Table 1.

Details of the methods for the differential state analysis of scRNA-seq data compared in this study

	Pseudobulk methods				Single-cell methods
	Pseudobulk methods that require built-in normalization		Pseudobulk methods that can be used with any normalization		Mixed models accounting for subjects as a random effect			Naïve methods that do not model subjects	Methods that have the option to use latent variables to correct for batches, etc.
Method name	DESeq2	edgeR	Limma	ROTS	MAST_RE	muscat_MM	NEBULA-LN	wilcoxon	MAST	LR	negbinom	poisson
Normalization	Median of ratios	TMM	TMM + voom	TMM + CPM + log2	No default (Log normalize)	Log normalize	Normalization factors from library sizes	Log normalize	Log normalize	Log normalize	Log normalize	Log normalize
Statistical tests	Negative binomial generalized linear model	Negative binomial model + empirical Bayes procedure	Linear model + empirical Bayes procedure	Reproducibility optimized test statistic	Two-part hurdle model with random effect for subject	lme4 linear mixed model with voom weights	Negative binomial mixed model	Wilcoxon rank sum test	Two-part hurdle model	Logistic regression	Negative binomial generalized linear model	Poisson generalized linear model
R packages (normalization, test)	DESeq2	edgeR	edgeR, Limma	edgeR, ROTS	MAST	muscat	nebula	Seurat	Seurat, MAST	Seurat	Seurat	Seurat
Filtering	Nonexpressed genes	Nonexpressed genes	Nonexpressed genes	Nonexpressed genes	Number cells expressing gene < subjects	Number cells expressing gene <20, Number cells in sample < 10	Genes with counts per cell <0.005	Nonexpressed genes	Nonexpressed genes	Nonexpressed genes	Number cells expressing genes <3	Number cells expressing genes <3
Version	1.32.0	3.34.0	3.48.0	1.20.0	1.18.0	01-06-2000	01-01-2007	4.0.2	4.0.2	4.0.2	4.0.2	4.0.2
References	[18]	[17]	[17, 19]	[17, 20]	[15]	[6, 21, 22]	[16]	[23]	[15, 23]	[23]	[23]	[23]