Gene set interpretation is a fundamental task in functional genomics, where researchers must derive biological insights from lists of genes identified in high-throughput experiments. Current approaches utilise statistical enrichment methods that query predefined functional databases, such as Gene Ontology [1] and KEGG pathways [2], to identify overrepresented biological processes [3,4]. Although powerful, current methods yield fragmented outputs, such as lists of enriched terms from various ontologies, leaving researchers to manually integrate these results to achieve functional insights, a process that is both inefficient and error-prone.

Exploratory gene set analysis has become increasingly challenging as the volume of annotated datasets grows. Researchers compare their findings not only with gene ontology term enrichments but also with signatures from knockdown experiments and diverse resources such as LINCS (Library of Integrated Network-based Cellular Signatures) [5] and the STRING database (STRING-DB) [6]. The main issue is that enrichment analysis tests for the over-representation of genes associated with specific terms. When gene lists overlap, the same genes often appear under many different functional terms. As a result, enrichment analysis can return many distinct-sounding terms that all point to the same underlying biology, making interpretation more difficult.

However, the challenge extends beyond simply removing duplicate information. Simplistic filtering of overlapping gene sets would obscure important biological relationships that only become apparent when analysing genes across the diverse resources mentioned above. These relationships often represent biological processes that bridge multiple databases and reveal insights not captured by any single resource. Consequently, manually curating enrichment outputs to identify both redundancies and meaningful biological patterns is not only error-prone and prohibitively time-consuming but also risks overlooking crucial biological connections.

We hypothesise that recently introduced topic modelling techniques can address this problem. By analysing how terms co-occur across texts, topic modelling reveals underlying themes without requiring predefined categories. These unsupervised statistical methods include Latent Dirichlet Allocation (LDA) [7] and Non-negative Matrix Factorization (NMF) [8], which identify recurring patterns in document collections. The resulting latent themes represent collections of related terms that frequently appear together and may correspond to biological processes, pathways, or functional modules not explicitly defined in current annotation databases.

Recent advances in large language models [9] (LLMs) create powerful new opportunities for gene set interpretation. LLMs have demonstrated remarkable capabilities in contextual understanding and natural language generation, potentially enabling automated synthesis of distributed biological knowledge. Several studies have explored this direction. Hu et al. [10] evaluated LLMs as “compressed databases” of biological literature, querying internal model weights to infer gene set function with external search used for post-hoc verification. GeneAgent [11] generates functional hypotheses from internal knowledge before fact-checking extracted claims against databases. llm2geneset [12] uses internal knowledge to dynamically construct gene set categories, with statistical testing applied afterwards.

Here we present GeneInsight, a tool that integrates LLMs with topic modelling to automate gene set interpretation. Our tool aggregates gene-specific annotations from the STRING database, applies topic modelling to identify coherent biological themes, and employs LLM-based summarisation to generate contextual interpretations of these themes. While existing LLM-based tools query the model’s internal knowledge to infer gene function, GeneInsight takes a different approach. We retrieve annotations directly from STRING and combine topic modelling with LLM summarisation to distil these into coherent biological themes, grounding interpretations in expert-curated database content.

GeneInsight uses a two-stage approach to extract and organise biological information from gene sets (Fig 1a). In the biological theme generation stage, the system collects functional annotations from the STRING database for each input gene, creating a collection of gene-specific descriptions. This textual corpus is subjected to cluster-based topic modelling, which groups similar annotations into clusters (topics) and identifies key terms for each cluster. An LLM then converts representative annotations from each cluster into interpretable biological themes. Themes are then prioritised using an overlap-ratio threshold combined with an empirical p-value calibrated against a size-stratified random-query null (S1 File, Materials and Methods).

The second summarisation stage begins with another round of cluster-based topic modelling to identify key themes. This approach refines these enriched themes by measuring how consistently they appear as cluster representatives across multiple runs of topic modelling. The software then extracts the final summary by selecting themes to include based on user-defined length preferences. A large language model creates a hierarchical summary where major biological themes appear as main headings with related subheadings grouped beneath them. The final interactive HTML report (Fig 1b) links theme descriptions to their corresponding gene annotations. This integration enables researchers to easily navigate between overarching biological processes and their specific components. Moreover, every theme is directly tied to the original STRING-derived annotation outputs, allowing users to trace each summary theme back to its supporting gene-level descriptions and source annotations.

We first characterised GeneInsight’s theme-prioritisation stage by comparing its retained themes with the STRING database functional enrichment application programming interface (API). Using identical underlying gene-level information and 1,000 Molecular Signatures Database (MSigDB) [14] gene sets, this comparison directly assessed our tool’s effectiveness in identifying important biological themes. We measured performance through metrics (Fig 1c) that evaluated both the diversity of identified concepts and the degree of overlap between methods.

GeneInsight consistently identified a larger number of enriched gene sets than the STRING database functional enrichment API across all filtering thresholds examined (Figs 2a, S1). While the two methods show strong positive correlation (r = 0.69-0.87), GeneInsight retained 2–3 times more terms than the STRING-DB API at every matched setting examined (S1 Fig), indicating that topic modelling and LLM summarisation capture biological themes not identified by standard term-by-term enrichment (Fig 2b).

To evaluate the conceptual breadth of identified terms, we measured semantic diversity using average pairwise distance between terms. This metric quantifies how conceptually distinct each term is from all others in the set, with higher values indicating coverage of a broader range of biological concepts rather than redundant or closely related processes. GeneInsight-derived terms demonstrated significantly greater semantic diversity compared to STRING-DB terms (Fig 2c), indicating their ability to capture a wider spectrum of biological information.

Next, we assessed the summarisation stage of GeneInsight, which converts prioritised themes into user-defined summaries. To evaluate theme preservation during summarisation at different levels (25, 50, 75, and 100 terms), we used a Top-k semantic similarity metric that focuses on each source term’s strongest matches to measure how well summaries capture essential biological concepts without being diluted by less relevant relationships. This approach identifies the k-nearest semantic neighbours for each term, better handling the imbalance between comprehensive source material and length-constrained summaries. GeneInsight demonstrated robust summarisation of key themes across different final summary counts (Fig 2d), with the stability of these metrics suggesting that our tool effectively identifies core biological concepts regardless of user-supplied summary length constraints.

To provide an orthogonal validation beyond the cosine similarity metric, which measures the directional similarity between text representations, we employed Earth Mover’s Distance [15] (MoverScore) (Fig 2e). Like cosine distance, this complementary approach assesses semantic similarity by measuring the minimum cost required to transform one text into another in the embedding space, capturing different aspects of semantic relationships. All Top-k values demonstrated MoverScores greater than 0.5, indicating that each extracted theme successfully captured substantial semantic content from the original enriched gene sets. The results showed consistently high correlation values (Fig 2f) between MoverScores and cosine similarity scores (0.93 - 0.94) across all theme configurations, confirming the robustness of our semantic similarity assessments.

We also confirmed that summarisation performance remained stable across documents of varying sizes (from <500 to >3000 terms), demonstrating that our tool’s summarisation capability is independent of input size (Fig 2g).

Next, we applied GeneInsight to an RNA-seq dataset from a murine model of mesothelioma treated with immunotherapy, specifically focusing on genes differentially expressed in treatment responders [13]. GeneInsight prioritised biological themes centred around Type I interferon signalling and monocyte-macrophage axis activation (S1 Data) which were not detected through standard Gene Ontology (GO) enrichment analysis. The importance of Type I interferon signalling was subsequently validated through mouse models using antibody-mediated interferon blockade experiments, which confirmed the functional relevance of these pathways in treatment response. The monocyte involvement themes identified by our tool were further validated through single-cell analysis, confirming GeneInsight’s capacity to identify biologically relevant signatures that effectively bridge bulk and single-cell approaches.

We then evaluated GeneInsight on a multi-omics dataset from the DREAM study [16,17], which included mesothelioma patients undergoing chemoimmunotherapy treatment. Using responder-specific genes identified through a NanoString panel and bulk RNA-seq time course data, GeneInsight successfully extracted stem cell-like signatures in T-cells (S2 Data) that were independently confirmed through orthogonal single-cell validation. In the original analysis, these stem-like signatures were only discovered after weeks of analysis involving differential abundance testing, manual inspection of CD8⁺ T cell subclusters, differential expression analysis and marker analysis comparing responders and non-responder populations [17]. GeneInsight streamlined this process by automatically identifying these key biological themes in a single analysis of 30 minutes, demonstrating how it can reduce analytical complexity and accelerate hypothesis generation from multi-omic datasets.

Finally, we evaluated GeneInsight using a published gene set [18] associated with the transcriptional response of neutrophils to Francisella tularensis infection. GeneInsight identified distinct biological themes from differentially expressed genes involved in glucose metabolism (S3 Data), suggesting a metabolic shift that aligns with neutrophil functional alterations during infection. This metabolic reprogramming pattern, particularly involving key glycolytic regulators such as PFKL, was validated in a follow-up study [19] using the same dataset 9 years later. This case demonstrates GeneInsight’s capacity to derive novel biological insights from existing datasets, potentially accelerating discovery timelines from functional genomics data.

GeneInsight is freely available as an open-source Python package under the MIT License at https://github.com/wlchin/geneinsight. The complete source code, documentation, installation instructions, and analysis workflows are accessible through this repository. The S1 File describes the methods associated with these analysis workflows in detail.

Several limitations should be considered when interpreting GeneInsight results. GeneInsight uses empirical p-values, calibrated against a size-stratified random-query null, as ranking scores for prioritising themes. The procedure assumes that the query genes can be treated as a random sample from a user-specified background gene set [20–22]. In practice, many biological inputs depart from this assumption. Differential-expression lists, curated pathways, STRING-derived neighbourhoods, and STRING-DB functional enrichment outputs often carry experimental, annotation, or network-derived structure that the random-query null does not model [6,20]. Gene–gene correlation, uneven annotation across genes, and overlap among functional categories can also affect calibration [20–22]. These are known limitations of over-representation analysis more broadly [23], and the empirical procedure used here calibrates the random-sampling component but not these additional sources of structure. We do not claim calibration against alternative nulls such as expression-matched or network-degree-matched randomisation. The empirical p-values should therefore be interpreted as a calibrated ranking score for prioritising themes, not as significance tests for individual genes or pathways.

GeneInsight also relies on LLM-generated summaries. Although the LLM is constrained to summarise only retrieved STRING-DB terms and the final HTML report links each theme back to its source annotations, hallucination or over-interpretation cannot be eliminated. Users should therefore verify that generated summaries accurately reflect the displayed source annotations and should treat GeneInsight outputs as hypothesis-generating summaries rather than definitive biological conclusions.

Looking ahead, GeneInsight’s algorithm is designed to process text-based annotations regardless of their source. While our current implementation uses the STRING database, the framework can readily incorporate PubMed abstracts, Gene Ontology terms, pathway descriptions from Reactome [24], and unpublished private datasets. This source-agnostic design allows GeneInsight to incorporate any text-based annotation source, providing an automated aggregation layer that synthesises diverse annotations into interpretable biological themes.