LLM-Assisted Functional Gene Annotation

LLM-Assisted Functional Gene Annotation Calvin Boyce1, Clara Pereira2, Gyeong Dae Kim3, Allegra Petti4* 1,2,3,4Department of Neurosurgery, Mass General Brigham, Boston, Massachusetts, USA. *Corresponding author(s). E-mail(s): ap[email protected]ard.edu; Contributing authors: cbo[email protected]ard.edu; cdp[email protected]ard.edu;[email protected]ard.edu; Abstract Functional gene annotation is a highly manual and subjective process that requires analysis of large amounts of statistical and literature results. Here we present Artificial Intelligence Gene Enrichment (AIGE), a careful automation of the current state of the art process used by bioinformatic experts to determine functions enriched in novel or experimentally derived gene lists. In 1206 test cases, AIGE is able to accurately recover 87% of biological functions, pathways, and cell types, is robust to noise contamination, and accurately assesses its own self-confidence. AIGE reports provide accurate functional annotations and encourage further research in a broad set of contexts. 1 Introduction Functional gene enrichment is a fundamental task in genomic and molecular biology, and an instrumental step in most standard bioinformatic analyses. Extensive work has been done to determine the functions and characteristics of individual genes[1,2], and to compile lists of genes associated with specific biological or molecular functions[3], but elucidating the phenotypic behavior of a novel gene set in the context of a specific biological question requires carefully designed experimentation. Large scale bioinformatic analyses often include dozens or even hundreds of novel gene lists for which individual experimentation to determine functional enrichment is infeasible. In these settings, approximate functional annotations are often obtained by comparing novel gene lists against databases of known gene lists to determine statistical significance of overlap via either Over-Representation Analysis (ORA) or Gene Set Enrichment 1 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint Analysis (GSEA)[4]. ORA uses Fisher’s exact test or a hypergeometric test to compare the overlap between a set of genes and a database of predefined gene lists, and is often paired with Bonferroni or false discovery rate statistical correction to account for multiple comparisons. GSEA determines the over or underrepresentation of a known set of genes in a ranked list of Differentially Expressed Genes (DEGs) using a Kolmogorov-Smirnov-like running sum. Each of these methods determines the likelihood of observing a degree of overlap between the novel gene set and the known gene lists under the assumption that genes are expressed uniformly and independently from one another, a strong assumption that leaves room for statistical innovation. Several tools and methods have been built to perform these statistical comparisons including ToppFun[5], Enrichr[6], clusterProfiler[7], g:Profiler[8], GSEA[4], and fgsea[9], providing enriched terms from Gene Ontology (GO)[3], Kyoto Encyclopedia of Genes and Genomes (KEGG)[10], Molecular Signatures Database (MSigDB)[11], and other curated lists of genes and their functional or pathway involvement. While these results give researchers a good statistical starting point, a proper functional enrichment analysis often includes extensive literature review to verify the functions of subsets of genes in the novel list, and an expert opinion to successfully synthesize the available information. This is a highly manual and subjective process that significantly slows down large-scale research and analysis. With the advent of Large Language Models (LLMs) and their ability to parse textual data at scale, several recent approaches have applied LLMs to bioinformatic tasks such as cell typing[12–16] and functional gene enrichment[17–20]. While the more novel approaches perform fundamental pre-training similar to that of Generative Pre-Trained (GPT) models[13], or integrate literature sources to augment biological knowledge[21], most approaches use flagship foundation models from sources like OpenAI, Anthropic, or DeepSeek out of the box and evaluate their performance reproducing Gene Ontology or other defined gene list names without fine tuning or training (i.e. zero-shot). This strategy takes advantage of the impressive capacity for LLMs to apply domain expertise to novel data in a flexible and scalable format. The current quintessential objection to using LLMs for zero-shot prediction is the inability to examine the reasons behind the token predictions of transformers, the fundamental unit of modern LLMs, leading to a lag in adoption of AI-based tools in biology and many other fields. While Chain of Thought (CoT) reasoning models provide some insight into the underlying processes of LLMs, recent research has cast skepticism on the reliability of reasoning models to represent their thinking process or generalize to truly novel data[22–24]. In an approach more analogous to classical neuroscience, researchers at Anthropic have recently pioneered an approach to examine “meta-features” (groups of neurons that “fire” or activate together) in their neural networks to better understand the process LLMs use to make decisions, but this field of research is still in its infancy[25]. Most recently, Wang et. al. published GeneAgent[26] which uses zero-shot LLM prediction to propose functions that may be enriched in a set of genes, then passes those proposed functions through an agent to support or refute their validity by comparing them against classic statistical tools and literature. This approach successfully minimizes the hallucinations of the LLM labeling the gene lists, but it also restricts the 2 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint capacity of the LLM to generate novel annotations not found in existing databases. The validity of a novel claim made by an LLM can only truly be determined experimentally, which is difficult to scale. Here we propose Artificial Intelligence Gene Enrichment (AIGE), an LLMaugmented automation of the current expert best practice for gene annotation. The model follows the same steps a human scientist would by parsing statistical sources provided by ToppFun, Enrichr, and g:Profiler, as well as gene summaries and literature findings from NCBI’s gene and PubMed databases to produce a holistic report that determines functions enriched in a novel set of genes while providing industry standard citations. This approach only utilizes the LLM for data synthesis, not for claim generation. It does not take advantage of the capacity for LLMs to propose previously unannotated functions, but it also eliminates the capacity for hallucination entirely. The AIGE model is publicly available in python and can be accessed here: https://github.com/calvinrboyce/ai gene enrichment. 2 Results 2.1 Model Architecture The AIGE model takes a list of genes as input as well as (optionally) a set of background genes to enrich against, a list of search terms for the PubMed literature search, and a brief context to provide the LLM to assist in report generation. It first sends the gene list and background set (if provided) to ToppFun[5], Enrichr[6], and g:Profiler[8] to obtain Over-Representation Analysis (ORA) results from databases including the Gene Ontology[3], KEGG[10], Reactome[27], WikiPathways[28], protein-protein interactions, and several cell atlases. Each tool uses slightly different statistical corrections to the classic hypergeometric test, with various strengths and weaknesses, so the AIGE model integrates p-values from all three tools to create an ensembled set of ORA results across each database. It then queries NCBI’s PubMed database to find recent articles that mention subsets of the provided genes, as well as optionally provided search terms (e.g. a paper studying glioblastoma that mentions 4 of the provided genes in an identified process). If the gene list is differentially ranked, or if the list is shorter than 10 genes, the top genes (up to 10) are queried against NCBI’s gene database to obtain summaries for each gene. ORA results from 12 databases (as obtained by the three statistical tools), literature results from PubMed (article titles, abstracts, and paragraphs), as well as gene summaries are then provided to an LLM (OpenAI’s gpt4.1-mini-2025-04-14)[29] to identify functional themes. Highly statistically significant ORA results are aligned with related literature findings to create a list of functional themes enriched in the input list of genes, each with a description of what the theme does biologically and why it was identified. The LLM also generates a confidence score between 0 and 1 reflecting the coherence and significance of each theme. The AIGE model then produces a high level summary of all the functions found to be enriched in the input. 3 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint Fig. 1 a, The AIGE model performs functional gene enrichment by sending the input list of genes to three statistical tools, ToppFun, Enrichr, and g:Profiler, as well as NCBI to fetch literature results, then sends all data to an LLM to identify themes in the results. b, The structure of the prompt includes system instructions, the main prompt, optional context, the list of genes, and the enrichment results. 2.2 Model Evaluation In the absence of experimental ground truth, we sought to validate the AIGE model by evaluating its ability to recreate labels for known lists of genes in several databases. Specifically, we evaluated its ability to recreate biological functions using the Biological Process branch of the Gene Ontology[3] and the Hallmark dataset from MSigDB[11], pathways using the Reactome dataset[27], and cell types using marker genes obtained from PanglaoDB[30]. We then evaluated the generated confidence scores to understand how the AIGE model assesses its own confidence by running the model on gene sets contaminated with randomly sampled genes (as in Hu et. al.[17]). 2.2.1 Biological Process Identification To validate the model’s ability to identify biological processes, we tested it on gene lists obtained from the Gene Ontology Biological Process database as well as MSigDB’s Hallmark dataset. Each gene list was provided to the model, which generated a list of themes found to be enriched in the gene set. Each of those themes was then compared to the database’s name for the gene list and evaluated semantically to determine how similar the proposed function was to the true function (following the work of Hu et. al.). We generated a percentile score for each term representing the similarity of the proposed name to the true name, as compared with all biological function names in the database. Thus, a percentile score of 95 indicates that the function proposed by the AIGE model is more similar to the database’s labelled function than it is to 95% of all biological functions represented in the database. Because the model has access to the GO database via its ORA tools, for each GO term we first blinded the model to the specific input term, then ran the analysis with masking so it couldn’t cite the exact input term in its results. Due to the hierarchical 4 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint nature of the GO database, the model still had access to parent and child terms in the GO graph, resulting in slight data leakage, so performance was expected to be high. The model was not provided search terms, background genes, or context during testing. When evaluated on 500 randomly sampled gene sets from GO with lists of genes ranging from 5 to 500 genes in length, the AIGE model was able to recover biological processes above the 95th percentile for 93% of terms (466 terms), with a median percentile score of 99.94. Fig. 2 a, The distribution of percentile scores for names generated by the AIGE model. 500 Gene Ontology terms were tested. b, The relationship between the length of the input gene list and the resulting percentile score. c, The distributions of semantic similarities between the proposed name and each term in the database for 10 randomly sampled test cases. The correct term name is indicated with a red dot. Terms displayed are: 1) Negative regulation of signal transduction by p53 class mediator, 2) Negative regulation of axon regeneration, 3) Maintenance of lens transparency, 4) Negative regulation of smoothened signaling pathway, 5) Macrophage activation, 6) Positive regulation of activated T cell proliferation, 7) Glycoprotein biosynthetic process, 8) Tricuspid valve morphogenesis, 9) Negative regulation of fatty acid biosynthetic process, 10) Sperm flagellum assembly. d, The percentage of database terms recovered at a given similarity percentile threshold. The number of terms recovered above 95% similarity is indicated in red. When evaluating on the MSigDB database, the model was completely blinded to MSigDB as a data source to eliminate data leakage. Because the MSigDB hallmark dataset only contains 50 terms, the model’s proposed names were semantically compared to the 50 MSigDB terms as well as the 30,786 GO Biological Process term names for a more representative semantic distribution. Of the 50 gene sets, the AIGE model was able to recover biological processes above the 95th percentile for 70% of terms (35 terms), with a median percentile score of 99.85. 5 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint Fig. 3 a, The distribution of percentile scores for names generated by the AIGE model. 50 MSigDB terms were tested. b, The relationship between the length of the input gene list and the resulting percentile score. c, The distributions of semantic similarities between the proposed name and each term in the database for 10 randomly sampled test cases. The correct term name is indicated with a red dot. Terms displayed are: 1) UV response up, 2) Bile acid metabolism, 3) Estrogen response early, 4) KRAS signaling up, 5) UV response dn, 6) Mitotic spindle, 7) Reactive oxygen species pathway, 8) Heme metabolism, 9) Notch signaling, 10) Peroxisome. d, The percentage of database terms recovered at a given similarity percentile threshold. The number of terms recovered above 95% similarity is indicated in red. 2.2.2 Pathway Identification Using the same validation scheme as above, we similarly evaluated the model on its ability to reproduce pathways from the Reactome database. Of the 2785 pathways in the database, we randomly sampled 480 terms to test. We blinded the model to the Reactome database as a whole (though it still had access to KEGG and WikiPathways), and computed percentile scores against the full list of pathways. The model was not provided search terms, background genes, or context during testing. The AIGE model was able to recover pathways above the 95th percentile for 86% of terms (413 terms) with a median percentile score of 99.75. In a successful case, the Reactome term “Zygotic genome activation (ZGA)” was mapped to the AIGE model’s proposed theme of “Embryonic Development and Zygotic Genome Activation (ZGA)”. In an unsuccessful case, the Reactome term “Post-translational protein phosphorylation” was mapped to “Signaling and Receptor Activity”. 6 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint Fig. 4 a, The distribution of percentile scores for names generated by the AIGE model. 480 Reactome terms were tested. The median percentile score is marked in red. b, The relationship between the length of the input gene list and the resulting percentile score. c, The distributions of semantic similarities between the proposed name and each term in the database for 10 randomly sampled test cases. The correct term name is indicated with a red dot. Terms displayed are: 1) RND1 GTPase cycle, 2) Nuclear events stimulated by ALK signaling in cancer, 3) Mtb iron assimilation by chelation, 4) Other semaphorin interactions, 5) Signaling by high-kinase activity BRAF mutants, 6) Dual incision in TC-NER, 7) Neurotransmitter release cycle, 8) Positive Regulation of CDH1 Gene Transcription, 9) Removal of aminoterminal propeptides from gamma-carboxylated proteins, 10) Fatty Acids bound to GPR40 (FFAR1) regulate insulin secretion. d, The percentage of database terms recovered at a given similarity percentile threshold. The number of terms recovered above 95% similarity is indicated in red. 2.2.3 Cell Typing Though the AIGE model was not designed to perform cell typing, it does have access to multiple cell atlases, and we evaluated its performance identifying 178 cell types using marker genes in the PanglaoDB dataset. Since the model does not have access to the PanglaoDB atlas, no masking was required during testing. The model was not provided search terms or background genes during testing, but the context phrase “Include a theme for likely cell type” was provided. Of the 178 cell types tested, the AIGE model was able to recover cell types above the 95th percentile for 77% of terms (137 terms) with a median percentile score of 99.44. 7 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint Fig. 5 a, The distribution of percentile scores for cell types generated by the AIGE model. 178 cell types were tested. The median percentile score is marked in red. b, The relationship between the number of provided marker genes and the resulting percentile score. c, The distributions of semantic similarities between the proposed cell type and each cell type in the database for 10 randomly sampled test cases. The correct cell type is indicated with a red dot. Cell types displayed are: 1) NK cells, 2) Hepatic stellate cells, 3) Ependymal cells, 4) Serotonergic neurons, 5) Pericytes, 6) Parathyroid chief cells, 7) Bergmann glia, 8) Dendritic cells, 9) Microfold cells, 10) Radial glia cells. d, The percentage of cell types recovered at a given similarity percentile threshold. The number of cell types recovered above 95% similarity is indicated in red. 2.2.4 Error Characterization For each test database we examined the 10 worst performing terms to gain insight into the types of mistakes made by the model and identified several mechanisms for failure. First, the semantic embedding model used to select the most likely theme from the output was at times prone to error. For example, for the GO term “Maintenance of lens transparency” (GO:0036438), the AIGE model identified several functional themes including “CHMP4B and Cellular Membrane Dynamics and Cataractogenesis”, but the semantic comparison instead selected the “Tissue and Anatomical Structure Homeostasis” theme as the most semantically similar to the GO term name. The comparison may have weighed the semantic similarity of “maintenance” and “homeostasis” more highly than the biological similarity of “lens transparency” and “cataractogenesis”, a problem which would readily be avoided when a human reviews the output of the model. Second, small annotations sometimes led the model to pick up on more general topics than labeled in the database. For example, The GO term “Positive regulation of granulocyte macrophage colony-stimulating factor production” contains only 8 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint 15 genes, for which the AIGE model proposed “Cytokine Production and Immune Response Regulation”. The model correctly identified the involvement of these genes with cytokine production but was unable to correctly identify GMCSF. These 15 genes are likely involved in the regulation of many cytokines, thus the proposed name was not incorrect but lacked specificity. Third, in some cases the heterogeneous nature of datasets like the MSigDB Hallmark dataset led to vague or improper annotations. For example, the MSigDB term “MYC targets v1” does not accurately represent a biological function or pathway, it simply lists genes that are regulated by MYC. So the AIGE model’s proposed “Regulation of Cell Cycle and DNA Replication Initiation” theme may indeed reflect the function of the genes listed, even if it is not similar to the term name. Fourth, database labels were at times ambiguous or inaccurate as in the case of the PanglaoDB cell type “Undefined placental cells”. In these cases, the model’s proposed “Imprinted Genes and Thyroid Hormone Metabolism - DIO3 and H19 Region” theme may be a more accurate annotation of the gene list than the assigned label. 2.2.5 Confidence Evaluation To contextualize the confidence scores provided for each theme in the model’s output, we examined the distribution of confidence scores produced across all test cases in all four databases (GO, MSigDB, Reactome, PanglaoDB), and determined distinct confidence modes (Figure 6a). Based on this distribution, we binned confidence scores above 0.93 as “High Confidence,” scores between 0.87 and 0.93 as “Medium Confidence,” and scores lower than 0.87 as “Low Confidence”. We then sampled 200 of our 480 Reactome test cases and generated a 50/50 mix test case (with half real genes from the database term and half randomly sampled genes from the superset of genes in the database), and a truly random test case (a random sample of genes from the database with the same length as the original gene list) for each test case. The AIGE model proposed high or medium confidence in 87% of pure test cases, 82% of mixed test cases, and 24% of random test cases, exhibiting only a mild reduction in confidence for the mixed cases, but a satisfactorily large reduction for random cases. However, analyzing the model’s accuracy for those same test cases revealed that the AIGE model recovered terms in the 95th percentile or higher for 86% of pure test cases, 84% of mixed test cases, and 28% of random test cases, indicating that while confidence didn’t decrease much for the mixed cases, neither did performance. At a 50% Signal to Noise Ratio (SNR), the AIGE model’s performance only dropped by 2%. The percentage of test cases labeled as high or medium confidence (87%, 82%, 24%), tracks closely with the accuracy of the model (86%, 84%, 28%), indicating that the confidence assessment is correctly evaluating performance. 9 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint M.L., Kwitek, A.E., Laulederkind, S.J.F., Tutaj, M.A., Vedi, M., Wang, S.- J., D’Eustachio, P., Aimo, L., Axelsen, K., Bridge, A., Hyka-Nouspikel, N., Morgat, A., Aleksander, S.A., Cherry, J.M., Engel, S.R., Karra, K., Miyasato, S.R., Nash, R.S., Skrzypek, M.S., Weng, S., Wong, E.D., Bakker, E., Berardini, T.Z., Reiser, L., Auchincloss, A., Axelsen, K., Argoud-Puy, G., Blatter, M.-C., Boutet, E., Breuza, L., Bridge, A., Casals-Casas, C., Coudert, E., Estreicher, A., Livia Famiglietti, M., Feuermann, M., Gos, A., Gruaz-Gumowski, N., Hulo, C., Hyka-Nouspikel, N., Jungo, F., Le Mercier, P., Lieberherr, D., Masson, P., Morgat, A., Pedruzzi, I., Pourcel, L., Poux, S., Rivoire, C., Sundaram, S., Bateman, A., Bowler-Barnett, E., Bye-A-Jee, H., Denny, P., Ignatchenko, A., Ishtiaq, R., Lock, A., Lussi, Y., Magrane, M., Martin, M.J., Orchard, S., Raposo, P., Speretta, E., Tyagi, N., Warner, K., Zaru, R., Diehl, A.D., Lee, R., Chan, J., Diamantakis, S., Raciti, D., Zarowiecki, M., Fisher, M., James-Zorn, C., Ponferrada, V., Zorn, A., Ramachandran, S., Ruzicka, L., Westerfield, M.: The gene ontology knowledgebase in 2023. Genetics 224(1), 031 (2023) https: //doi.org/10.1093/genetics/iyad031 https://academic.oup.com/genetics/articlepdf/224/1/iyad031/59147104/iyad031.pdf [4] Subramanian, A., Tamayo, P., Mootha, V.K., Mukherjee, S., Ebert, B.L., Gillette, M.A., Paulovich, A., Pomeroy, S.L., Golub, T.R., Lander, E.S., Mesirov, J.P.: Gene set enrichment analysis: A knowledge-based approach for interpreting genome-wide expression profiles. Proceedings of the National Academy of Sciences 102(43), 15545–15550 (2005) https://doi.org/10.1073/pnas.0506580102 https://www.pnas.org/doi/pdf/10.1073/pnas.0506580102 [5] Chen, J., Bardes, E.E., Aronow, B.J., Jegga, A.G.: Toppgene suite for gene list enrichment analysis and candidate gene prioritization. Nucleic Acids Research 37, 305–311 (2009) https: //doi.org/10.1093/nar/gkp427 https://academic.oup.com/nar/articlepdf/37/suppl 2/W305/3956724/gkp427.pdf [6] Chen, E.Y., Tan, C.M., Kou, Y., Duan, Q., Wang, Z., Meirelles, G.V., Clark, N.R., Ma’ayan, A.: Enrichr: interactive and collaborative html5 gene list enrichment analysis tool. BMC Bioinformatics 14 (2013) https://doi.org/10.1186/ 1471-2105-14-128 [7] Yu, G., Wang, L.-G., Han, Y., He, Q.-Y.: clusterprofiler: an r package for comparing biological themes among gene clusters. OMICS: A Journal of Integrative Biology 16(5), 284–287 (2012) https://doi.org/10.1089/omi.2011.0118 https://doi.org/10.1089/omi.2011.0118. PMID: 22455463 [8] Raudvere, U., Kolberg, L., Kuzmin, I., Arak, T., Adler, P., Peterson, H., Vilo, J.: g:profiler: a web server for functional enrichment analysis and conversions of gene lists (2019 update). Nucleic Acids Research 47(W1), 191–198 (2019) https://doi.org/10.1093/nar/gkz369 https://academic.oup.com/nar/article-pdf/47/W1/W191/28879887/gkz369.pdf 16 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint [9] Sergushichev, A.A.: An algorithm for fast preranked gene set enrichment analysis using cumulative statistic calculation. bioRxiv (2016) https://doi.org/10.1101/ 060012 https://www.biorxiv.org/content/early/2016/06/20/060012.full.pdf [10] Kanehisa, M., Goto, S.: Kegg: Kyoto encyclopedia of genes and genomes. Nucleic Acids Research 28(1), 27–30 (2000) https://doi.org/10.1093/nar/28.1.27 https://academic.oup.com/nar/article-pdf/28/1/27/9895154/280027.pdf [11] Liberzon, A., Birger, C., Thorvaldsd´ottir, H., Ghandi, M., Mesirov, J., Tamayo, P.: The molecular signatures database hallmark gene set collection. Cell Systems 1(2015) https://doi.org/10.1016/j.cels.2015.12.004 [12] Hou, W., Ji, Z.: Assessing gpt-4 for cell type annotation in single-cell rna-seq analysis. Nature Methods 21 (2024) https://doi.org/10.1038/s41592-024-02235-4 [13] Rizvi, S.A., Levine, D., Patel, A., Zhang, S., Wang, E., He, S., Zhang, D., Tang, C., Lyu, Z., Darji, R., Li, C., Sun, E., Jeong, D., Zhao, L., Kwan, J., Braun, D., Hafler, B., Ishizuka, J., Dhodapkar, R.M., Chung, H., Azizi, S., Perozzi, B., Dijk, D.: Scaling large language models for next-generation single-cell analysis. bioRxiv (2025) https://doi.org/10.1101/2025.04.14.648850 https://www.biorxiv.org/content/early/2025/04/17/2025.04.14.648850.full.pdf [14] Yang, C., Zhang, X., Chen, J.: Large language model consensus substantially improves the cell type annotation accuracy for scrnaseq data. bioRxiv (2025) https://doi.org/10.1101/2025.04.10.647852 https://www.biorxiv.org/content/early/2025/04/17/2025.04.10.647852.full.pdf [15] Liu, J., Xu, S., Zhang, L., Zhang, J.: Single-Cell Omics Arena: A Benchmark Study for Large Language Models on Cell Type Annotation Using Single-Cell Data (2024). https://arxiv.org/abs/2412.02915 [16] Wang, X., Tan, R., Wang, B., Cristea, S.: Biological reasoning with reinforcement learning through natural language enables generalizable zero-shot cell type annotations. bioRxiv (2025) https://doi.org/10.1101/2025.06.17.659642 https://www.biorxiv.org/content/early/2025/06/24/2025.06.17.659642.full.pdf [17] Hu, M., Alkhairy, S., Lee, I., Pillich, R.T., Fong, D., Smith, K., Bachelder, R., Ideker, T., Pratt, D.: Evaluation of large language models for discovery of gene set function. Nature Methods 22 (2025) https://doi.org/10.1038/ s41592-024-02525-x [18] Khan, T., Yurieva, M., Kabeer, B.S.A., Toufiq, M., Rinchai, D., Palucka, K., Chaussabel, D.: Deep functional profiling of gene sets using large language models: A blueprint for tailored, context-aware functional annotation. bioRxiv (2024) https://doi.org/10.1101/2024.12.12.628275 https://www.biorxiv.org/content/early/2024/12/17/2024.12.12.628275.full.pdf 17 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint [19] Joachimiak, M.P., Caufield, J.H., Harris, N.L., Kim, H., Mungall, C.J.: Gene Set Summarization using Large Language Models (2024). https://arxiv.org/abs/ 2305.13338 [20] Zhu, J., Wang, R.Y., Wang, X., Azevedo, R., Moreno, A., Kuhn, J.A., Khan, Z.: Enhancing gene set overrepresentation analysis with large language models. Bioinformatics Advances 5(1), 054 (2025) https://doi.org/10. 1093/bioadv/vbaf054 https://academic.oup.com/bioinformaticsadvances/articlepdf/5/1/vbaf054/62403000/vbaf054.pdf [21] Jin, Q., Yang, Y., Chen, Q., Lu, Z.: Genegpt: augmenting large language models with domain tools for improved access to biomedical information. Bioinformatics 40(2), 075 (2024) https://doi.org/10.1093/ bioinformatics/btae075 https://academic.oup.com/bioinformatics/articlepdf/40/2/btae075/56803364/btae075.pdf [22] Chen, Y., Benton, J., Radhakrishnan, A., Uesato, J., Denison, C., Schulman, J., Somani, A., Hase, P., Wagner, M., Roger, F., Mikulik, V., Bowman, S.R., Leike, J., Kaplan, J., Perez, E.: Reasoning Models Don’t Always Say What They Think (2025). https://arxiv.org/abs/2505.05410 [23] Shojaee, P., Mirzadeh, I., Alizadeh, K., Horton, M., Bengio, S., Farajtabar, M.: The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (2025). https://arxiv.org/ abs/2506.06941 [24] Lawsen, A.: Comment on The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (2025). https://arxiv.org/abs/2506.09250 [25] Lindsey, J., Gurnee, W., Ameisen, E., Chen, B., Pearce, A., Turner, N.L., Citro, C., Abrahams, D., Carter, S., Hosmer, B., Marcus, J., Sklar, M., Templeton, A., Bricken, T., McDougall, C., Cunningham, H., Henighan, T., Jermyn, A., Jones, A., Persic, A., Qi, Z., Thompson, T.B., Zimmerman, S., Rivoire, K., Conerly, T., Olah, C., Batson, J.: On the biology of a large language model. Transformer Circuits Thread (2025) [26] Wang, Z., Jin, Q., Wei, C.-H., Tian, S., Lai, P.-T., Zhu, Q., Day, C.-P., Ross, C., Leaman, R., Lu, Z.: Geneagent: self-verification language agent for gene-set analysis using domain databases. Nature Methods 22 (2025) https://doi.org/10. 1038/s41592-025-02748-6 [27] Milacic, M., Beavers, D., Conley, P., Gong, C., Gillespie, M., Griss, J., Haw, R., Jassal, B., Matthews, L., May, B., Petryszak, R., Ragueneau, E., Rothfels, K., Sevilla, C., Shamovsky, V., Stephan, R., Tiwari, K., Varusai, T., Weiser, J., Wright, A., Wu, G., Stein, L., Hermjakob, H., D’Eustachio, P.: The reactome pathway knowledgebase 2024. Nucleic Acids Research 52(D1), 672–678 (2023) 18 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint https://doi.org/10.1093/nar/gkad1025 https://academic.oup.com/nar/articlepdf/52/D1/D672/55039828/gkad1025.pdf [28] Agrawal, A., Balcı, H., Hanspers, K., Coort, S.L., Martens, M., Slenter, D.N., Ehrhart, F., Digles, D., Waagmeester, A., Wassink, I., AbbassiDaloii, T., Lopes, E.N., Iyer, A., Acosta, J., Willighagen, L.G., Nishida, K., Riutta, A., Basaric, H., Evelo, C., Willighagen, E.L., Kutmon, M., Pico, A.: Wikipathways 2024: next generation pathway database. Nucleic Acids Research 52(D1), 679–689 (2023) https://doi.org/10.1093/nar/gkad960 https://academic.oup.com/nar/article-pdf/52/D1/D679/55040703/gkad960.pdf [29] OpenAI: ChatGPT (gpt-4.1-mini-2025-04-14). Large language model (2025). https://platform.openai.com/docs/models/gpt-4.1-mini [30] Franz´en, O., Gan, L.-M., Bj¨orkegren, J.L.M.: Panglaodb: a web server for exploration of mouse and human single-cell rna sequencing data. Database 2019, 046 (2019) https://doi.org/10. 1093/database/baz046 https://academic.oup.com/database/articlepdf/doi/10.1093/database/baz046/28277084/baz046.pdf [31] Chen, J.: Gene Lookup API. https://toppgene.cchmc.org/API/lookup [32] Chen, J.: Functional Enrichment API. https://toppgene.cchmc.org/API/enrich [33] Ma’ayan, A.: Enrichr API Documentation. https://maayanlab.cloud/Enrichr/ help#api [34] Raudvere, U.: g:Profiler Package. https://pypi.org/project/gprofiler-official/ [35] Sayers, E.: The E-utilities In-Depth: Parameters, Syntax and More (2009). https: //www.ncbi.nlm.nih.gov/books/NBK25499/#chapter4.ESearch [36] Sayers, E.: The E-utilities In-Depth: Parameters, Syntax and More (2009). https: //www.ncbi.nlm.nih.gov/books/NBK25499/#chapter4.EFetch [37] Contributors, T.B.: Biopython Package. https://biopython.org/docs/1.76/api/ Bio.Entrez.html [38] OpenAI: Responses API. https://platform.openai.com/docs/api-reference/ responses/create?lang=python [39] OpenAI: Structured Output. https://platform.openai.com/docs/guides/ structured-outputs?api-mode=responses [40] Ontology, G.: Gene Ontology Files. https://geneontology.org/docs/ download-ontology/ [41] Ontology, G.: Gene Annotations. https://current.geneontology.org/products/ 19 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint pages/downloads.html [42] MSigDB: Hallmarks Collection. https://www.gsea-msigdb.org/gsea/msigdb/ human/collections.jsp [43] Pathways, R.: Reactome Downloads. https://reactome.org/download-data [44] PanglaoDB: Cell Type Markers. https://panglaodb.se/markers.html [45] Liu, F., Shareghi, E., Meng, Z., Basaldella, M., Collier, N.: Self-Alignment Pretraining for Biomedical Entity Representations (2021). https://arxiv.org/abs/ 2010.11784 [46] Huggingface: Transformers Package. https://huggingface.co/docs/transformers/ en/index [47] Cursor: Cursor - The AI Code Editor (2024). https://www.cursor.com 20 .CC-BY-NC 4.0 International licenseperpetuity. It is made available under a preprint (which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in The copyright holder for thisthis version posted September 2, 2025. ; https://doi.org/10.1101/2025.08.06.668957doi: bioRxiv preprint