Gene lists no longer return just an enrichment table: a conversational agent links statistics, literature and annotation into an interpretation you can question

Enrichment results are no longer only a table of terms sorted by adjusted p-value; they come with a written interpretation tied to the biological context the user supplies, and the user can keep asking follow-up questions to refine it. Tools such as DAVID, Enrichr and clusterProfiler return lists of enriched terms and do not address the biological question being asked; connecting those terms to a specific question still required manually searching databases, literature and gene function information. A feature comparison against those tools plus GeneAgent and BRAD shows Enrich.AI covering enrichment, LLM interpretation, a web app, multiple visualizations, literature search, gene information, pathway libraries and human-in-the-loop support together, while each compared tool covers only part of that set.

The whole workflow is driven by a large language model agent that decides which tools to call and with which parameters, then synthesizes the multi-step results into an interpretation. The ReAct architecture advances the model through a Thought-Action-Observation loop for up to seven hops or until it has gathered the needed information, and tool call parameters are constrained by typed Pydantic input schemas to reduce parameter-level fabrication. The base model is Gemini 2.5 Flash at temperature 0.4, orchestrated by LangGraph; enrichment runs through Enrichr using Fisher's exact test with Benjamini-Hochberg correction, with results filtered by adjusted p-value.

Interpretation reliability is measured with a reviewable evaluation procedure rather than illustrated by examples alone. The evaluation decomposes model interpretations into individual claims, uses Claude Sonnet 4.5 as a judge to assess faithfulness against retrieved evidence, checks hallucination rate against external databases the tool could not access, and examines reproducibility across repeated runs. The evaluation used the 50 MSigDB Hallmark gene sets, each undergoing enrichment analysis and literature search; 65.6% of claims were grounded in retrieved evidence, 32.1% reflected established background biology, 2.3% were unsupported, the hallucination rate was 2.64% (54 of 2,047 database-decidable claims), no fabricated citations appeared across 90 runs, and enrichment results were consistent across replicates with variation mainly in literature search.

The demonstrated scope is: a user submits a gene or protein list with optional tissue, cell type or disease context; the platform runs enrichment over more than 220 Enrichr libraries, searches over 40 million publications in Europe PMC, retrieves gene annotations from MyGene.info, and returns a summary, detailed analysis, result tables, six interactive plots and a reasoning trace. The evaluation covers 50 MSigDB Hallmark gene sets. The platform is freely available as a Streamlit web application with default access, also accepts user-supplied API keys from several model providers, can run locally, and does not collect or share information beyond an active session. The authors state that future development will add more enrichment methods and visualizations.

Enrichment currently runs exclusively through Enrichr, and rank-based approaches such as GSEA are not yet supported. Faithfulness and hallucination rate are judged by Claude Sonnet 4.5, making this a model-judging-model evaluation. Enrichment results were consistent across replicate runs, with variation mainly in the literature search step. The authors also note that the model cannot always make an independent judgment of the biological plausibility of results, which is why control of the reasoning process stays with the user.