NEWS

---

enrichit 0.2.7 (2026-10-28)

• GSEA no longer turns a degenerate input into a silently useless table: eps outside [0, 1] is rejected (it returned p = eps > 1), an all-zero ranking is an error instead of a frame of ES = NaN, a gene set covering the whole ranked list is excluded as fgsea does, and gene sets dropped for having no estimable p-value are counted instead of vanishing (2026-10-27)
• the multilevel "too few same-side permutations" guard is a rate, not a count, so nPermSimple below 10 no longer turns every pathway NA; default results are unchanged (2026-10-27)
• qvalueCutoff is applied row by row: one unestimable Storey q-value used to disable the cutoff for the whole table, so qvalueCutoff = 0.2 returned terms with q = 0.9 (clusterProfiler#710, #584, #614, #332). Related: a failed estimate is now a message() rather than a warning(), and NA pvalue/p.adjust no longer yield all-NA rows (2026-10-27)
• NSEA's permutation null no longer counts skipped permutations (nPerm reports what ran). Not reachable from ordinary input, so this is hardening (2026-10-27)
• ORA results keep DESize, UniverseSize and SetSize as columns, so GeneRatio == Count/DESize and friends are checkable from the table (clusterProfiler#703, #523, #618, #723) (2026-10-27)
• bayes_enrich(): default false_negative is 0.9, not 0.1, which drove every posterior to 0 on real KEGG data. Also rejects burnin + thin > n_iter and warns when no candidate reaches posterior_cutoff (2026-10-26)
• dsea() and dsea_scores(): design-aware GSEA and per-sample ssGSEA scores, with the contrast never guessed; as_ranked() and as_genes() adapt DESeq2 objects for it (2026-10-18)
• filter_gene_sets() is vectorized and ora_gson() skips the background intersection when the background was not restricted: 14,209 GO BP sets 14.0 s -> 0.9 s, single enrichGO() 19 s -> 4.2 s, 13-cluster compareCluster() 381 s -> 197 s. ora() returns geneID in input order instead of hash-set order (2026-10-18)
• remove the golden-reference script and data for the retired camera/roast/fry engines, and the design note for the never-implemented dsea_nb set test (2026-10-18)

enrichit 0.2.6 (2026-10-06)

• fix permutation-based GSEA p-values in the fixed and adaptive engines (method = "sample" / "permute") for the default two-sided (scoreType = "std") calculation: the numerator counted only same-sign permutations while the denominator divided by all of them, so p-values were too small under a symmetric null. The denominator now counts only permutations on the same side of zero as the observed ES, following the main fgsea convention (2026-10-06, closes #3)
• update the RcppEigen sparse-matrix interface in rwr_eigen_cpp() from MappedSparseMatrix to Map<SparseMatrix> for Eigen 5 compatibility (2026-10-06, PR #2)
• add a reportNA argument to gsea_gson() so pathways with NA multilevel p-values can be retained (2026-10-04, closes DOSE#88)
• fix GSEA leading-edge edge cases with one rank = 0L sentinel for zero-overlap, N_R == 0 and non-finite running-score paths; align as_gseaResult() with the same contract and document the gsea_leading_edge_details() return value (2026-10-04–2026-10-05, closes DOSE#46)
• keep qvalue as NA when estimation fails instead of falling back to p.adjust, and document the pvalue / p.adjust / qvalue semantics in as_enrichResult() and as_gseaResult() (2026-10-05, addresses DOSE#24)

enrichit 0.2.5 (2026-09-24)

• ORA results carry an oddsRatio column (Fisher's exact 2x2 odds ratio, next to FoldEnrichment). FoldEnrichment is a ratio of proportions; the odds ratio is the effect size the hypergeometric test is built on. It is the plain cross-product odds ratio and differs slightly from fisher.test()$estimate, which reports the conditional MLE (2026-09-22)
• report non-finite gene statistics instead of aborting with missing value where TRUE/FALSE needed. gsea_gson() checked is.unsorted() before validating its input, and a bare if (NA) stopped there, so an NA / NaN / Inf never reached the finiteness check that explains the problem. gseaScores() now validates too and returns NA rather than aborting when the running score is not finite (2026-09-22)
• add generic constructors as_enrichResult() / as_gseaResult() to build 'enrichit' result objects from external result tables (enrichr, g:Profiler, WebGestalt, fgsea, ...), with canonical column aliases, ID de-duplication, p-value validation, derived statistics (RichFactor / FoldEnrichment / zScore) when the query and background are known, and missing GSEA detail columns recomputed via gsea_leading_edge_details() (2026-09-21)

enrichit 0.2.4 (2026-09-17)

• stop emitting spurious no package '<...>' was found warnings when no input gene can be mapped (2026-09-17). check_gene_id() routed its notices through yulab.utils::yulab_msg(), which builds the package citation banner and expects a package name, so it called packageDescription() on the notice text itself and emitted three such warnings per run. The notices now go through message(), yulab.utils::yulab_msg() is no longer imported, and the sample of expected gene IDs no longer contains NA for sets with fewer than 100 genes

enrichit 0.2.3

• calibrate NSEA significance testing by switching the default to a whole-pipeline permutation null instead of GSEA's label-permutation test (2026-08-23). The legacy pipeline ran the label-permutation test on the network-diffused scores; diffusion induces strong autocorrelation between neighbouring genes, which violates the exchangeability assumption of that null and produces anti-conservative p-values (~2.5x the nominal false-positive rate). significance = "whole_pipeline" rebuilds the null of the enrichment score by permuting gene labels, re-diffusing with RWR and recomputing the score; significance = "internal" keeps the earlier formulation
• add significance, nPerm, seed and exponent to nsea() and nsea_gson(), and expose pvalueCutoff / pAdjustMethod in nsea_gson(). nPerm (default 1000) sets the number of whole-pipeline permutations with a smallest estimable p-value of 1/(nPerm + 1); seed makes the null reproducible; nsea_gson() now applies pvalueCutoff to both the raw and the adjusted p-value, matching gsea_gson()
• extend prepare_network() to accept a mechgraph object as well as an edge-list data.frame / matrix or a sparse matrix. The edges table must have from / to columns; a numeric score or weight column supplies the edge weight, otherwise unit weights are used. The class is duck-typed by name, so enrichit does not hard-depend on the mechgraph package
• add a fast C++ kernel gsea_es_all_cpp() that computes the classic weighted enrichment score for many gene sets at once from the hit positions only, cheap enough to evaluate on every whole-pipeline permutation
• share single implementations of the RWR diffusion (.nsea_diffuse()) and TF-IDF specificity weighting (.nsea_specific_weights()); in signed mode positive and negative seeds propagate in a single linear solve. Robust node-name extraction guards against Matrix accessor quirks on serialized sparse matrices
• add regression tests for whole-pipeline reproducibility and p-value calibration on null data

enrichit 0.2.2

• align gsea_gson() p-value filtering with historical clusterProfiler/DOSE behaviour: pvalueCutoff now requires both the raw and the adjusted p-value to pass (previously only the raw p-value was filtered), restoring significant pathway counts comparable to clusterProfiler <= 4.18.x (2026-08-14)
• clarify and harden the seed interface of GSEA (2026-08-14). gsea() treats seed = TRUE as a fixed default seed (consistent with the C++ default) instead of silently coercing it to the integer 1; gsea_gson() exposes an explicit seed argument; seed = FALSE (default) draws a fresh seed from R's RNG on each run, while a numeric seed (or set.seed() before the call) makes the result reproducible. Regression tests assert identical results across runs with a fixed seed

enrichit 0.2.1 (2026-08-04)

• fix gsea() to intersect gene sets with names(geneList) before applying minGSSize / maxGSSize, so the size filter constrains the actual overlap rather than the raw gene-set size; also guard gsea_gson() against NA pvalue rows leaking into the result table (2026-08-04, clusterProfiler#824)

enrichit 0.2.0 (2026-07-01)

• multi-omics integration at pathway and topology level (2026-06-24): aggregate_enrichment() for Late Fusion over several enrichResult / gseaResult objects; correlation-aware p-value aggregation (method = "brown") and method = "weighted_mean" for signed statistics with layer-specific weights in aggregate_omics(); mnsea() / mnsea_gson() for multi-layer network-based enrichment, with prepare_multilayer_network(), propagate_multilayer() and collapse_multilayer_scores(); mnseaResult to store diffusion results, collapsed scores, layer weights and cached explanation tables; get_mnsea_contribution() and extract_mnsea_subnetwork() for explanation-ready extraction. Provides the engine layer for clusterProfiler::mnseGO(), mnseKEGG(), mnseMKEGG() and mnseWP()
• Early Integration helpers (2026-06-23): aggregate_omics(), harmonize_ids() and select_features_for_ora(), with a conflict_policy argument ("keep_all" / "strict" / "penalty") for directional conflicts in signed statistics; early fusion before ora(), gsea() and nsea() through a decoupled aggregation layer
• get_omics_contribution() and classify_omics_pattern() for multi-omics contribution tracing
• Weighted Enrichment Analysis (2026-06-23): a weight argument in ora(), ora_gson(), gsea() and gsea_gson(); Weighted ORA uses Wallenius' noncentral hypergeometric distribution via BiasedUrn, Weighted GSEA fuses external weights with the ranked statistics
• Network-based Set Enrichment Analysis (2026-06-23): nsea() and nsea_gson() based on Random Walk with Restart, with mode = "signed" for bidirectional propagation; prepare_network() for parsing and normalizing edge lists or sparse matrices; fast RWR via RcppEigen sparse multiplication; a zero-dependency integration strategy for diffusion followed by multilevel GSEA. Provides the engine layer for clusterProfiler::nseGO(), nseKEGG(), nseMKEGG() and nseWP()
• align multilevel GSEA rank scaling with fgsea::prepareStats() to reduce drift against the long-used fgsea backend (2026-06-22): replace the fixed * 1e6 scaling in prepare_gsea_inputs() with fgsea-style total-weight normalization and integer rounding, with a regression test against fgsea::fgseaMultilevel()
• exclude zero-overlap gene sets from ora_gson() before multiple-testing correction (2026-06-22): Count = 0 rows stay out of p.adjust / qvalue so ORA results match historical DOSE / clusterProfiler behaviour, resolving inflated adjustment in enricher(), enrichKEGG() and compareCluster() (clusterProfiler#821, #819)

enrichit 0.1.5 (2026-06-16)

• add Bayesian term selection for ORA results (2026-06-16): bayes_enrich() estimates posterior probabilities of active explanatory terms from enrichResult objects, bayes_summary() returns them ordered by posterior. Candidate term spaces may come from top ranked rows, significant as.data.frame() rows, all x@result rows, or explicit term IDs. The output stays an enrichResult with posterior, posterior odds, Bayesian rank, active flag and covered-gene columns

enrichit 0.1.4 (2026-04-08)

• fix GSON helpers (2026-04-06): TERM2NAME() maps gsid to term names, TERMID2EXTID() returns gene vectors in the requested term order, and missing terms return character(0). Resolves malformed Description.* columns in clusterProfiler::groupGO()

enrichit 0.1.3 (2026-03-10)

• fix gsea_gson() (2026-03-10): force sorting of geneList so the result object is consistent with the input, which resolves incorrect metric/colour alignment in enrichplot::gseaplot()

enrichit 0.1.2 (2026-02-26)

• improve robustness of calculate_qvalue() (2026-02-02): handle a missing qvalue package, retry with different parameters, and handle invalid p-values (NA, infinite, out of range)
• validate p-values in ora_gson() (2026-02-02), warning on values outside [0, 1]
• use phyper instead of summing dhyper for better performance and precision in the C++ ORA p-value calculation (2026-02-02)

enrichit 0.1.1 (2026-02-01)

• setReadable() converts gene IDs to any type, not only SYMBOL (2026-01-21)
• add an organism slot to compareClusterResult (2026-01-20)

enrichit 0.1.0 (2026-01-18)

• fix bugs in gsea_gson() and ora_gson() (2026-01-11): fill missing columns (e.g. qvalues) with NA, and handle NA or duplicate gene-set IDs in result rownames
• calculate_qvalue() returns NA instead of NULL when qvalue estimation fails (2026-01-11)
• update ora_gson() output columns (2026-01-11)

enrichit 0.0.9 (2026-01-10)

• add leading edge analysis for GSEA (2026-01-10)

enrichit 0.0.8 (2025-12-22)

• fix p-value calculation in multilevel GSEA (2025-12-10), by learning the source code of 'fgsea'

enrichit 0.0.7

• add gseaScores() to calculate GSEA scores for a single gene set (2025-12-07)

enrichit 0.0.6

• add vignette (2025-12-07)

enrichit 0.0.5

• implement the multilevel GSEA algorithm (2025-12-06)

enrichit 0.0.4

• implement a simplified adaptive early-stopping GSEA algorithm (2025-12-05): run an initial batch of permutations, stop if p > threshold, otherwise increase the permutation count geometrically until the p-value stabilizes or the maximum is reached

enrichit 0.0.3

• implement ora_gson() and gsea_gson() as replacements for enricher_internal and GSEA_internal (2025-12-05)
• move helper functions and class definitions from DOSE to enrichit (2025-12-05), to extend this package as the base package for the clusterProfiler family

enrichit 0.0.2

• gsea(): Gene Set Enrichment Analysis using C++ via Rcpp (2025-12-04)

enrichit 0.0.1

• ora(): fast Over-Representation Analysis using C++ via Rcpp (2025-12-03)