| dsea | R Documentation |
dsea() takes counts and a design and returns GSEA results for a contrast:
it transforms the counts, reduces the design to a per-gene statistic, ranks
the genes and runs the GSEA engine. dsea_scores() stops one step earlier and
returns the gene set by sample score matrix instead.
dsea(
x,
gene_sets,
design,
contrast = NULL,
transform = c("vst", "rlog"),
min_size = 10L,
max_size = 500L,
...
)
dsea_scores(
x,
gene_sets,
transform = c("vst", "rlog", "logcpm"),
min_size = 10L,
max_size = 500L,
exponent = 1,
...
)
x |
a counts matrix, or a |
gene_sets |
a named list of gene identifier vectors, identifiers matching
|
design |
for |
contrast |
a column name or index of |
transform |
|
min_size, max_size |
gene set size limits, applied to the overlap with the genes actually present. |
... |
for |
exponent |
for |
Why the ranking is what it is. Genes are ranked by the difference in mean transformed expression between the two sides of the contrast. That looks like the weakest available choice and it is the one the benchmark supported: ranking by a moderated t or by DESeq2's Wald statistic was no better under uniform or heterogeneous effects, and materially worse when the affected genes were the noisier ones (recall 0.826 for Wald against 0.979 for the mean difference under an eightfold dispersion multiplier; across a 1x/2x/4x series the Wald statistic falls 0.995 to 0.950 while the mean difference stays flat). GSEA's statistic already equalises genes by rank, so layering a variance weight on top discards real signal at the extremes rather than adding power.
Which column may be the contrast. A design column encodes a
comparison only when it is not constant and not a group indicator under
cell-means coding. The intercept is rejected because ranking by it orders
genes by mean expression, which turns the analysis into "highly expressed
pathways"; a column of ~ 0 + condition is rejected for the same reason,
since its coefficient is one group's mean rather than a difference. A
continuous column (a dose slope, say) is allowed.
The transform uses design(x), the ranking uses design. When
x is a DESeqDataSet, vst(blind = FALSE) reads the design stored in the
object, while the ranking reads the design argument. If the two differ the
call warns rather than blending two models: refit the object with the intended
design, or pass a counts matrix so the transform can use design.
What the design is used for. For a two-group contrast the ranking is the group mean difference. For a design with nuisance terms it is the contrast coefficient from the fitted linear model, so batch and other covariates are removed before ranking rather than ignored.
What this is not. It is a design-aware front end to GSEA, not a new set test. Directions are relative to the ranking: a set scoring high sits toward the top, which is the direction the contrast was oriented in.
dsea() returns the data.frame from gsea(), sorted by p-value, with two
columns added: padj, the Benjamini-Hochberg adjusted p-value, and
contrast, the design column that was tested. Everything else is GSEA's own
output, so the table can be filtered and plotted like any other GSEA result.
dsea_scores() returns a list with scores (gene sets by samples) and meta
recording the transform and scorer settings. meta$set_sizes reports the
number of members actually scored, after intersecting with the data, and
meta$raw_set_sizes the input size, so the two can be compared.
gsea, ssgsea_scores
## Not run:
dds <- DESeq2::DESeq(DESeq2::DESeqDataSetFromMatrix(counts, coldata, ~ condition))
design <- model.matrix(~ condition, coldata)
res <- dsea(dds, gene_sets, design = design, contrast = "conditionB")
head(res[order(res$pvalue), ])
sc <- dsea_scores(dds, gene_sets)
dim(sc$scores)
## End(Not run)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.