Choosing a method

CytePool’s tests compare samples, not cells: each takes one profile per sample, so that a difference is measured against the variation between donors. Cells of one donor share its genetics and environment, and a test that treats them as independent replicates finds differences that do not hold across donors Squair et al., 2021Zimmerman et al., 2021Murphy & Skene, 2022. Pool the cells of each sample and cell type into a cube first (from_cells); the methods below then run on its pseudobulk profiles.

Genes that change in each cell type¶

Each runs on every group (cell type) of a cube, one at a time, and returns a DifferentialExpressionResult with a table for every cell type and contrast (DifferentialExpressionResult.table), a summary, and which samples each fit used and why the others were left out. The limma User’s Guide (sections 15.4 and 15.5) says that limma-trend usually works well when the largest library size is not more than about three times the smallest, and that voom is more powerful when library sizes are quite variable. Pseudobulk library sizes grow with the number of cells pooled, which can differ severalfold between samples, and limma_voom is the default.

A fixed effect per donor (~ donor + condition) compares the samples of each donor with each other, which removes the differences between donors from the noise. It cannot test a variable that is constant within each donor, such as sex: that variable is confounded with the donor effects, and Design.check says so. A random intercept per donor (dream_voom) keeps the correlation between the samples of a donor and still estimates donor-level effects. Mixed models, REML and dream explains the model.

One matrix of counts¶

Bulk RNA-seq, or one cell type taken out of a cube, is a samples x genes matrix of counts. Three families of methods test it:

• limma: voom, fit_linear_model, contrasts_fit, then moderated_statistics or treat for a fold-change threshold. A bulk experiment can also become a cube with one group (from_bulk), on which limma_voom runs the same steps.

• edgeR’s quasi-likelihood F-tests: fit_quasi_likelihood and quasi_likelihood_f_test, negative binomial models with a moderated quasi-dispersion for every gene Chen et al., 2025.

• DESeq2: fit_deseq2, with deseq2_wald_test for one coefficient or contrast and deseq2_likelihood_ratio_test for several coefficients at once, such as every level of a factor Love et al., 2014.

All three moderate the variability of each gene by borrowing from the others, and they agree on genes with clear effects. They differ in the scale they model (log counts per million, or counts) and in how they shrink: limma moderates residual variances towards a prior, edgeR moderates quasi-dispersions, DESeq2 shrinks dispersions towards a trend in the mean count. Count models gives the equations. In this version, edgeR’s and DESeq2’s methods run on one matrix at a time, not on every group of a cube.

Other questions¶

Choices that matter in every method¶

• Samples with few cells. A profile pooled from a handful of cells is noisy. min_cells (10 by default) leaves such samples out of a cell type’s fit, and min_counts sets the least total count a sample needs; the scverse best practices keep samples with at least 10 cells and 1,000 counts Heumos et al., 2023. prior_weights="n_cells" instead keeps every sample and weights it by its number of cells Hoffman et al., 2026.

• Genes with low counts. Genes are filtered by expression with the design, by edgeR’s rule (filter_by_expression; Chen et al. (2016)), before voom’s trend is fitted.

• Outlying samples. quality_weights=True estimates a weight for every sample from how well it follows the model, instead of removing it Liu et al., 2015, and robust=True protects the prior from genes with outlying variances Phipson et al., 2016.

• Many cell types. p-values are adjusted within each cell type and contrast; adjust_across_groups=True adds an adjustment over every cell type (Multiple testing).