta-analysis that CytePool’s meta-analysis follows.