lisaR separates gene-set enrichment from category-level inference. The quantities answer different questions and should not be used interchangeably.
Differential expression is calculated upstream. lisaR uses the
supplied signed ranking and gene-set memberships to run GSEA with
fgsea. It retains the gene-set normalised enrichment score
(NES), raw P value, adjusted P value and leading edge. The adjusted P
values are supplied by fgsea; they are not category-level P
values.
By default, GSEA uses the multilevel route. The selected ranking, gene-set size limits and other settings are recorded in the run. Gene-level DE adjusted P values, gene-set adjusted P values and category adjusted P values refer to different hypotheses.
A category’s mean_NES is the mean of its distinct
individually significant assigned sets with finite NES, at the
configured gene-set cutoff. It is not a category significance test.
Likewise, min_padj, the smallest member-set adjusted P
value, is not a category false-discovery rate.
Gene sets can overlap and belong to several categories. These assignments do not create independent observations. Point size counts gene sets, not samples or independent biological mechanisms.
For each DE analysis, lisaR applies robust Hommel closed testing
(hommel(..., simes = FALSE)) to the raw gene-set P values.
The family contains all eligible unique gene-set hypotheses across the
configured semantic collections, including unclassified sets. HALLMARKS
is excluded and retains its direct gene-set results. Eligibility, not
the GSEA significance cutoff or NES direction, defines this family.
Repeated instances of the same hypothesis across collections count once, using the largest P value conservatively. Hypothesis identity preserves the pathway identifier, analysis, ranked universe, effective membership and scoring settings. Eligible hypotheses without a usable raw P value remain in the family with P = 1; they are not silently dropped.
Robust closed testing allows arbitrary dependence, assuming valid individual input P values. It does not correct an inappropriate upstream design or an invalid individual test.
The adjusted category P tests the null that none of the category’s evaluable member sets is enriched. It is already family-adjusted; lisaR does not apply another BH correction to it.
The category card reports the exact value. Stars show only the strongest threshold reached:
| Mark | Adjusted category P |
|---|---|
* |
≤ 0.05 |
** |
≤ 0.01 |
*** |
≤ 0.001 |
**** |
≤ 0.0001 |
A significant category supports at least one enriched member set. It does not establish coherent activation of the whole category or the direction of its effect. An unmarked category can be nonsignificant or not evaluable.
At the default alpha of 0.05, d/N gives a simultaneous 95% lower confidence bound: at least d of N eligible member hypotheses with usable raw P values are supported as enriched. The tests use the conservative deduplicated family P values described above. A zero bound does not demonstrate absence; N = 0 means not evaluable.
This is a minimum supported count, not the exact true count or the number passing the GSEA display cutoff. Do not add these bounds across overlapping categories. Check the card’s analysed/total coverage and missing-P information.
Each DE analysis has its own inference family. A difference in mean NES, or significance in one analysis but not another, does not establish a between-analysis difference. ON minus PRE and other LISA profile comparisons remain descriptive. Testing a condition difference requires an appropriate upstream statistical model.
Goeman et al. (2019), Simultaneous control of all false discovery proportions in large-scale multiple hypothesis testing, Biometrika 106, 841–856. doi:10.1093/biomet/asz041.
See the hommel documentation and fgsea documentation for their algorithms. The licence and citation guide provides the software and study attribution.
Return to Advanced guides.