Test differential transcript usage

test_usage() tests whether the share of a gene's expression carried by each isoform differs between groups of samples, with one of three established frameworks:

• "diffsplice": edgeR quasi-likelihood fits and diffSplice, the pipeline of the 2025 Nucleic Acids Research paper below. Fast, and the one designed for counts divided by the read-ambiguity overdispersion.

• "satuRn": quasi-binomial models of transcript usage (Gilis et al., 2021), suited to many samples and to single-cell data.

• "DEXSeq": negative binomial models of exon (here transcript) usage (Anders et al. 2012), the slowest and often the most conservative.

Recent benchmarks of usage tests (2024 and 2025) rank these among the best; none dominates, so the backend is an argument. With inferential replicates in the object (see add_uncertainty()), the counts are divided by the overdispersion that the read ambiguity adds (see inferential_overdispersion()) before the test (scale = TRUE).

Transcripts are kept when at least min_count reads and a share of min_prop of the gene fall on them in as many samples as the smallest group, and genes need two such transcripts.

The gene-level p-value combines the transcript p-values of a gene with Simes' method, and both levels are adjusted with the Benjamini-Hochberg procedure within each contrast, so that the three backends report the same columns.

With more than two groups, every contrast of contrasts is tested and the results are stacked, one block per contrast. "diffsplice" and "satuRn" fit all the groups once and test each contrast on that fit, so the dispersions use every sample; "DEXSeq" compares a model with and without a group effect, so it is fitted on the samples of the two groups of each contrast.

contrasts = "omnibus" gives one p-value per transcript for a difference among any of the groups, with contrast = "omnibus" and no fold-change (usage_log2FC is NA; look at the contrasts for the direction). "DEXSeq" compares its models with and without the group effect on all the samples. edgeR and satuRn test one contrast at a time, so for them the p-value is the Simes combination of the contrasts of every level against the first: the groups can only differ if some level differs from the first, and those contrasts share the reference, so they are positively correlated and Simes' test keeps its error rate.

Usage

test_usage(
  self,
  group,
  levels = NULL,
  contrasts = NULL,
  method = "diffsplice",
  assay = "counts",
  scale = NULL,
  min_count = 10,
  min_prop = 0.1,
  workers = 1L,
  gene_name = NULL,
  ...
)

Arguments

self

An IsoformicExperiment.

group

Column of col_data() with the groups to compare.

levels

Levels of group to use, reference first. Samples of other groups are ignored. NULL takes all the levels of group.

contrasts

The comparisons to test: NULL (default) compares every level with the first one; "pairwise" tests every pair of levels (the later level against the earlier one); "omnibus" tests whether usage differs among any of the groups (see Details); otherwise a list of c(numerator, denominator) pairs of levels, or one such pair. With two levels the default is the single contrast second against first.

method

"diffsplice", "satuRn" or "DEXSeq". The backend must be installed.

assay

Counts assay to test.

scale

Logical, divide the counts by the overdispersion of the replicates. NULL (default) does it when the object has replicates. In a read-split experiment (technical replicates of the same libraries, see the uncertainty article) this removed most false positives of "diffsplice" and "DEXSeq" and made their calls reproducible, so keep it on.

min_count, min_prop

Filters described above.

workers

Number of workers for satuRn and DEXSeq (parallel on Unix-like systems).

gene_name

Gene names or IDs to keep in the result. The filter is applied after the test.

...

Unused.

Value

A tibble with one row per tested transcript and contrast: gene_id, gene_name, transcript_id, transcript_name, transcript_type, contrast ("numerator vs denominator"), usage_log2FC (the change in usage on the log2 scale: for "diffsplice" the transcript against the other transcripts of the gene, for "satuRn" the log2 odds ratio, for "DEXSeq" the usage fold-change), pvalue, qvalue, gene_pvalue, gene_qvalue, method and scaled.

References

Dividing out quantification uncertainty enables assessment of differential transcript usage with limma and edgeR (2025). Nucleic Acids Research, doi:10.1093/nar/gkaf1305.

Gilis J et al. (2021). satuRn: scalable analysis of differential transcript usage for bulk and single-cell RNA-sequencing applications. F1000Research 10:374.

Anders S, Reyes A, Huber W (2012). Detecting differential usage of exons from RNA-seq data. Genome Research 22:2008-2017.

See also

test_productive_fraction(), test_apa(), de_confidence()

Examples

iso <- example_isoformic()
if (requireNamespace("edgeR", quietly = TRUE)) {
  res <- test_usage(iso, group = "condition", method = "diffsplice")
  head(res)
  table(res$contrast)
}
#> 
#> control vs preeclampsia 
#>                     164