Application of Large Language Models for Annotating Genes into Reactome Pathways

Wu, G.; Matthews, L.; Boyer, N.; Milacic, M.; Beavers, D.; Li, N. T.; May, B.; Rothfels, K.; Shamovsky, V.; Stephan, R.; Gillespie, M.; Hermjakob, H.; DEustachio, P.; Stein, L. D.

doi:10.64898/2025.12.20.695723

bioRxiv · 10.64898/2025.12.20.695723

Application of Large Language Models for Annotating Genes into Reactome Pathways

Wu, G.·Matthews, L.·Boyer, N.·Milacic, M.·Beavers, D.·Li, N. T.·May, B.·Rothfels, K.·Shamovsky, V.·Stephan, R.·Gillespie, M.·Hermjakob, H.·DEustachio, P.·Stein, L. D.

View original source ↗Export metadata ↓

Abstract

Reactome is the most comprehensive, open source, open access biological pathway knowledgebase, widely used in the research community. To ensure the highest quality of its content, human pathway data in Reactome is manually curated. However, manual curation is labor-intensive, time-consuming, and increasingly difficult to keep up with the ever-growing biomedical literature. Large language model (LLM)-driven artificial intelligence (AI) technologies are transforming many fields, including bioinformatics resource development. Applying LLM/AI technologies in Reactome may offer a powerful way to scale curation and consolidate pathway-related data into a single resource. This manuscript describes the first stage of our attempt to adopt LLM/AI technologies for Reactome manual curation. We developed an LLM workflow that can assist curators in adding new genes to existing pathways and refining the functional annotations of existing ones. The workflow predicts pathways in which genes are likely to function, identifies PubMed-indexed literature that may support these predictions, generates text summaries describing potential molecular mechanisms, and extracts functional relationships among biological entities from full-text PDF papers. To validate the workflow output, we used a computational approach based on semantic similarity between LLM workflow-generated summaries and Reactome manual annotations. The results show significant enrichment of high-similarity matches. Manual evaluation of 19 genes indicated that more than half of the outputs are useful for supporting curation. Based on these results, we developed an enhanced workflow that incorporates protein-protein interaction data, facilitating Reactomes reaction-based annotation. In summary, our initial adoption of LLM/AI technologies produced encouraging results and provides a practical framework for integrating AI-assisted methods into Reactomes curation pipeline. The strategies described here may be broadly applicable to community knowledgebases in general.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Wu, G., Matthews, L., Boyer, N., Milacic, M., Beavers, D., Li, N. T., May, B., Rothfels, K., Shamovsky, V., Stephan, R., Gillespie, M., Hermjakob, H., DEustachio, P., Stein, L. D.. 2025-12-23. Application of Large Language Models for Annotating Genes into Reactome Pathways. https://doi.org/10.64898/2025.12.20.695723

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Uncovering Viral Signals Across Cell Populations in Colorectal Cancer

Viral sequences have been detected in colorectal cancer, but their cellular and anatomical distribution remains poorly resolved. We aimed to establish a cell-type-resolved map of viral reference-associated RNA signals across colorectal cancer, tumour-adjacent normal tissue, and healthy colon. We repurposed human-unmapped reads from single-cell RNA sequencing libraries generated in ten publicly available studies. We retained cell-barcode and unique molecular identifier information, removed duplicate molecules, and aligned reads to a curated human-associated viral reference using three aligners. Viral signals were analysed across cell types, tissue conditions, and colorectal segments. We used participant-grouped repeated cross-validation to evaluate tissue-condition classification and screened an independent paired colorectal cancer proteomics cohort for compatible peptide matches. The study included 547 libraries from 112 participants, including 49 with matched colorectal cancer and tumour-adjacent normal samples, and approximately 415,000 cells spanning ten major cell populations. Viral reference-associated RNA signals differed across epithelial, immune, stromal, and vascular compartments, tissue conditions, and anatomical regions of the colon. Of 318 independent healthy-reference viral-load comparisons, seven involving six virus-cell-type groups remained significant after Benjamini-Hochberg correction, including plasma-cell HHV-7 and HHV-6B and mast-cell HCV-2. Cell-type-resolved viral features distinguished colorectal cancer, tumour-adjacent normal tissue, and healthy colon in participant-grouped validation; random forest achieved a macro-AUC of 0.957 (SD 0.028), with HERV-K113 as the leading quantitative feature. An independent proteomic screen identified peptide matches compatible with 20 candidate viral groups. We established a cell-type-resolved map of viral reference-associated RNA signals across colorectal tissue conditions and anatomical regions. By identifying candidate viral signals in epithelial and immune compartments, this map provides a basis for investigating virus-host pathways involved in tumour biology, immune regulation, and antigen-presentation states, and could help prioritise future combination strategies pairing virus-directed or virus-mimetic approaches with standard treatment or immune checkpoint blockade.

bioinformatics↗

bioRxiv2026

Statistical significance testing for spatial domain detection in spatial transcriptomics

In spatial transcriptomics, domain-detection methods group platform-specific spatial measurement units--cells, bins or multicellular spots--into spatial domains for comparing molecular programs, cellular composition and interactions among tissue regions. However, the number of candidate domains detected in the same tissue can vary with user-specified parameters or model-selection criteria. Whether candidate domains should be merged can be framed as a test of a single-domain null hypothesis, under which each gene has constant mean expression across all spatial units in the combined region. However, testing candidate domains on the same gene-expression data used to detect them may induce post-selection bias and yield anti-conservative p-values, increasing the risk of oversegmentation. To address these challenges, we introduce SigDoTS (statistical Significance testing for Domain deTection in Spatial transcriptomics), a hypothesis-testing framework for spatial domain detection. SigDoTS uses hierarchical testing to determine which candidate domains to retain or merge, with adjustment for family-wise error control. It accounts for gene-gene covariance and spatial dependence and computes p-values by rerunning the upstream domain-detection pipeline on datasets simulated under the fitted single-domain null. In simulations, SigDoTS controlled the type I error rate at the root node under the single-domain null and improved recovery of the true spatial domains when multiple domains were present. In human dorsolateral prefrontal cortex, SigDoTS retained superficial and deep Layer 6 domains but merged Layer 3 candidate domains across different initial numbers of candidate domains. In ovarian and colorectal cancers, SigDoTS retained epithelial domains with different gene-expression programs and local immune environments, together with immune, stromal and vascular domains.

bioinformatics↗

bioRxiv2026

ASD Neuro Agentic (ANA): An Agentic Framework for Feature Discovery and Cross Site Generalization in Autism fMRI Classification

Classifiers of autism spectrum disorder (ASD) from resting state functional MRI rarely generalize across sites, and many reported accuracies are inflated by data leakage.We show that an agentic artificial intelligence framework, ASD Neuro Agentic (ANA), whose language model agents propose features from the literature and researcher input while fixed tools perform every calculation, can search auditably for site independent classifiers while controlling the leakage. In 1,028 participants from 17 ABIDE I centers, ANA tested 71 pipelines and selected a ComBat harmonized ensemble of tangent space connectivity models from six parcellations and 98 public connectivity derivatives. In leave one center out testing, the ensemble reached an area under the curve of 0.814 (95% CI, 0.773 to 0.841), outperforming ASD DiagNet and ASD SAENet. Predictive connectivity was mainly reduced in ASD. Agentic AI offers a systematic and efficient approach to exploring, evaluating, and identifying pipelines as model selection grows complex.

bioinformatics↗