NEWS

---

clusterProfiler 4.21.5

interpret()

• interpret() takes prior: another tool's answer (a SingleR or scGPT label, a TF-activity ranking), sent after the evidence as a claim to check rather than as evidence. The record adds conclusion, prior_verdict and prior_rationale. On PBMC 10k with SingleR labels, 12 of 14 verdicts on misassigned subtypes were Corrected and 1 of 14 on correct ones (2026-11-02, Mon)
• reduce_redundancy() defaults to 0.7, the default of the simplify() it calls, and its documentation now describes what simplify() does rather than the earlier greedy selection. Cisplatin GO ORA keeps 81 of 141 terms (0.5: 40), geneList GO ORA 76 of 195; KEGG lists are kept whole (2026-11-02, Mon)

GO annotation cache

• The built GO gene-set object is cached per (organism, ont, keytype) on top of the annotation cache: 1.4 s of a 2.8 s warm BP call, and compareCluster() pays it once per cluster (13 clusters: 39.5 s against 19.7 s, results identical). Up to three objects are kept, least recently used first out, and an evidence-filtered annotation is never cached (2026-10-29, Thu)

pathway-level mechanism evidence

• pathway_evidence() tests enriched pathways against measured perturbation data instead of hunting for transcription factors. Given candidate drivers it resolves each to its knockdown datasets and reports A/B/C with a statement column holding the claim a perturbation experiment could refute. Candidates are supplied by the caller on purpose: picking them automatically carried no usable signal (2026-10-30, Fri)
• Level A requires replication across two cell lines, a direction that agrees with the pathway, and a pathway direction that does not come from the pathway's own members. Before this, a member-derived direction earned B on a tautology (the agreeing fraction is at least 50% by construction) and significance alone earned A even when the perturbation moved the pathway the way it was already running (2026-10-30, Fri)
• reason records why a pathway stopped short of A -- single-cell-line, direction-conflict, no-dataset, too-small, not-significant, no-test-completed, member-derived, low-support and the input-side cases -- so "nothing found" is distinguishable from "nothing could be looked at". A failed GPSAdb call is no-test-completed with a warning, not weak evidence, and a timeout no longer reads as "no perturbation test identified a driver" (2026-10-31, Sat)
• mechanism_preflight() decides before a run whether the data can carry a driver claim, and returns the numbers behind the decision. On eleven real datasets it blocked 7 of 7 that gave no usable answer and 0 of 4 that did, and every reason was visible in advance (2026-11-03, Mon)
• mechanism_report() turns reason into "what to do next" and gives the candidate's full inventory, separating data, analysis and "no experiment would change this". It stops recommending experiments when the preflight says attribution is not supported, because adding a cell line cannot rescue a dataset with 23 differentially expressed genes, and it resolves aliases so a report on GR finds NR3C1's dataset (2026-11-03, Mon)
• mechanism_report() gains network, min_overlap and global_effect_de. module groups drivers whose regulator-target sets overlap, because paralogues, subunits of one complex and parallel pathway arms cannot be told apart by an expression experiment: on 17 benchmark datasets this covers the true driver 7/10 against 6/10 for the single top-1 and lifts module-level PPV from 19% to 54%, recovering HIF1A-EPAS1, STAT1-IRF1, XBP1-ATF6, TP53-TP73 and NR3C1-AR with no prior annotation. global_effect_reason marks a driver whose knockdown moved more than global_effect_de genes (default 3000), since such a perturbation drags any pathway of this size and a supported verdict on it says "large perturbation", not "specific driver". Neither field changes the A/B/C class (2026-11-08, Sat)
• mechanism_reproduction() reports how many of a driver's independent perturbation datasets move the pathway the same way, and pathway_evidence(reproduction = ) uses it to word a single-cell-line row. The majority sign is the mode of the per-dataset signs rather than sign(sum(shift)), so one dataset with a large effect cannot overturn two smaller ones that agree, and datasets whose members are all NA are dropped before k is counted (2026-11-07, Sat)
• mechanism_discover() ranks candidate transcription factors for a profile, so pathway_evidence() gets a shortlist instead of searching 265 candidates at 1-2 s per query. On 17 experiments the true driver ranked first in 43%, the top five in 71% and the top ten in 93% (median 2) among the 14 drivers a typical network contains; the other three are absent from CollecTRI. The shortlist is not a verdict: verifying the top five let 70% of true drivers through but also 40% of the decoys, a positive predictive value near 19%, and the decoys that pass are the driver's neighbours -- ATF4 and ATF6 beside ER stress, EPAS1 beside HIF1A, IRF1 downstream of STAT1. The score also rises with the number of annotated targets, even among candidates that are not the driver; n_targets and share report this rather than correcting it (2026-11-07, Sat)
• The attribution route explored on branch mechanism-decompose is not merged. Five statistics over one reference matrix landed on the same recall/precision trade-off, the best pointing at the true driver in 6 of 9 datasets against 6 of 9 for this implementation, with two failure modes no reweighting fixed. dev_docs/mechanism-attribution-postmortem.md records the five things not to try again (2026-11-03, Mon)
• Fixes: a level-B row now says a single-cell-line limit is the data rather than the evidence, while a direction conflict keeps its own wording; a member-derived direction no longer reads "100% of members moved up, below the support floor of 50%"; reason reports too-small ahead of low-support; coherent-refuted is separated from coherent-untested; an empty result, an outage and an unreachable level A each warn instead of passing as findings; print() runs for pathway_evidence and a column subset no longer breaks it; empty and NA tokens are not counted as members; alias lookup runs in both directions, so drivers = "GR" finds NR3C1 (2026-10-31, Sat)

mechanism analysis: collect, adjudicate, write

• evidence() grades a gene list against upstream collections and returns an evidenceBundle whose $regulators carries evidence_kind (measured, predicted, observed or none) and a level. supported requires one measured source corroborated by a second: counting sources is not counting evidence when two predictions agree (2026-10-25, Sun)
• adjudicate(), formerly consensus(), decides which claims survive a change of evidence type and returns verdicts, the evidence behind them, diagnostics, parameters and a provenance fingerprint; interpret() requires an adjudication, so that decision cannot be skipped on the way to prose (2026-10-28, Wed)
• upstream = "perturbagen" asks GPSAdb which measured perturbation moves the input genes -- the only source that reaches level A. upstream = "tfunion" adds ENCODE ChIP-seq, ChEA 2022, TF_Perturbations and TRRUST: 2869 sets naming 315 factors C3:TFT has no set for, including EPAS1, NFKB1, TP53, CTCF, ESR1, JUN and FOS (2026-10-27, Tue)
• mechanism_writable(), mechanism_xref(), collapse_findings(), mechanism_graph() with plot_mechanism(), mechanism_sets(), consensus_findings(), mechanism_significant() and mechanism_alias_regulator() are exported, and simplify() removes redundancy from any collection rather than only GO (#743, 2026-10-26, Mon)
• Fixes: perturbation_effect no longer calls an activator a repressor when the caller's own contrast knocks that factor down, and measured_direction now reads the same gene set it does; unsupported_claims() no longer drops real symbols such as P53 and P300, no longer exempts a rejected regulator because its symbol was also an input gene, and no longer excuses one whose symbol prefixes an unrelated gene; print() reports coverage and a failed source instead of silence, and no longer counts aliases against candidate regulators (2026-10-29, Thu)

clusterProfiler 4.21.4

fixes

• The headline verdict could be produced by a symbol collision. TRANSFAC's NF1 is Nuclear Factor I while the HGNC symbol NF1 is neurofibromin, and matched on the string the motif term and a knockout dataset "agreed". A collided label can no longer reach supported (2026-10-25, Sun)
• A protein name that maps to a gene is no longer treated as a family label. TRANSFAC names nuclear receptors by their protein -- GR, PXR, FXR, HIF2A -- none of which are HGNC symbols, so the perturbation channel skipped them as families and GR could not reach supported on a glucocorticoid experiment, where it is the one answer that is certain (2026-10-26, Mon)
• Regulators are looked up in perturbation databases under every name they are known by, in both directions: the collections say GR where the database says NR3C1, and HIF1 where it says HIF1A. Querying one spelling returned nothing, which is indistinguishable from "never perturbed" (2026-10-27, Tue)
• mechanism_is_family_label() no longer reads a hyphenated symbol as a family: TRANSFAC writes NKX2_5 where HGNC writes NKX2-5, which also stripped those genes of their measured evidence kind (2026-10-25, Sun)
• The perturbation sweep no longer runs concurrently: GPSAdb returns different data when queried in parallel and reports nothing unusual, so the supported set flipped between runs. options(clusterProfiler_perturb_workers = 4) takes the trade knowingly (2026-10-26, Mon)
• A p-value tie between two datasets for one regulator is settled by system match, then p-value, then genes moved, then dataset id -- the last unique, so the ordering is total rather than dependent on arrival order (2026-10-26, Mon)
• Corroboration now requires the sources to cite different datasets, not merely different source names: perturbagen and biotrigger are two views of one GPSAdb experiment, so when both cited the same dataset the server was agreeing with itself (2026-10-26, Mon)
• evidence_similarity() and evidence_biotrigger() are exported but no longer part of the bundle, after measuring what they contribute: biotrigger raised the supported count without adding a transcription factor, and the additions were genes with no bearing on the system under study (2026-10-26, Mon)
• microRNA collections are no longer tested by default. Predicted seed matches gave 7 significant miRNAs among the up-regulated genes and none among the down ones, the opposite of what a repressor predicts, so the direction that would make the result interpretable is absent (2026-10-25, Sun)
• upstream = also accepts "cellstate" (C8) and "tissue" (C4:CGN), whose set identifiers already name the entity and are passed through unchanged rather than through the transcription-factor rules, which would turn a cardiac signature into the gene HPX (2026-10-25, Sun)
• network = TRUE is documented as informational: $ppi carries the edges but nothing consumes them into a verdict, so the channel is reference material rather than corroboration (2026-10-28, Wed)
• evidence_ppi() grades STRING edges by the channel behind them: escore or dscore at 0.4 or above gives B, anything else C, and no STRING channel measures a perturbation so an edge never reaches A. pscore is deliberately unused -- it is 0 even for NFKB1-NFKBIA, which carries escore = 0.999 (2026-10-25, Sun)
• The session cache directory is no longer forced to tempdir() on attach, which overwrote a directory set in .Rprofile and made downloaded collections miss their cache (2026-10-25, Sun)
• Switching ontology no longer evicts the GO annotation cache. The cache was keyed on the ontology and stored only the last call's annotation, so BP then MF then BP rebuilt it every time: five alternating enrichGO() calls took 85.5 s of which the enrichment itself is 0.4 s, and now take 11.6 s, with identical results (2026-10-27, Tue)
• A gseaResult is now tested with GSEA, not re-tested with ORA. Taking the union of core_enrichment and running enricher() on it discarded the ranking and the sign, so the "GSEA path" was the ORA path with a different input and the two agreed against ground truth because they were the same test (2026-10-27, Tue)
• The two front ends can legitimately disagree: on the airway list GR_Q6 has ORA padj 0.0002 and GSEA padj 0.166, both correct, because one asks about over-representation among significant genes and the other about moving monotonically with the phenotype. Use ORA for the first question, GSEA for the second (2026-10-27, Tue)
• An interrupted Enrichr download no longer poisons the cache: the union libraries count as cached only when a .meta marker sits beside the GMT, as the MSigDB files already did. A truncated file used to be reused, shrinking the collection and grading every regulator outside it absent instead of failed (2026-10-28, Wed)
• tft combined with upstream = "tfunion" is an error rather than a silent no-op. tft = "gtrd" is how a caller escapes the family and consensus labels no perturbation dataset can answer for, and dropping it turned a 504-set search into a 2869-set one without a word (2026-10-28, Wed)
• enrichDAVID() applies qvalueCutoff row by row: an NA q-value means Storey estimation failed for that row, so it bypasses this cutoff instead of switching the cutoff off for every term (clusterProfiler#710, #584, #614, #332, 2026-10-27, Tue)
• gseKEGG() reports the same category / subcategory columns as enrichKEGG(); a reader that filtered on category silently behaved differently depending on which function produced the result (#787, 2026-10-26, Mon)

interpret()

• $pathways carries pathway_padj, the pathway's own enrichment p-value, beside padj_within_pathway, which is the attribution's. The two answer different questions and without the first the narrative context could only order by attribution strength: on the hypoxia contrast that put "carbohydrate metabolic process" above "response to hypoxia" (2026-10-28, Wed)
• An adjudicated run now adds one fixed sentence to its system prompt: only supported findings may be stated as mechanisms, single_source is a hypothesis, absent is not "does not exist", and untested is not an absence. It is appended rather than substituted, so the four legacy input types keep the prompt they had byte for byte -- a test compares the strings rather than trusting it (2026-10-28, Wed)
• The pathway layer now reaches the narrative. interpret(adjudication) lists which regulator explains each enriched pathway, in its own prompt section marked as context, and only for entities the adjudication already graded -- so an attribution can never introduce a regulator or make a claim sound stronger. The "context, not evidence" instruction travels inside that section rather than in the system prompt, which would have changed the bytes of the four legacy input types (2026-10-28, Wed)
• network = TRUE is documented as reference material: $ppi carries the graded STRING edges but nothing turns an edge into a corroboration, so it changes no verdict. A test pins that, so the assumption cannot creep back (2026-10-28, Wed)
• fix interpret() reverting a caller's adjudication. It re-ran adjudicate() with the default min_independent in three places -- the prompt builder and the two post-hoc evidence tables -- so a caller who adjudicated at a stricter setting had it silently overridden, which is the one thing making adjudication a separate step exists to prevent. The carried verdict table now travels through the whole call (2026-10-28, Wed)
• interpret() gains evidence_status(), a deterministic grounding check: every gene symbol the interpretation names is classified as observed or out_of_evidence. No model is called, so unlike a second model the check is exact (2026-09-26, Sat)
• interpret(), interpret_agent() and interpret_hierarchical() stamp their results with an execution record, read back with provenance(): the model actually addressed, the decoding parameters and whether the response confirmed them, digests of the prompts and response, token usage and finish_reason. A remote model is not a versionable dependency, and without the record a transport failure arrives as an ordinary result (2026-09-03, Thu)
• interpret() reads per-gene statistics through gene_stats, a differential-expression table whose gene, log fold change and adjusted p-value columns are recognised by name or row names, and lists the genes of the enriched terms with those values (2026-10-18, Sun)
• interpret() no longer truncates the genes it sends: every input gene is listed, every fold change of the sent terms' genes is given, and an agent cluster without enriched terms lists every input gene. n_pathways stays NULL, because on five public lists a lower cap named fewer themes; lower reduce_threshold instead (2026-10-11, Sun)
• interpret() now interprets an enrichment result and nothing else, so its task and prior arguments are removed; combining enrichment evidence with other tools moves to interpret_agent(prior = ), whose Synthesizer returns conclusion and prior_verdict (2026-09-30, Wed)
• reduce_redundancy() and reduce_go_redundancy() are exported and interpret() takes reduce / reduce_threshold, so every significant theme reaches the model once: GO by Wang semantic similarity, other collections by Jaccard overlap (2026-09-20, Sun)
• interpret() and interpret_agent() take reasoning, passed through to the provider and recorded in the execution stamp; left NULL the request is byte-identical to one made before the argument existed (2026-09-20, Sun)
• interpret_agent() gains an optional Reviewer agent (review = TRUE) that audits the draft, reporting grounding separately from it (2026-09-03, Thu)
• interpret() and interpret_agent() warn when add_ppi = TRUE and the STRING query fails, instead of silently interpreting without PPI evidence (2026-09-30, Wed)
• fix interpret() to read gene names from the object's symbols rather than from a mismatched identifier column; an unparsed model response is reported instead of returning NULL (2026-09-26, Sat)
• digest is now in Suggests, used for the prompt and response digests of provenance() (2026-10-23, Fri)

clusterProfiler 4.21.3

• fix get_kegg_species() and kegg_taxa() for KEGG's retirement of the organism special database name: https://rest.kegg.jp/list/organism now answers HTTP 400 while the rest of the REST API keeps working, so the weekly "Update KEGG Data" workflow and any user calling search_kegg_organism(use_internal_data = FALSE) failed with "cannot open URL ... HTTP status was '400 Bad Request'". Both now read /list/genome, which returns the same organisms as <T number>\t<code>; <scientific name> (<common name>); the species list is parsed from the name and code of that single column. An organism without a common name keeps its scientific name in common_name, as it did before (2026-10-18, Sun)
• restore_hybrid_result() and restore_entrez_result_ids() map a gene set's IDs back to the source keyType with one vectorized pass instead of one labels[ids] + unname() per set. On a BP result (~5,900 tested sets, 390k ID slots) that loop cost 7.5 s per call and compareCluster() calls it once per cluster; the vectorized form takes 0.2 s and is byte-identical, including list names, empty sets and geneID order (pinned by a test). hybrid_ids_for_input() likewise replaces a per-id mapping$source == id scan with a single split and named lookup (negligible wall time at cluster-sized inputs, but no longer quadratic in the mapping size). Measured on the PBMC 10k 13-cluster SYMBOL compareCluster() together with the enrichit fixes: 140 s -> about 70 s (run-to-run 69-78 s), with 7,489 rows, p-values (difference 0), Counts and geneID sets identical (2026-10-18, Sun)
• enrichGO() caches the hybrid GSON it builds for a non-ENTREZID keyType when some input IDs cannot be mapped to Entrez. Without a universe, that build reads every key of the OrgDb and therefore depends only on (organism, ont, keyType), but compareCluster() calls enrichGO() once per cluster and rebuilt it every time (7.7 s per call on human GO BP, PBMC 10k, 13 clusters, 2026-10-18). The hybrid path itself is unchanged and still fires ONLY when an input ID has no Entrez mapping -- an all-mappable input keeps the existing fast path and never builds hybrid data -- while a universe-scoped or evidence-filtered call neither reads nor writes this cache (2026-10-18, Sun)
• read.gmt.pc() now reads the fields of a Pathway Commons gene set description by their labels rather than by splitting the string at every "; ". A pathway name that holds a semicolon, such as Phospholipase C-mediated cascade; FGFR2, is no longer cut short and its datasource, organism and idtype stay in their own columns; previously the extra pieces were merged into the last field, so the fields were shifted one to the left and nothing said so. A description that declares fewer fields than expected returns NA for the missing one instead of failing the whole read, and enrichPC() and gsePC() take the species from the first gene set that declares an organism (#789, 2026-10-08, Thu)
• fix enrichGO() with a non-ENTREZID keyType when some source IDs cannot be mapped to Entrez: the memory-saving Entrez annotation remains the fast path, while unmapped source IDs are added through direct native-key annotations and analyzed in one hybrid GSON; when universe is omitted, the effective annotated hybrid background is constructed and reported by message instead of silently dropping IDs (#829, 2026-10-03, Sat)
• restore the original source keyType consistently after the Entrez optimization and readable conversion, including result metadata and projected gene-set memberships (#829, 2026-10-03, Sat)

clusterProfiler 4.21.2

• enrichMKEGG() and gseMKEGG() accept a GSON object for organism, as enrichKEGG()/gseKEGG() already did, so KEGG module enrichment can run against a locally built annotation (gson_KEGG()) when KEGG is unreachable; they also now stop with "organism should be a species name or a GSON object" instead of an unrelated error from the organism mapper. The organism documentation for the KEGG functions now mentions the GSON option at all (2026-09-22, Tue, #623)
• enrichGO() takes an evidence argument to restrict the GO annotation by evidence code before the analysis, e.g. evidence = c("IDA", "IPI", "IMP") to keep manually supported annotations or evidence = "IEA" for electronic ones only; dropping IEP is the usual way to avoid circular reasoning when the gene sets come from co-expression clustering. Because the annotation (and hence the background) changes, the result is not simply a subset of the unfiltered run, and filtered runs are not cached (2026-09-22, Tue, #160)
• enrichGO() no longer aborts with missing value where TRUE/FALSE needed when an OrgDb's organism cannot be identified — an AnnotationHub OrgDb, for example. The cached-GO-data check compared the cached organism against get_organism(OrgDb) in a bare if(), and that comparison is NA for such an OrgDb, so the second call in a session failed with an error that named none of this; an unidentifiable organism is now treated as "not cached" and the annotation is rebuilt (2026-09-22, Tue, #232)
• Gff2GeneTable() now reads Ensembl/GTF-style GFF3 as well as NCBI/RefSeq files: it accepts a gene_id= attribute (Ensembl) in addition to GeneID= (NCBI) and takes the gene symbol from Name=/gene_name= when gene= is absent. Previously an Ensembl file left every key NA, and merging two all-NA key columns is a cartesian product — useful-looking but wrong on a small file and negative length vectors are not allowed on a real genome. A file with neither identifier style now stops with a message saying what is expected (2026-09-22, Tue, #193)
• add_go_level() appends a level column giving the GO level of every enriched term (1 = the ontology root), so results can be kept to a band of levels with e.g. subset(res, level >= 3 & level <= 6); the level is taken from the ontology the term itself belongs to, so ont = "ALL" and compareCluster() results are labelled term by term (2026-09-22, Tue, #793)
• simplify() now infers the ontology from the GO IDs when the result object does not carry one, instead of refusing: this covers a gseaResult whose @setType holds a collection name rather than an ontology (MSigDB C5, a gson GO file, a TERM2GENE built from GO annotations) as well as an enrichResult from enricher(), which previously required assigning x@ontology by hand; results spanning several ontologies are simplified as GOALL, and non-GO gene sets are still rejected with a message that says so (2026-09-22, Tue, #753, #369)
• compareCluster() now warns when a non-character universe is supplied: it wraps each per-cluster call in suppressMessages(), which used to swallow the underlying enrichment function's "universe is not in character and will be ignored" message, so a numeric universe appeared to be accepted while having no effect (2026-09-22, Tue, #654)
• gsePC() and gseWP() now expose an eps argument and forward it to GSEA() (and on to enrichit::gsea_gson()), matching GSEA(), gseGO(), gseKEGG() and gseMKEGG() (2026-09-22, Tue)
• fix gson_cpd() to use the compound KEGG REST endpoint; enrichKEGG(organism = 'cpd') had been failing since KEGG retired /link/cpd/pathway, which now returns HTTP 400 (2026-09-22, Tue, #828)
• kegg_rest() now reports a failed download where it happens, telling an unreachable network apart from an HTTP error status and from an empty response; previously these surfaced much later as an obscure error (2026-09-22, Tue)
• download.KEGG.Path() and download.KEGG.Module() now stop with an actionable message when the downloaded annotation is empty, or when the gene/pathway table shares no ID with the pathway-name table, instead of returning an empty annotation that downstream surfaced as a misleading "No gene can be mapped"; this replaces an is.null() guard that could never fire (2026-09-22, Tue, #146, #378, #561, #646)
• align GSEA significance filtering with the historical behavior of clusterProfiler <= 4.18.x: pvalueCutoff now requires both the raw p-value and the adjusted p-value to pass the cutoff, so significant-pathway counts from GSEA()/gseGO()/gseKEGG()/gseMKEGG() are again comparable with the DOSE/fgsea backend (2026-08-14, Thu)
• restore the seed argument on GSEA(), gseGO(), gseMKEGG() and gseKEGG() and forward it to enrichit::gsea_gson(), so permutation-based GSEA results can be reproduced with a fixed seed (or set.seed() before the call); the argument had been dropped when the enrichit engine was introduced (2026-08-14, Thu)
• fix get_data_from_KEGG_db() to strip species suffix from pathway names when use_internal_data=TRUE, matching the online path behaviour (2026-08-04, Tue, #783)
• fix enrichGO() to check whether the OrgDb supports ENTREZID before attempting the ENTREZID-remapping optimisation; custom OrgDb packages without ENTREZID keytype (e.g., non-model organisms) now fall back to the original get_GO_data path instead of throwing an error (2026-08-04, Tue, #823)

clusterProfiler 4.21.1

• support UniProt accession -> KO conversion in bitr_kegg() via KEGG REST and export gson_KO() for KEGG Orthology workflows (2026-06-24, Wed)
• add nseGO(), nseKEGG(), nseMKEGG(), nseWP(), mnseGO(), mnseKEGG(), mnseMKEGG(), and mnseWP() as high-level topology-aware enrichment wrappers powered by enrichit, with metadata backfilling and lightweight wrapper tests (2026-06-24, Wed)
• re-export aggregate_omics(), aggregate_enrichment(), harmonize_ids(), and select_features_for_ora() from enrichit for high-level multi-omics workflows (2026-06-24, Wed)
• get_ppi_network to download and cache the full STRING PPI network for a specific species (2026-06-23, Tue)
  • the getPPI is now alias to get_ppi
• fix GSEA wrappers to restore eps, forward extra arguments such as sampleSize to enrichit::gsea_gson(), and align the GSEA method documentation with the current enrichit interface (2026-06-22, Mon, #822)
• re-export bayes_enrich() and bayes_summary() from enrichit (>= 0.1.5) for Bayesian term selection and posterior-ranked summaries of ORA enrichment results (2026-06-16, Tue)

clusterProfiler 4.20.0

• Bioconductor RELEASE_3_23 (2026-04-29, Wed)

clusterProfiler 4.19.8

• fix: map non-ENTREZID universe in enrichGO (2026-04-22, Wed)

clusterProfiler 4.19.7

• interpret(), interpret_agent(), and interpret_hierarchical() now use aisdk's global default model when model = NULL, so users can switch the package-wide default with aisdk::set_model() while still overriding per call with an explicit model argument (2026-03-31, Tue)

clusterProfiler 4.19.6

• update ko2name() to robustly parse KO names via KEGG REST, support vector input with deduplication, and return NA when NAME is missing (2026-02-25, Wed)
• bug fixed for plot.interpret (2026-02-05, Thu)

clusterProfiler 4.19.5

• interpret() prompt optimized with 'Comparative Analysis' and 'Rule of Exclusion' to better distinguish cell types with shared functions (e.g. NK vs CD8+ T cells) using specific marker genes (2026-01-22, Thu)
• fixed a bug in interpret() where empty interpretation results were returned due to incorrect list structure handling in process_enrichment_input (2026-01-22, Thu)
• interpret() now considers specific marker genes in cell type annotation to avoid key markers being overshadowed by general pathways (2026-01-21, Wed)
• optimize enrichGO() to avoid memory boom when keyType is not ENTREZID (2026-01-21, Wed, #805)
• gson_GO_local() to support local GO annotation by adding ancestral terms (2026-01-21, Wed)
• interpret() implements a gene-based fallback mode for clusters with no enriched pathways, ensuring comprehensive analysis (2026-01-20, Tue)
• plot() method for interpretation object to visualize the LLM-inferred regulatory network using ggtangle (2026-01-20, Tue)
• interpret_agent() supports multi-agent system (Deep Mode) for interpretation (2026-01-20, Tue)
  • Agent Cleaner: Filters noise and selects relevant pathways
  • Agent Detective: Identifies key regulators and functional modules using PPI/TF data
  • Agent Synthesizer: Synthesizes findings into a coherent narrative
• interpret() supports 'Knowledge-Guided Interpretation' (2026-01-20, Tue)
  • add_ppi parameter to integrate PPI network and identify hub genes
  • gene_fold_change parameter to incorporate expression levels
  • Mixed-source enrichment analysis support (e.g. Pathways + TFs) for causal integration
  • LLM-guided network refinement to output core regulatory networks
• interpret_hierarchical() for hierarchical interpretation (e.g. Major -> Minor clusters) (2026-01-20, Tue)
• interpret() now supports prior parameter for reference-guided interpretation (e.g. from SingleR/scGPT) (2026-01-20, Tue)

clusterProfiler 4.19.4

• interpret() now supports task parameter to specify the task: 'interpretation', 'annotation' and 'phenotyping' (2025-01-18, Sat)
• interpret() supports enrichResult, gseaResult, compareClusterResult and list of enrichment results (2025-01-18, Sat)

clusterProfiler 4.19.3

• instead of packing KEGG cache data in the package, we now download it from https://yulab-smu.top/clusterProfiler (2025-12-15, Mon)
• add github action to automatically update KEGG cache data (2025-12-09, Tue)
• use 'enrichit' as engine for enrichment analysis (2025-12-07, Sun)

clusterProfiler 4.19.2

• use 'quarto' as vignette engine (2025-11-20, Thu)
• update KEGG cache data (2025-11-20, Thu, #792)
  • number of KEGG pathway with category information: 582
  • Number of species: 11344

clusterProfiler 4.19.1

• bug fixed in enrichPC (2025-11-01, Sat, #789)

clusterProfiler 4.18.0

• Bioconductor RELEASE_3_22 (2025-10-30, Thu)

clusterProfiler 4.16.0

• Bioconductor RELEASE_3_21 (2025-04-17, Thu)

clusterProfiler 4.15.2

• more general regular pattern to remove species suffix in KEGG pathway name (2025-02-27, Thu)
• remove input duplicated genes in groupGO() (2024-11-29, Fri, #741)

clusterProfiler 4.15.1

• simplify() keeps the most informative term if there exist multiple terms that meets the criteria (2024-11-29, Fri, #744)
• add 'RichFactor', 'FoldEnrichment' and 'zScore' in enrichDAVID() result (2024-11-12, Tue)
• update DAVID Web Service URL to make enrichDAVID() work properly (2024-11-09, Sat)
  • https://davidbioinformatics.nih.gov/content.jsp?file=WS.html
• add new citation (2024-11-07, Thu)
  • https://doi.org/10.1016/j.xinn.2024.100722

clusterProfiler 4.14.0

• Bioconductor RELEASE_3_20 (2024-10-30, Wed)

clusterProfiler 4.13.4

• re-export DOSE::enrichDO() and DOSE::gseDO() (2024-10-01, Tue)

clusterProfiler 4.13.3

• fixed bug in enrichPC() (2024-08-26, Mon)

clusterProfiler 4.13.2

• fixed bug of gson_KEGG() (2024-08-19, Mon)

clusterProfiler 4.13.1

• update functions to access PathwayCommons data (2024-08-11, Sun, gson#9)
• use yulab.utils::yulab_msg() for startup message (2024-07-26, Fri)
• update kegg_category information (7 categories and 572 subcategories) (2024-07-26, Fri)
  • Cellular Processes (36)
  • Drug Development (75)
  • Environmental Information Processing (41)
  • Genetic Information Processing (39)
  • Human Diseases (99)
  • Metabolism (190)
  • Organismal Systems (92)

clusterProfiler 4.12.0

• Bioconductor RELEASE_3_19 (2024-05-15, Wed)

clusterProfiler 4.10.1

• bug fixed in parsing KEGG category (2024-03-07, Thu, #664)
• update citation (#656) and wikipedia data URL (2024-01-10, Wed, #633)

clusterProfiler 4.10.0

• Bioconductor RELEASE_3_18 (2023-10-25, Wed)

clusterProfiler 4.9.5

• fixed R check (2023-10-18, Wed)

clusterProfiler 4.9.4

• use check_installed() to check package dependency (2023-09-08, Fri, #621)
• use yread() in WikiPathway utilities (2023-09-07, Thu)

clusterProfiler 4.9.3

• enrichKEGG() and gseKEGG() now supports organism = 'cpd' to accept KEGG Compound ID (2023-08-31, Thu)
  • gson_cpd() and gson_ko()
• use yulab.utils::yread() to parse file (2023-08-15, Tue)
• supports Pathways Common (2023-08-02, Wed, #613)

clusterProfiler 4.9.2

• append_kegg_category() function to add KEGG pathway category information to KEGG enrichment result and now it is the default behavior of enrichKEGG() and gseKEGG() (2023-07-12, Wed)
• parse KEGG Pathway Category information (2023-07-11, Tue)
• mv parse_gff() to GOSemSim::read.gaf() and re-export (2023-07-10, Mon)
• mv buildGOmap() to `GOSemSim::buildGOmap() and re-export

clusterProfiler 4.9.1

• getPPI() to query PPI network from 'stringdb' (2023-05-15, Mon)
• getTaxID() and getTaxInfo() functions to query taxonomy information (2023-05-14, Sun)

clusterProfiler 4.8.0

• Bioconductor RELEASE_3_17 (2023-05-03, Wed)

clusterProfiler 4.7.2

• change wikiPathways link. (2023-03-10, Fri)
• update get_data_from_KEGG_db() for the KEGG api changes (2023-03-05, Sun)
• removing species info at the end of KEGG pathway names (2023-03-05, Sun)

clusterProfiler 4.7.1

• update according to the KEGG api changes (2023-03-01, Wed)

clusterProfiler 4.6.0

• Bioconductor 3.16 release

clusterProfiler 4.5.3

• GSEA() supports GSONList object (2022-09-21, Wed)
• enricher() supports GSONList object (2022-09-06, Tue)

clusterProfiler 4.5.2

• support passing a GSON object to enricher(USER_DATA) and GSEA(USER_DATA) (2022-8-01, Mon)
• gson_kegg_mapper() allows building a gson object from outputs of KEGG Mapper service (2022-07-29, Fri, #492)
• fix show method for compareClusterResult (2022-06-21, Tue, #473)
• gson_KEGG() download latest KEGG and output a GSON object (2022-06-08, Wed)
• support passing a GSON object to gseKEGG(organism)
• support passing a GSON object to enrichKEGG(organism) (2022-06-06, Mon)

clusterProfiler 4.5.1

• follow KEGG api upgrade that change from http to https (2022-06-06, Mon)
• use 'wininet' to download KEGG data when .Platform$OS.type = "windows" (2022-06-03, Fri)
• mv read.gmt and read.gmt.wp to the 'gson' package and reexport these two functions from 'gson' (2022-04-28, Thu)
• fix compareCluster when fun = enrichPathway(2022-4-28, Thu)

clusterProfiler 4.4.0

• Bioconductor 3.15 release

clusterProfiler 4.3.4

• fix enrichGO , gseGO and groupGO when keyType = 'SYMBOL' && readable=TRUE(2022-4-9, Sat)

clusterProfiler 4.3.3

• parse GAF file to prepare GO annotation data (esp for proteomic study) (2022-03-08, Tue, #397, #418, #421, #442)
• bug fixed in compareCluster() (2022-01-27, Thu, #424)

clusterProfiler 4.3.2

• bug fixed in extract_params() (2022-01-12, Wed, #392, @amcdavid)
• make simplify() works for gseGO() in compareCluster()
• support formula interface for GSEA methods in compareCluster() (2022-01-04, Tue, @altairwei, #416)

clusterProfiler 4.3.1

• compareCluster() supports GSEA algorithm (2021-12-11, Sat)
• update error message of download.KEGG.Path() and download.KEGG.Module()(2021-11-21, Sun)
• update simplify() function to support ont = ALL (2021-10-27, Wed)

clusterProfiler 4.2.0

• Bioconductor 3.14 release

clusterProfiler 4.1.4

• import yulab.utils (2021-08-20, Fri)

clusterProfiler 4.1.3

• Remove Human Gut Microbiome dataset as the functionalities are provided in https://github.com/YuLab-SMU/MicrobiomeProfiler (2021-08-15, Sun)

clusterProfiler 4.1.2

• update citation and DESCRIPTION (2021-08-15, Sun)
• update kegg_species.rda and allow online download using KEGG api (2021-08-14, Sat)

clusterProfiler 4.1.1

• add citation (new paper published on The Innovation) (2021-07-04, Sun)

clusterProfiler 4.0.0

• Bioconductor 3.13 release

clusterProfiler 3.99.1

• Add new data set, DE_GSE8057, which contains DE genes obtained from GSE8057 (2020-03-08, Mon)

clusterProfiler 3.99.0

• Add KEGG enrichment analysis of Human Gut Microbiome data (2021-02-20, Sat)

clusterProfiler 3.19.1

• setting default timeout to 300 for downloads (2021-02-05, Fri)
• fixed download method setting
• capable of setting KEGG download method via options(clusterProfiler.download.method = METHOD) (2020-12-31, Thu)

clusterProfiler 3.18.0

• Bioconductor 3.12 release (2020-10-28, Wed)

clusterProfiler 3.17.5

• update [[.compareClusterResult (2020-10-14, Wed)

clusterProfiler 3.17.3

• internal suports of enrichment analyses using WikiPathways (2020-09-09, Wed)
  • enrichWP for ORA analysis
  • gseWP for GSEA analysis
  • get_wp_organisms for listing supported organisms
  • read.gmt.wp for parsing gmt file downloaded from wikiPathways

clusterProfiler 3.17.2

• use libcurl if capable (2020-09-08, Tue)
  • https://github.com/YuLab-SMU/clusterProfiler/pull/290

clusterProfiler 3.17.1

• bug fixed of extract_params (2020-08-18, Tue)
  • https://github.com/YuLab-SMU/clusterProfiler/issues/282

clusterProfiler 3.16.0

• Bioconductor 3.11 release

clusterProfiler 3.15.3

• incorporate clusterProfiler.dplyr (2020-03-12, Thu)
  • arrange, filter, group_by, mutate, rename, select, slice and summarize

clusterProfiler 3.15.2

• remove Suggests of KEGG.db as it will be deprecated in Bioconductor 3.11 (2020-01-14, Tue)
• optimize enrichGO to use less memory (2019-12-13, Fri)
• re-implement read.gmt without using GSEABase, and my own version is much more fasta :)

clusterProfiler 3.15.1

• e.g. user can pass fun=enrichGO to compareCluster without quoting enrichGO (2019-12-02, Mon)
• add keytype and readable info in compareCluster output
• mv compareClusterResult class defintion to DOSE (2019-11-02, Sat)
• mv fortify, barplot and dotplot for compareClusterResult to enrichplot.

clusterProfiler 3.14.0

• Bioconductor 3.10 release

clusterProfiler 3.12.0

• Bioconductor 3.9 release

clusterProfiler 3.11.1

• asis parameter in [.compareClusterResult (2018-12-24, Mon)
  • https://github.com/GuangchuangYu/enrichplot/issues/17

clusterProfiler 3.10.0

• Bioconductor 3.8 release

clusterProfiler 3.9.2

• re-export DOSE::gsfilter and DOSE::setReadable (2018-05-25, Fri)

clusterProfiler 3.9.1

• change color scheme of dotplot of compareClusterResult back to red->purple (2018-05-17, Thu)
  • https://support.bioconductor.org/p/108996/

clusterProfiler 3.8.0

• Bioconductor 3.7 release

clusterProfiler 3.7.1

• uniprot_get function (2018-01-30, Tue)
• import enrichplot (2018-01-29, Mon)