Run functional enrichment on a gene list

Hand Claude Code a list of gene symbols (typically the significant hits from a DE analysis or a CRISPR screen); get back an Enrichr-backed enrichment table across GO Biological Process, KEGG, Reactome, and disease libraries, with a short natural-language synthesis of the dominant pathways and citations to the database accessions.

   
Problem class Data analysis
Subject areas Molecular and Cellular Biology, Immunology and Microbiology, Drug Repurposing and Discovery
Evidence level Reported
Complexity One skill or MCP
Availability Fully open
Compute Laptop

Problem

Functional enrichment is the canonical “so what does this gene list mean” step that follows almost every transcriptomics, proteomics, or screen experiment. The mechanics are unglamorous: hit Enrichr (or g:Profiler, or DAVID) with the gene symbols, sort the term tables by adjusted p-value, and translate the top hits into a paragraph a wet-lab collaborator can act on. The cost is the swivel-chairing — exporting the DE table, pasting symbols into a web form, copying tables back into the analysis notebook — and the interpretation pass at the end is where naïve LLM use most often goes wrong: vanilla GPT-4 will name plausible-sounding pathways for almost any gene list, including random ones. Solved looks like: paste a gene-symbol list and the species, get back a single Markdown report with the top enriched terms per library, each cited to its database accession, and a written summary anchored to those tabulated terms — not to the model’s prior.

  1. Install the gget Claude Skill. From the K-Dense marketplace:

    /plugin marketplace add K-Dense-AI/claude-scientific-skills
/plugin install gget@claude-scientific-skills
pip install gget

  2. Hand Claude the gene list and the report contract. A minimal prompt:

    I have 187 upregulated genes from a DE run (human, padj<0.05,
log2FC>1). Use the gget skill to run gget.enrichr on this list
against the following Enrichr libraries:
  - GO_Biological_Process_2023
  - KEGG_2021_Human
  - Reactome_2022
  - MSigDB_Hallmark_2020
  - DisGeNET
For each library, return the top 10 terms by adjusted p-value.
Save each table to results/enrichment/<library>.csv. Then write
results/enrichment/SUMMARY.md with one paragraph per library
citing only terms that actually appear in the saved CSVs (term
name + adjusted p-value + database accession).

Gene list:
<paste symbols, one per line>

    gget enrichr calls the Enrichr REST API directly; no API key is needed for typical interactive use.

  3. Verify before believing. Ask Claude to cross-check every claim in SUMMARY.md against the saved CSVs:

    Re-read results/enrichment/SUMMARY.md. For every pathway or
disease named, confirm it appears in the corresponding CSV with
adjusted p-value < 0.05. Flag any sentence whose claim cannot
be grounded in the tables and rewrite it.

    This explicit verification pass is the recipe’s main hallucination mitigation — it mirrors the design pattern GeneAgent benchmarks (see Evidence).

  4. Run a negative-control gene set. Sample ~200 random protein-coding genes and re-run the same prompt. Any “enrichment” that survives BH correction on the random set is signalling background bias in the library (e.g., long-gene bias in GO) rather than biology. Worth doing once per project, not once per analysis.

  5. For ranked enrichment (GSEA), drop to GSEApy. gget enrichr is over-representation analysis on an unranked hit list. If you have a full ranked DE table and want gene-set enrichment proper (running enrichment score, leading-edge analysis), call GSEApy from the same Claude session — it ships with Enrichr-compatible gene-set libraries via gp.prerank().

  6. Hand off to downstream interpretation. The CSV tables are the audit trail. Pair with the build-target-dossier recipe when an enriched pathway points back to a specific target you want to characterise next.

Why this assembly

Rung 2 of the simplicity ladder. The gget skill wraps gget enrichr so Claude calls a single function with the right library names and gets back a pandas DataFrame; the model never invents a pathway because the API call is what populates the table. Plain Claude Code (rung 1) would either confabulate Enrichr results or write a one-off requests block against the Enrichr endpoint — fine occasionally but error-prone and not reproducible. A toolbelt (rung 3) buys nothing because the enrichment + interpretation flow is single-source. Rung 4 (autonomous systems) is the wrong tier for a step measured in seconds. The verification pass in step 3 is the explicit hallucination-mitigation pattern; it is the same self-grounding loop that GeneAgent (NIH/Nature Methods 2025) benchmarks at 84% claim support against 1,106 gene sets.

Availability

Fully open. The gget skill is OSS in K-Dense-AI/scientific-agent-skills; the underlying pachterlab/gget library is BSD-2-Clause; Enrichr is free for academic and non-profit use. No subscription, no institutional licence. Commercial users should review the Enrichr terms of service — the API is free to call but the underlying libraries have their own licences (some KEGG mirrors are commercial-restricted).

Compute requirements

Laptop. The whole workflow is HTTP requests against Enrichr; a 200-gene list across five libraries returns in 10–30 seconds end-to-end. No GPU. Disk is trivial (the CSV tables are a few hundred KB). Rate limits are generous for interactive use but if you batch over hundreds of gene sets, throttle to ~1 request/second to stay within Enrichr’s etiquette.

Evidence

Reported. The strongest reference for the assembly class is GeneAgent (Wang et al., Nature Methods 22:1677, 2025, DOI:10.1038/s41592-025-02748-6; PMID 40721871), a self-verification LLM agent that queries Enrichr and other curated databases to ground gene-set claims; across 1,106 gene sets it lifted ROUGE-L on MSigDB from 0.239 ± 0.038 (GPT-4 alone) to 0.310 ± 0.047 (GeneAgent), with 84% of 15,848 generated claims supported by database evidence and 92% of self-verification decisions judged correct by two human experts on a 132-claim sample. The recipe here is a smaller-grain composition (one skill, no autonomous loop) but the underlying database-grounding pattern is the same. Complementary anchors: Hu et al., Nature Methods 21:2353, 2024 — “Evaluation of large language models for discovery of gene set function” (DOI:10.1038/s41592-024-02525-x) shows GPT-4 names common gene-set functions with high specificity but only when grounded against a database; Joshi et al., llm2geneset preprint 2024-11 (DOI:10.1101/2024.11.11.621189) shows LLM-generated gene sets can be used as Enrichr-compatible inputs. No peer-reviewed benchmark of “Claude + gget skill + Enrichr” against hand-written gget enrichr code is known — the agent loop adds reproducibility and the verification pass, not new statistics.

Alternatives considered

  • Plain Claude Code with the Enrichr REST API. Works for one-off analyses but the model has to re-derive the right userListId ⇒ enrich flow and the right library shortnames each time, and is less likely to retain the per-library result tables in a reproducible form. Reach for it only when the gget skill isn’t installed and the analysis is throwaway.
  • GSEApy directly. GSEApy covers Enrichr over-representation, ranked GSEA, single-sample GSEA, and Biomart conversions in one Python package. It is the right answer when the analysis is ranked-list GSEA (step 5) or batched across hundreds of gene sets. It is not yet wrapped as a Claude skill in the catalog, so calls fall back to plain Claude + Python.
  • g:Profiler / WebGestalt. Both are stronger than Enrichr on some libraries (g:Profiler has tighter multi-test correction; WebGestalt has better term-grouping). Neither has a Claude skill in the catalog; if you need them, drop to plain Claude + the respective REST APIs. Consider proposing skill wrappers if this becomes routine.
  • GeneAgent directly. ncbi-nlp/GeneAgent is publicly available; reach for it when the deliverable is gene-set naming with verified evidence claims (the use case the paper benchmarks) rather than a tabular enrichment report. It is not yet catalogued as a Claude skill.
  • An autonomous-science system (Biomni). Overkill for a single enrichment step. Biomni includes Enrichr-class tools alongside dozens of others; reach for it only when enrichment is one node in a larger autonomous loop.

See also

Sources

Tried this recipe?

Share feedback — what worked, what didn’t, what you’d change. The form opens with this recipe pre-selected and a link back to this page.