diff_mean_test: Non-parametric differential expression test for sparse...

View source: R/differential_expression.R

diff_mean_testR Documentation

Non-parametric differential expression test for sparse non-negative data


Non-parametric differential expression test for sparse non-negative data


  compare = "each_vs_rest",
  R = 99,
  log2FC_th = log2(1.2),
  mean_th = 0.05,
  cells_th = 5,
  only_pos = FALSE,
  only_top_n = NULL,
  mean_type = "geometric",
  verbosity = 1



A matrix of counts; must be (or inherit from) class dgCMatrix; genes are row, cells are columns


The group labels (e.g. cluster identities); will be converted to factor


Specifies which groups to compare, see details; default is 'each_vs_rest'


The number of random permutations used to derive the p-values; default is 99


Threshold to remove genes from testing; absolute log2FC must be at least this large for a gene to be tested; default is log2(1.2)


Threshold to remove genes from testing; gene mean must be at least this large for a gene to be tested; default is 0.05


Threshold to remove genes from testing; gene must be detected (non-zero count) in at least this many cells in the group with higher mean; default is 5


Test only genes with positive fold change (mean in group 1 > mean in group2); default is FALSE


Test only the this number of genes from both ends of the log2FC spectrum after all of the above filters have been applied; useful to get only the top markers; only used if set to a numeric value; default is NULL


Which type of mean to use; if 'geometric' (default) the geometric mean is used; to avoid log(0) we use log1p to add 1 to all counts and log-transform, calculate the arithmetic mean, and then back-transform and subtract 1 using exp1m; if this parameter is set to 'arithmetic' the data is used as is


Integer controlling how many messages the function prints; 0 is silent, 1 (default) is not


Data frame of results


This model-free test is applied to each gene (row) individually but is optimized to make use of the efficient sparse data representation of the input. A permutation null distribution us used to assess the significance of the observed difference in mean between two groups.

The observed difference in mean is compared against a distribution obtained by random shuffling of the group labels. For each gene every random permutation yields a difference in mean and from the population of these background differences we estimate a mean and standard deviation for the null distribution. This mean and standard deviation are used to turn the observed difference in mean into a z-score and then into a p-value. Finally, all p-values (for the tested genes) are adjusted using the Benjamini & Hochberg method (fdr). The log2FC values in the output are log2(mean1 / mean2). Empirical p-values are also calculated: emp_pval = (b + 1) / (R + 1) where b is the number of times the absolute difference in mean from a random permutation is at least as large as the absolute value of the observed difference in mean, R is the number of random permutations. This is an upper bound of the real empirical p-value that would be obtained by enumerating all possible group label permutations.

There are multiple ways the group comparisons can be specified based on the compare parameter. The default, 'each_vs_rest', does multiple comparisons, one per group vs all remaining cells. 'all_vs_all', also does multiple comparisons, covering all groups pairs. If compare is set to a length two character vector, e.g. c('T-cells', 'B-cells'), one comparison between those two groups is done. To put multiple groups on either side of a single comparison, use a list of length two. E.g. compare = list(c('cluster1', 'cluster5'), c('cluster3')).


clustering <- 1:ncol(pbmc) %% 2
vst_out <- vst(pbmc, return_corrected_umi = TRUE)
de_res <- diff_mean_test(y = vst_out$umi_corrected, group_labels = clustering)

sctransform documentation built on Oct. 19, 2023, 9:08 a.m.