Adjusted p-value. A p-value corrected for the number of tests, here usually by the method of Benjamini and Hochberg, which controls the false discovery rate. Multiple testing
B-statistic. The log-odds that a gene’s coefficient or contrast is not zero. Empirical Bayes moderation
Competitive and self-contained tests. A competitive gene-set test asks whether a set’s genes change more than the other genes; a self-contained test asks whether they change at all. Gene-set tests
Contrast. A combination of a design’s coefficients that answers one question, such as stimulated minus control. It is estimable when the samples determine it. Designs, linear models and contrasts
Counts per million (CPM). Counts divided by the effective library size, times one million; log-CPM is their log2, after a small count is added. Count models
Cube. CytePool’s container of pseudobulk profiles: for every sample, group and feature, sums over cells, with tables that describe the samples, groups and features. The cube and its additive layers
Design. The sample table with the model to fit: an R formula for the fixed effects, random effects in lme4’s notation, and named contrasts. Designs, linear models and contrasts
Digest. The SHA-256 hash of one part of a cube, such as a table or one group of a layer, from which the fingerprint is built. Cube files and provenance records
Dispersion. The parameter of a negative binomial distribution that sets how much the variance exceeds the mean; its square root is the biological coefficient of variation. Count models
Empirical Bayes moderation. Shrinking each gene’s variance towards a value estimated from all genes, which steadies tests with few samples. The prior degrees of freedom say how much the shared value counts. Empirical Bayes moderation
False discovery rate. The expected share of false calls among all calls. Multiple testing
Feature. A measured variable, usually a gene: a column of a samples x features matrix.
Fingerprint. A digest of a cube’s values, its tables, cell numbers and layers included, that names the cube however it was made. Cube files and provenance records
Gene set. A named list of genes, such as a pathway or a signature, from a collection such as MSigDB, often read from a GMT file. Gene-set tests
Group. The second axis of a cube, usually a cell type: the cells of one sample are pooled per group. A bulk cube has one group.
Heterogeneity. How much the studies of a meta-analysis disagree beyond their sampling error, measured by the between-study variance tau², Cochran’s Q and I². Meta-analysis
Kenward-Roger’s method. An adjustment of the standard errors of a mixed model’s estimates, and of the degrees of freedom of their tests, for the uncertainty of its variance components in small samples. Mixed models, REML and dream
Layer. One statistic of a cube, summed over cells: counts, detected cells, or sums of normalized and log-normalized values. The cube and its additive layers
Library size. The total count of a sample, or of a profile; times a normalization factor, the effective library size.
Normalization factor and size factor. Corrections for sequencing depth and composition: edgeR multiplies library sizes by normalization factors (such as TMM), and DESeq2 replaces them by size factors. Count models
Over-representation analysis. A test of whether the genes called significant hold more of a set’s genes than chance would give, by the hypergeometric distribution against a universe of genes. Gene-set tests
Precision weight. A weight for each observation, the inverse of its predicted variance, such as voom’s. Precision weights
Provenance record. The record of where a cube’s data came from (its source), how it was built and every operation since, each naming its input by fingerprint. Cube files and provenance records
Pseudobulk profile. The sums over the cells of one sample and group, which treats the sample, not the cell, as the unit of replication. Choosing a method
Quasi-dispersion. In edgeR’s quasi-likelihood model, a gene’s own multiplier of the negative binomial variance. Count models
Random intercept. A shift shared by the samples of one level of a grouping factor, such as a donor, drawn from a normal distribution whose variance is estimated. Mixed models, REML and dream
Reference level. The first level of a factor, against which the coefficients of its other levels are differences. Designs, linear models and contrasts
REML. Restricted maximum likelihood, which estimates variance components without the bias that estimating the fixed effects gives maximum likelihood. Mixed models, REML and dream
Residual degrees of freedom. The number of samples a fit uses minus the number of coefficients it estimates: what is left to estimate the variance.
Rotation test. A test that compares a statistic with its values under random rotations of a linear model’s residual space, which leave the data’s distribution unchanged under the null hypothesis. Gene-set tests
Sample. The unit of replication: a donor, or a donor under one condition. Its cells are pooled into one profile per group.
Sample quality weight. A weight for each sample, from how closely its values follow the model across genes. Precision weights
Satterthwaite’s degrees of freedom. An approximation to the degrees of freedom of a t-test in a mixed model. Mixed models, REML and dream
Source record. The caller’s record of where a cube’s data came from, kept in its provenance as given. Cube files and provenance records
TREAT. A test that a fold change exceeds a threshold, rather than that it is not zero. Empirical Bayes moderation