Paper deep dive

LLMBDC: Language Model for Biological Domains Oriented Clustering of Gene Ontology

Ximing Ran, Jie Xu, Peng Jin, Zhaohui Qin, Zhexing Wen, Jiaying Lu

Year: 2026Venue: arXiv preprintArea: q-bio.GNType: PreprintEmbeddings: 51

Intelligence

Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%

Last extracted: 8/4/2026, 3:48:52 AM

Summary

The paper introduces LLMBDC, a training-free framework that uses large language models (LLMs) to cluster Gene Ontology (GO) terms into higher-order biological domains (BioDomains). It leverages zero-shot semantic reasoning and confidence scoring to overcome the redundancy and lack of context in traditional GO enrichment analysis. Benchmarked against methods like REVIGO and SapBERT on Alzheimer's disease and Fragile X syndrome datasets, LLMBDC demonstrated significantly improved precision, recall, and clustering accuracy (ARI and NMI).

Entities (11)

LLMBDC · softwareframework · 98%Gene Ontology · ontology · 95%Jiaying Lu · person · 95%Ximing Ran · person · 95%Zhexing Wen · person · 95%Alzheimer's disease · disease · 92%Fragile X syndrome · disease · 92%BioDomains · concept · 90%Cauchy Combination Test · statistical-method · 88%REVIGO · softwaretool · 85%SapBERT · softwaremodel · 85%

Relation Signals (10)

LLMBDC → clusters → Gene Ontology

confidence 95% · cluster GO terms into BioDomains

LLMBDC → uses → Large Language Models

confidence 95% · LLMBDC... leverages zero-shot semantic reasoning of LLMs

LLMBDC → appliedto → Fragile X syndrome

confidence 92% · Benchmarked across Alzheimer's disease (AD) and Fragile X syndrome (FXS)

LLMBDC → appliedto → Alzheimer's disease

confidence 92% · Benchmarked across Alzheimer's disease (AD) and Fragile X syndrome (FXS)

Jiaying Lu → authored → LLMBDC

confidence 90% · Jiaying Lu... LlmBdc: Language Model for Biological Domains Oriented Clustering of Gene Ontology

Ximing Ran → authored → LLMBDC

confidence 90% · Ximing Ran... LlmBdc: Language Model for Biological Domains Oriented Clustering of Gene Ontology

Zhexing Wen → authored → LLMBDC

confidence 90% · Zhexing Wen... LlmBdc: Language Model for Biological Domains Oriented Clustering of Gene Ontology

LLMBDC → outperforms → REVIGO

confidence 90% · LLMBDC improved ARI from 9.7% to 73.3% (AD) and from 15.7% to 66.6% (FXS) over REVIGO

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Gene Ontology (GO) enrichment analysis is a foundational tool for translating large-scale genomic data into biological insights, but typically yields hundreds of redundant terms that obscure overarching themes. Existing summarization tools rely on fixed similarity metrics (REVIGO, GOSemSim, clusterProfiler::simplify()), gene-overlap measures (Metascape), or static hierarchy mappings (GO-slim), and therefore cannot incorporate biological context. Manual curation provides context-aware grouping but is subjective and labor-intensive. A scalable, context-aware framework is needed to cluster GO terms into interpretable higher-order biological domains. Here we present LLMBDC (Large Language Model for Biological Domains Oriented Clustering of Gene Ontology), a training-free framework that leverages zero-shot semantic reasoning of LLMs with confidence scoring to cluster GO terms into BioDomains using only ontology information at inference time. Benchmarked across Alzheimer's disease (AD) and Fragile X syndrome (FXS) against six baseline methods including SapBERT, LLMBDC achieved substantially higher precision, recall, and clustering performance. Against ground-truth annotations, LLMBDC improved ARI from 9.7% to 73.3% (AD) and from 15.7% to 66.6% (FXS) over REVIGO, with corresponding NMI gains from 59.9% to 73.4% (AD) and 66.0% to 79.5% (FXS). A Cauchy combination test further confirmed that aggregated BioDomains retained statistically significant functional signals. LLMBDC provides a scalable, reproducible, and interpretable route to context-aware, system-level interpretation of GO enrichment results while preserving biological specificity.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

50,104 characters extracted from source content.

Expand or collapse full text

LlmBdc: Language Model for Biological Domains Oriented Clustering of Gene Ontology Ximing Ran Emory University Atlanta, GA, USA ximing.ran@emory.edu Jie Xu Emory University Atlanta, GA, USA jie.xu@emory.edu Peng Jin Emory University Atlanta, GA, USA peng.jin@emory.edu Zhaohui Qin Emory University Atlanta, GA, USA zhaohui.qin@emory.edu Zhexing Wen ∗ Emory University Atlanta, GA, USA zhexing.wen@emory.edu Jiaying Lu ∗ Emory University Atlanta, GA, USA jiaying.lu@emory.edu Abstract Gene Ontology (GO) enrichment analysis is a foundational tool for translating large-scale genomic data into biological insights, but typically yields hundreds of redundant terms that obscure overar- ching themes. Existing summarization tools rely on fixed similar- ity metrics (REVIGO,GOSemSim,clusterProfiler::simplify()), gene-overlap measures (Metascape), or static hierarchy mappings (GO-slim), and therefore cannot incorporate biological context. Man- ual curation provides context-aware grouping but is subjective and labor-intensive. A scalable, context-aware framework is needed to cluster GO terms into interpretable higher-order biological do- mains. Here we present LlmBdc (Large Language Model for Bio- logical Domains Oriented Clustering of Gene Ontology), a training- free framework that leverages zero-shot semantic reasoning of LLMs with confidence scoring to cluster GO terms into BioDomains us- ing only ontology information at inference time. Benchmarked across Alzheimer’s disease (AD) and Fragile X syndrome (FXS) against six baseline methods including SapBERT, LlmBdc achieved substan- tially higher precision, recall, and clustering performance. Against ground-truth annotations, LlmBdc improved ARI from 9.7% to 73.3% (AD) and from 15.7% to 66.6% (FXS) over REVIGO, with correspond- ing NMI gains from 59.9% to 73.4% (AD) and 66.0% to 79.5%(FXS). A Cauchy combination test further confirmed that aggregated BioDomains retained statistically significant functional signals. LlmBdc provides a scalable, reproducible, and interpretable route to context-aware, system-level interpretation of GO enrichment results while pre- serving biological specificity. ∗ Corresponding authors: Zhexing Wen (zhexing.wen@emory.edu), Jiaying Lu (jiaying.lu@emory.edu) Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full cita- tion on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy other- wise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from permissions@acm.org. BioKDD’26, Jeju, Korea © 2026 Copyright held by the owner/author(s). Publication rights licensed to ACM. ACM ISBN 978-1-4503-X-X/2018/06 https://doi.org/X.X CCS Concepts •Appliedcomputing→Bioinformatics; •Computingmethod- ologies→Knowledge representation and reasoning. Keywords Functional Genomics Analysis, Large Language Models, Gene On- tology Clustering ACM Reference Format: Ximing Ran, Jie Xu, Peng Jin, Zhaohui Qin, Zhexing Wen, and Jiaying Lu. 2026. LlmBdc: Language Model for Biological Domains Oriented Clustering of Gene Ontology. InProceedings of 25th International Workshop on Data Mining in Bioinformatics (BioKDD’26).ACM, New York, NY, USA,10pages. https://doi.org/X.X 1 Introduction With the rise of high-throughput technologies, genomic data have become an increasingly important component of modern biologi- cal research, providing rich information on multi-layered gene reg- ulation across a wide range of conditions, including disease states, developmental stages, and environmental perturbations [ 5,27]. A key step in interpreting these data is to understand the biolog- ical functions associated with molecular features of interest, as identified through various analytical approaches. Gene Ontology (GO) analysis [25] is one of the widely used approaches for this purpose. It identifies significantly overrepresented biological func- tions within a set of genes, providing a structured vocabulary to characterize biological processes, molecular functions, and cellu- lar components – collectively referred to as GO terms [2]. These GO terms offer functional context for genes or gene-associated sig- nals, helping researchers elucidate the mechanisms underlying the studied condition or phenotype. To perform GO analysis, researchers commonly rely on tools such as DAVID [ 22], clusterProfiler [28], or Enrichr [4], which ap- ply statistical methods (e.g., hypergeometric testing) to identify en- riched terms. These tools generate output tables listing significant GO terms alongside adjusted p-values, fold enrichment, and associ- ated genes. While biologically informative, such results often pose two major challenges: (1) the output can be overwhelmingly long, sometimes consisting of hundreds of terms for large gene sets, and (2) the terms are frequently redundant, with overlapping entries (e.g., “immune response” and “regulation of immune response”) arXiv:2608.00099v1 [q-bio.GN] 30 Jul 2026 BioKDD’26, August 09–13, 2026, Jeju, KoreaRan et al. that fragment biological interpretation. Tools like REVIGO [23] attempt to address these issues by clustering semantically simi- lar terms, but they rely on fixed algorithms that cannot adapt to user-defined biological themes or align with the specific focus of a study. To overcome this limitation, some researchers manually an- notate GO terms into custom functional macro-categories. While this enables hypothesis-driven interpretation, the process is labor- intensive, subjective, and poorly reproducible across annotators. In recent years, large language models (LLMs) [ 1,8,30] have emerged as powerful tools capable of understanding and generat- ing human-like language. Trained on massive corpora, LLMs can interpret natural language input, draw on contextual and semantic relationships, and generate structured or free-form responses to a wide range of queries. These capabilities open new opportunities for tasks that involve interpreting and organizing complex biolog- ical information—such as GO term annotation [3]—where flexibil- ity, contextual understanding, and scalability are critical. In this study, we developed atraining-freeandcustomizableAI- enhanced framework that leverages the capabilities of LLMs to sys- tematically annotate Gene Ontology (GO) terms into user-defined biological domains (BioDomains). We refer to this framework as LlmBdc(LargeLanguageModel forBiologicalDomains Oriented Clustering of Gene Ontology). The key insight behind LlmBdc is to cast the GO-to-BioDomain annotation task as a textual seman- tic ranking problem [14,29]. LlmBdc first adaptively generates a prompt describing the request to return a ranked list of the most appropriate BioDomains to each query GO term. LlmBdc then har- nesses the reasoning capabilities of LLMs to perform the ranking based on semantic relevance and prior biomedical knowledge. To evaluate its performance, we applied the proposed LlmBdc to two distinct expert-annotated datasets: one focused on Alzheimer’s disease (AD), a neurodegenerative disorder, containing 7,120 GO terms annotated to 19 BioDomains according to Gregory et al. [ 3]; and another focused on fragile X syndrome (FXS), a genetic neu- rodevelopmental disorder, containing 516 GO terms, which we man- ually annotated to 21 BioDomains. We demonstrated that LLM- driven annotation outperforms baseline methods as well as other models, in both accuracy and consistency. These results support the conclusion that LLMs represent a reliable, efficient, and user- friendly solution for GO-to-BioDomain annotation. Together, our findings show that LLM-based annotation provides a scalable and generalizable approach for organizing GO enrichment results into biologically meaningful domains. By reducing manual effort, en- hancing reproducibility, and enabling flexible, context-specific in- terpretation, LlmBdc bridges the gap between statistical outputs and biological insight. Its ease of use and adaptability make it broadly applicable across diverse research areas involving gene set analysis. More broadly, this work lays the foundation for integrating LLM- driven reasoning into functional genomics workflows, offering a powerful new paradigm for interpreting high-throughput data in a biologically coherent and researcher-driven manner. 2 Related Work 2.1 Biological Domain Annotation Biological domain (BioDomain) annotation offers a systems-level approach to grouping detailed GO term annotations into broader functional categories, which is especially useful for complex dis- eases like Alzheimer’s disease (AD). National Institute on Aging (NIA) and Alzheimer’s Association (A) developed theCommon Alzheimer’s and Related Dementias Research Ontology (CADRO)[ 19], which organizes AD-relevant processes—such as inflammation, me- tabolism, and neuronal function, into standardized domains. Build- ing on this framework, Gaiteri et al. [3] created a genome-wide pipeline that combines genetic association data, predicted variant effects, and multi-omic analyses to rank genes and assign them to 19 curated domains using manually selected GO terms. While effective, this method relies on static mappings that may become outdated or lack flexibility across diseases. Beyond these curated domain mappings, enrichment analyses often produce long lists of functionally redundant or overlapping pathways, making interpretation challenging. To address this is- sue, several computational methods have been developed to clus- ter or simplify enrichment results, thereby improving clarity and biological interpretability. These methods fall into three main cat- egories. First,semantic similarity-based methods leverage the structure of the GO graph and information‑content metrics to com- pute pairwise similarity between terms and remove redundancy ac- cordingly. Representative tools includeREVIGO[ 23],GOSemSim[31], and thesimplify()function inclusterProfiler[28]. Second, gene-set similarityapproaches shift focus from term-level se- mantics to the biological overlap of underlying gene sets. Meth- ods such asMetascape[ 32] compute functional or gene-overlap similarity between the gene sets associated with enriched terms, providing an alternative consolidation strategy. Third,ontology- structure-driventools reduce complexity by leveraging the hier- archical nature of the ontology itself. For example,GO-slim[21] maps specific GO terms to a curated subset of broader, high-level categories, thus offering a compressed but informative summary of enrichment results. While these computational methods offer valu- able complexity reduction, they also present significant limitations in practice. The number and size of clusters generated can be highly variable and often suboptimal for interpretation. For instance,RE- VIGOaccepts only up to 2,000 GO terms at a time and, when given around 500 enriched pathways, may still generate over 200 clusters. Such fragmentation limits the utility of the output, as the user is left with a large number of small or overly specific clusters that re- quire further manual interpretation. Moreover, the granularity and stability of clusters are sensitive to parameter settings and method choice, leading to inconsistent results across different analyses. 2.2 LLMs in Gene Ontology Clustering Gene Ontology (GO) clustering can be viewed as a specialized in- stance of biomedical text summarization, in which fine-grained ontology terms must be condensed and organized into coherent higher-level functional groupings. Large language models (LLMs) have demonstrated strong potential on this broader class of biomed- ical ontology annotation tasks, and progress has gradually nar- rowed from generic biomedical representation learning toward GO- specific reasoning. Early efforts were driven by domain-specialized pretrained LLMs that learn dense semantic biomedical representations. Models such LlmBdc : LLM for Gene Ontology ClusteringBioKDD’26, August 09–13, 2026, Jeju, Korea LLMBDC Clustering Pipeline GO termBioDomain list System Content Disease Description Query LLM GO Name & Definition GO Clustering result Task Introduction GO Term Information GO Gene Set Similarity GO Structure ChatGPT DeepSeek Qwen Best Practice for LLMBDC Validation dataset Alzheimer’s Disease (AD) Fragile X Syndrome (FXS) GO term Biologist Annotated BioDomain LLM Annotated BioDomain Model Temperature ChatGPT-4o-mini DeepSeek-reasoner Qwen3-0.6B Best Fundamental Model for LLMBDA Qwen3-8B Best Model Setting GO Name & Definition GO Structure GO Gene Set Similarity Disease Description Unknown in Output Model Temperature Low :optional prompt input Result Example GO term: B cell activation (GO:0042113) Biologist Annotated BioDomain: Immune Response Model result: Top1. Immune Response Model result: Top2. Cell Cycle Model result: Top3. Apoptosis Biological Domain Significance Test Zero-shot Adaptive Prompting Generation Figure 1: Overview of LlmBdc framework. as SapBERT [12], BioGPT [15], and PubMedBERT [7] provide foun- dational embeddings that are widely reused in downstream on- tology tasks, and have been further specialized for GO through structure-aware variants. GoBERT [17], for instance, couples BERT with the GO graph via two GO-informed pre-training objectives, jointly capturing explicit and implicit relations among gene func- tions. While effective, such specialization typically demands large- scale supervised corpora [ 16], which are scarce in the supervision- sparse regimes characteristic of curated biological domains. A second line of work moves away from fine-tuning toward prompting and agentic reasoning—paradigms well aligned with the summarization view of GO clustering, yet so far underexplored for biomedical ontology annotation. Prompt-based methods such as HiPrompt [14] achieve accurate normalization of clinical terms without parameter updates. Multi-agent frameworks extend LLMs from term-level prediction to gene-set–level functional summariza- tion: GeneAgent [26] and the large-scale evaluation of [10] show that LLMs can generate coherent functional summaries directly from groups of genes. Despite these advances, none of these ap- proaches has been adapted to the specific summarization problem of grouping GO terms into custom, user-defined biological domains— the central task we address. We close this gap by framing GO-to-BioDomain assignment as a zero-shot biomedical annotation task built on top of GO term clus- tering: GO terms are first grouped into semantically coherent clus- ters, and an LLM is then prompted to identify the closest biological theme for each cluster and align it with a higher-level BioDomain. Both the clustering-level theme inference and the term-level do- main assignment are carried out through the same LLM-driven an- notation process, removing the need for static training corpora or pre-defined term–domain mappings. 3 Methods 3.1 Problem setup Enrichment analyses often produce long lists of functionally redun- dant or overlapping pathways, making it difficult to identify over- arching biological themes. In this study, we consider the general task ofGO clustering over biological domains, where the goal is to map enriched GO terms into a set of broader, human-interpretable biological themes. Formulating this as a domain assignment task enables consistent comparison, flexible modeling, and statistical summarization at the domain level. Definition 3.1 (GO Clustering over biological domains).Given a term푡 푖 from a domain-specific ontology (e.g., Gene Ontology) and a list of user-specified biological domainsD= 푑 1 , 푑 2 , ... , 푑 |D| , GO clustering of biological domains aims to assign the most appropri- ate domain labels( ̂ 푑 1 , ̂ 푑 2 , ... )to푡 푖 . 3.2 LlmBdc: Large Language Model for Biological Domains Oriented Clustering of Gene Ontology Following the problem definition, LlmBdc푓 휃 (⋅)treats the cluster- ing problem as a ranking task. Specifically, given a GO term푡and a set of candidate biological domainsD, the model returns a ranked list of domain assignments: ( ̂ 푑 1 , ̂ 푑 2 , ... , ̂ 푑 푘 ) = 푓 휃 (푡,D),(1) where the model parameters휃are obtained from released model weights and require no additional fine-tuning. All code and techni- cal details for this study are publicly available at https://anonymous. 4open.science/r/LLMBDC_auto-615E/. 3.2.1 Adaptive Prompt Generation.LlmBdc leverages a large lan- guage model (LLM) to perform zero-shot GO clustering over bio- logical domains. The key insight is that LLMs can be rapidly adapted BioKDD’26, August 09–13, 2026, Jeju, KoreaRan et al. to new tasks through prompt-based inference without any train- ing [14,29]. The first step in LlmBdc is to construct a prompt픓 from the input GO term푡and the candidate domainsD: 픓 = Ψ(푡,D),(2) whereΨdenotes the prompt generation function tailored for domain- level GO clustering. We implementΨusing a template-based adap- tive scheme. Fig.2presents an example prompt for clustering the GO term “Meiotic Chromosome Condensation” into biological do- mains relevant to Fragile X Syndrome. The hyperparameterTop_k is set to 5 and can be adjusted to balance between output diversity and focus. Beyond the basic prompt structure, we also explore en- riched variants that incorporate additional contextual knowledge— such as domain-specific disease descriptions, GO term definitions, and associated gene sets—as shown in Fig.1. We evaluate these variants empirically in Sec.4. # System prompt to activate relevant domain knowledge in the language model You are a biomedical ontology expert with deep knowl- edge of Gene Ontology, pathway enrichment, and biolog- ical domain classification. # Task Introduction Your task is to assign the most appropriate high-level biological domain(s) to a given GO term, based on its name and functional context.From the list of Biodomains, choose thetop 5labels that best fit this term—ranked most-to-least appropriate. BioDomain Candidates: DNA Repair; Autophagy; Cell Cycle; RNA Metabolism; Epigenetic Regulation ... # GO Term Information GO Term: Meiotic Chromosome Condensation GO Definition: Compaction of chromatin structure prior to meiosis in eukaryotic cells. # Optional prompt input [e.g.,Disease Description] You are a biomedical ontology expert specializing in Frag- ile X Syndrome (FXS) research, the most common inher- ited form of intellectual disability and a leading genetic cause of autism spectrum disorders. FXS is caused by CGG repeat expansion in the FMR1 gene, leading to loss of FMRP and resulting in cognitive impairment, behavioral challenges, and synaptic dysfunction. Your task is to clas- sify Gene Ontology (GO) terms in the context of FXS pathology. # LLM result Answer:LLM∶[”Cell Cycle”,”DNA Repair”,”Structural Stabilization”,”RNA Metabolism”,”Autophagy”] Figure 2: An example of adaptive prompt with task descrip- tion and test query are provided to LLM. 3.2.2 Zero-Shot GO Clustering over Biological Domains.Given the generated prompt픓, we query an instruction-tuned LLM to pro- duce a ranked list of biological domains for a given GO term. For- mally, the LLM outputs: 푊 = 퐿푀 휗 (픓) = 퐿푀 휗 (Ψ(푡,D)),(3) where푊is a sequence of words,Ψis the prompt generator, and휗 represents the frozen parameters of the pre-trained LLM. No train- ing or fine-tuning is required, making this process fully zero-shot. 3.2.3 LLM Output Parsing.The output푊 = (푤 1 , 푤 2 , ... , 푤 푙 )is a raw textual sequence, which needs to be converted into a struc- tured list of ranked biological domains( ̂ 푑 1 , ̂ 푑 2 , ... , ̂ 푑 푘 ). Domain names may vary in length and structure (e.g., “Apoptosis”, “Metal Binding and Homeostasis”). To extract a clean ranked list, we apply a pars- ing functionΦ(⋅): ( ̂ 푑 1 , ̂ 푑 2 , ... , ̂ 푑 푘 ) = Φ(푊 ).(4) We improve parsing reliability by including explicit formatting in- structions in the prompt (e.g., JSON-style output). If the response is not well-structured,Φuses fallback rules such as keyword match- ing and text normalization. 3.2.4 Statistical Aggregation of Enrichment Scores by Domain.Af- ter clustering GO terms into biological domains using the LLM, we aggregate their statistical significance at the domain level. Instead of reporting p-values per term, we summarize them per domain cluster. We use theCauchy Combination Test(CCT) [13], which is suitable for combining dependent p-values. Given a cluster퐶 푗 with 푝-values푝 1 , 푝 2 , ... , 푝 푛 , the test computes: 푇 푗 = 푛 ∑ 푖=1 tan ( 0.5 − 푝 푖 ) 휋 ,(5) 푃 푗 = 1 2 − 1 휋 arctan(푇 푗 ).(6) This produces a single, robust enrichment p value푃 푗 that sum- mary all the pathways in the cluster. We use the p value here for each domain, avoiding assumptions of independence across terms. When compared with expert-curated domain annotations, our ag- gregated scores align well, supporting the biological relevance of the clustering. 4 Experimental Results 4.1 Benchmark Dataset Description Table 1: Benchmark Datasets overview. Category AD FXS #GO Terms7,120 516 #BioDomains19 21 #Avg Domain/GO 1.04 1 #GO w/ 1 Domain 6,834 516 #GO w/ 2 Domain 286 0 To evaluate the performance of LlmBdc, we employed two datasets curated by expert biologists. These datasets serve as ground truth LlmBdc : LLM for Gene Ontology ClusteringBioKDD’26, August 09–13, 2026, Jeju, Korea Table 2: Performance on AD and FXS datasets, reported as mean±standard deviation (in %). DatasetModelPrecision@1 Precision@2 Recall@1 Recall@2 AD Random5.54 ± 0.405.48 ± 0.135.36 ± 0.43 10.57 ± 0.28 GO Similarity5.72 ± 0.235.48 ± 0.155.51 ± 0.24 10.55 ± 0.34 GO Similarity Search20.48 ± 0.12 12.93 ± 0.19 20.03 ± 0.12 25.17 ± 0.34 Graph Traversal20.93 ± 0.56 12.84 ± 0.30 20.25 ± 0.52 24.82 ± 0.54 EditDist8.55 ± 0.007.25 ± 0.008.26 ± 0.00 13.97 ± 0.00 SapBERT48.67 ± 0.00 31.90 ± 0.00 47.16 ± 0.00 61.52 ± 0.00 LlmBdc83.87±0.27 46.80±0.10 83.87±0.27 93.60±0.19 FXS Random1.04 ± 0.231.04 ± 0.231.04 ± 0.23 2.08 ± 0.45 GO Similarity1.43 ± 0.681.27 ± 0.421.43 ± 0.68 2.53 ± 0.85 GO Similarity Search8.25 ± 0.236.50 ± 0.418.25 ± 0.23 13.00 ± 0.81 Graph Traversal13.71 ± 0.117.31 ± 0.17 13.71 ± 0.11 14.62 ± 0.34 EditDist3.12 ± 0.003.22 ± 0.003.12 ± 0.00 6.43 ± 0.00 SapBERT38.21 ± 0.00 25.93 ± 0.00 38.21 ± 0.00 51.85 ± 0.00 LlmBdc78.41±0.25 46.54±0.26 78.41±0.25 93.08±0.52 for BioDomain annotation. Table1provides an overview of these two datasets. Specifically, •AD:The first dataset consists of GO term annotations related to Alzheimer’s disease (AD) risk, sourced from prior studies [3]. The annotations are based on the CADRO framework [19], specifi- cally covering both general aligned and drug trial classification BioDomains. A total of 19 biological domains were annotated, each associated with AD pathology. •FXS:The second dataset is derived from a gene set variation anal- ysis [9] of Fragile X syndrome (FXS) disease. FXS is a genetic disorder characterized by intellectual disability and developmen- tal delay. Biologists conducted functional annotation on 516 GO terms that showed significant differential between healthy con- trols and FXS patients. These GO terms were classified into 21 bi- ological domains, providing a comprehensive view of the molec- ular disruptions associated with FXS. Both conditions involve complex neurological dysfunction but rep- resent distinct etiologies: AD is a late‑onset neurodegenerative dis- order, while FXS is a neurodevelopmental disorder. This contrast allows us to evaluate LlmBdc across heterogeneous biological con- texts, thus assessing its generalizability beyond a single disease framework. 4.2 Exp 1: Comparing with Existing Biological Domain Clustering Methods To validate the effectiveness of LlmBdc, we compare it against a range of zero-shot biological domain annotation baselines across two benchmark datasets: Alzheimer’s Disease (AD) and Fragile X Syndrome (FXS). These baselines are grouped into three main methodological categories: (a) Gene-set similarity-based methods: •GOSimilarity: this method ranks biological domains by Jaccard similarity between the gene set of a GO term and those of can- didate domains. If no gene overlap is found, a random domain is assigned. •GO Similarity Search: this method extends GO similarity by in- corporating a neighborhood-based strategy. For each BioDomain, a neighborhood is constructed by selecting its top 10 most sim- ilar GO terms based on gene set overlap. Domain assignment is then performed by measuring the similarity between the input GO term and these BioDomain-specific neighbors. (b) GO structure-based methods: •Graph Traversal: this method maps GO terms to biological do- mains by tracing their ancestors in the GO hierarchy. If an ances- tor matches a known domain, it is assigned; otherwise, a random domain is used. Multiple domain matches are all included. It is worth noting that we found that 9 out of 19 biological domains in the AD dataset and 7 out of 21 in the FXS dataset can be located in GO hierarchy. (c) GO term semantic similarity-based methods: •EditDist[20]: this method uses Levenshtein edit distance to as- sign biological domains to GO terms based purely on string sim- ilarity between their names. •SapBERT[12]: this method represents a semantic embedding- based biological domain idea, where GO terms and candidate domain names are embedded using SapBERT, and cosine simi- larity is used to rank the candidates. Specifically, SapBERT is a pretrained Transformer model optimized for biomedical concept linking, pre-trained on PubMed-derived biomedical texts. We also include aRandommethod that assigns domains uniformly at random as a naive baseline. Table 2presents the benchmarking results. Following the prob- lem setup in Sec.3.1, we adopt the ranking based metrics includ- ing Precision@K and Recall@K as the main evaluation metrics (the higher the better). Across both datasets, LlmBdc with DeepSeek- reasoner clearly outperforms all compared methods in both pre- cision and recall. Notably, classical methods relying on gene set or semantic similarity fall short in domain generalization, while LlmBdc effectively captures biological relevance with no task-specific training. BioKDD’26, August 09–13, 2026, Jeju, KoreaRan et al. Table 3: Domain-level comparison between expert and LlmBdc on the AD dataset. BioDomainTruth p LlmBdc p Truth Rank LlmBdc Rank Mitochondrial Me- tabolism 3.68E-10 5.69E-1011 Synapse8.01E-05 8.62E-0522 Cell Cycle3.19E-04 1.45E-01 314 Endolysosome3.88E-04 3.84E-0444 Structural Stabiliza- tion 1.42E-03 2.11E-04 53 Lipid Metabolism 1.47E-03 1.35E-03 65 DNA Repair4.17E-03 4.36E-03 78 Metal Binding & Homeostasis 8.26E-03 3.98E-03 87 Oxidative Stress9.28E-03 1.19E-0299 Proteostasis9.74E-03 1.62E-021010 Myelination2.61E-02 2.42E-021111 APP Metabolism5.94E-02 5.50E-021212 Vasculature7.89E-02 8.21E-021313 Apoptosis1.90E-01 2.84E-01 1416 Tau Homeostasis2.11E-01 2.11E-011515 RNA Spliceosome 7.81E-01 4.84E-01 1617 Immune Response1.001.001718 Epigenetic1.001.001819 Autophagy1.001.48E-03 196 73.4% 73.4% 100% 9.7% 59.9% 64.3% 66.6% 79.5% 100% 15.7% 66% 37% ADFXS ARINMICoverageARINMICoverage 0.00 0.25 0.50 0.75 1.00 Score Method LLMBDC REVIGO Figure 3: Domain-level signficance rank comparison be- tween LlmBdc and expert-based GO clustering. Most top- ranked domains by LlmBdc are consistent with expert pri- oritization. 4.2.1 In-depth analysis.We further assess the alignment between LlmBdc-generated clustering and expert-defined domain-level GO enrichment using statistical evidence. Specifically, we compare the domain-level Cauchy Combination Test (CCT) [ 13]푝-values from LlmBdc to those derived from expert-labeled GO clusters across two benchmark datasets: Alzheimer’s Disease (AD) and Fragile X Syndrome (FXS). As shown in Fig. 3, the top-ranked biological do- mains prioritized by LlmBdc closely match the expert-derived clus- ters in both datasets. To further support this finding, we present domain-level푝-values and rankings side-by-side in Tables 3and4. Across both datasets, LlmBdc reliably ranks the most biologically relevant domains at the top. In AD,Mitochondrial Metabolismand Synapseare the top two in both expert and LlmBdc results. In FXS, Table 4: Domain-level comparison between expert and LlmBdc on the FXS dataset. BioDomainTruth p LlmBdc p Truth Rank LlmBdc Rank Immune System and Inflammation 1.09E-05 1.06E-0511 Epigenetic Regula- tion 5.34E-05 4.75E-05 23 Structural Stabiliza- tion 6.16E-05 2.64E-0532 Neurotransmission and Synaptic Regu- lation 7.47E-05 7.47E-0544 Autophagy1.09E-04 1.27E-0455 Signal Transduc- tion 2.12E-04 2.69E-0467 DNA Repair3.47E-04 7.75E-0479 Protein Metabolism and Trafficking 4.47E-04 5.46E-03 816 System and Devel- opmental Process 7.51E-04 4.65E-04 98 Apoptosis1.84E-03 1.92E-031012 Response to Stimu- lus 1.84E-03 1.61E-03 1111 Gliogenesis and Glial Differentia- tion 1.98E-03 1.97E-031213 Mitochondrial Function and Me- tabolism 2.34E-03 2.00E-031314 RNA Metabolism 3.58E-03 3.48E-031415 Neurodevelopment and Neuronal Dif- ferentiation 4.92E-03 5.64E-031518 Intracellular Traf- ficking and Or- ganelle Dynamics 5.19E-03 2.45E-04 166 Lipid Metabolism 5.31E-03 5.55E-03 1717 Transcription and Translation Ma- chinery 5.66E-03 8.36E-031819 Molecular Trans- port and Homeosta- sis 6.06E-03 9.06E-04 1910 Cell Cycle1.11E-02 2.34E-022021 Cell Adhesion and Interaction 3.72E-02 2.31E-02 2120 domains likeImmune System,Structural Stabilization, andNeuro- transmissionshow strong agreement. While minor mismatches oc- cur, such as inCell CycleorProtein Metabolism, the overall ranking correlation indicates that LlmBdc effectively recovers expert-level prioritization without manual input. LlmBdc : LLM for Gene Ontology ClusteringBioKDD’26, August 09–13, 2026, Jeju, Korea Table 5: Performance on AD and FXS datasets, reported as mean±standard deviation (in %). DatasetModelPrecision@1 Precision@2 Recall@1 Recall@2 ADLlmBdc (Qwen3-0.6B)19.55 ± 0.17 14.97 ± 0.07 18.72 ± 0.17 28.44 ± 0.10 LlmBdc (Qwen3-8B)77.02 ± 0.15 44.20 ± 0.05 75.29 ± 0.15 85.35 ± 0.12 LlmBdc (ChatGPT-4o-mini)77.39 ± 0.68 43.55 ± 0.34 77.39 ± 0.68 87.09 ± 0.69 LlmBdc (DeepSeek-reasoner)83.87±0.27 46.80±0.10 83.87±0.27 93.60±0.19 FXSLlmBdc (Qwen3-0.6B)5.31 ± 4.238.16 ± 7.935.31 ± 4.23 16.31 ± 1.59 LlmBdc (Qwen3-8B)67.77 ± 0.23 39.31 ± 0.06 67.77 ± 0.23 78.62 ± 0.11 LlmBdc (ChatGPT-4o-mini)36.26 ± 0.94 23.76 ± 0.61 36.26 ± 0.94 47.51 ± 1.22 LlmBdc (DeepSeek-reasoner)78.41±0.25 46.54±0.26 78.41±0.25 93.08±0.52 4.3 Exp 2: Comparing with Existing GO Clustering Method To assess the advantages of our proposed LlmBdc over traditional clustering approaches, we compared it against REVIGO [23], a widely used method for reducing redundancy in GO enrichment results. REVIGO groups semantically similar GO terms based on informa- tion‑content measures and does not rely on any external domain knowledge. In contrast, LlmBdc leverages an LLM to assign GO terms to user‑defined biological domains, even when the domain list is not previously aligned with the ontology. To provide a fair and challenging comparison, we ran both methods under the same “unknown biodomain list” setting (i.e.,meaning models was not given any predefined mapping), and evaluated how well their out- puts match ground‑truth domain annotations. We conducted a sec- ond benchmarking analysis based on three quantitative metrics: Adjusted Rand Index (ARI), Normalized Mutual Information (NMI), and Coverage. Using the unknown biodomain list as input, we compared the semantic clusters produced by REVIGO against the biodomain assignments generated by LlmBdc. Across both datasets, LlmBdc consistently outperformed REVIGO. For the AD dataset, LlmBdc achieved an ARI of 0.73, NMI of 0.73, and full Coverage (100%), whereas REVIGO obtained substantially lower scores (ARI = 0.10, NMI = 0.60) and covered only 64% of the enriched GO terms. A similar pattern was observed in FXS: LlmBdc reached ARI = 0.67, NMI = 0.80, and 100% Coverage, compared to REVIGO’s ARI = 0.16, NMI = 0.66, and 37% Coverage. These results demonstrate that even without preset biodomain information, LlmBdc produces clusters that are more consistent, information-rich, and comprehensive than those generated by REVIGO. 4.4 Ablation Study We systematically evaluated the optimal configuration for deploy- ing LlmBdc in GO clustering over biological domains using two benchmark datasets: Alzheimer’s Disease (AD) and Fragile X Syn- drome (FXS), with expert annotations as reference. Our analysis focused on three factors: decoding temperature, the use of an “Un- known” label, and prompt content. Temperature and “Unknown” Label.As shown in Fig. 4aand Fig.4b, a low decoding temperature (0.0 or 0.1) yields better pre- cision and recall, suggesting that deterministic decoding ensures more consistent clustering. Additionally, disabling the “Unknown” label option improves recall by encouraging full domain assign- ments, even in uncertain cases. Prompt Input.We tested four prompt input formats and found that using only the GO term name and its definition led to the best results. Adding GO structure, gene set metadata, or disease descriptions did not improve performance and sometimes reduced accuracy, likely due to added noise. Model Choice of LLM.Table5reports the performance of vari- ous LLM models used in LlmBdc. We evaluate both open-source models (Qwen3-0.6B and Qwen3-8B [30]) and commercial mod- els (ChatGPT-4o-mini [1] and DeepSeek-reasoner [8]). DeepSeek- reasoner consistently outperforms all others, achieving the highest precision and recall on both datasets. RecommendedConfiguration.Our findings suggest that the best practice for LlmBdc is to use DeepSeek-reasoner with a low tem- perature (0.0–0.1),disable the “Unknown” label, and only in- clude the GO name and definition in the prompt. This configura- tion provides the most accurate and interpretable clustering per- formance. 5 Discussion FlexibilityandCustomizability.A key strength of LlmBdc is its flexibility and customizability. Unlike static clustering tools such as REVIGO, which group GO terms based on fixed algorithms or ontology tree distances, LlmBdc can incorporate custom-defined BioDomains and include optional comprehensive information—such as disease-specific descriptions or GO term hierarchy structures— into its annotation process. Currently, BioDomains are defined by domain experts, such as biologists or clinical researchers, based on prior knowledge and the specific context of the study. In the fu- ture, these BioDomains could be generated with the assistance of large language models (LLMs), reducing manual effort and improv- ing scalability. This makes LlmBdc well suited for researchers who wish to organize GO terms according to domain-specific biological themes, such as “Neural Circuit Assembly,” “Stress Response Path- ways,” or “Mitochondrial Function,” depending on the study focus. Another practical advantage of LlmBdc lies in its ability to flexi- bly accommodate updates to the GO database. Because it operates through natural language reasoning rather than relying on fixed ontology structures or predefined hierarchical mappings, LlmBdc can seamlessly adapt to changes in the GO term catalog. As the BioKDD’26, August 09–13, 2026, Jeju, KoreaRan et al. Input Type GO name GO structure GO similarity Disease Precision@1 0.40.60.8 Precision@2 0.20.30.4 Recall@1 0.40.60.8 Recall@2 AD FXS 0.40.60.8 Model & Condition ChatGPT With Unknown ChatGPT Without Unknown DeepSeek With Unknown DeepSeek Without Unknown (a) Unknown label effect FXS − Precision@1FXS − Precision@2FXS − Recall@1FXS − Recall@2 AD − Precision@1AD − Precision@2AD − Recall@1AD − Recall@2 0.00.10.50.90.00.10.50.90.00.10.50.90.00.10.50.9 0.80 0.84 0.88 0.92 0.4 0.5 0.6 0.7 0.8 0.9 0.72 0.75 0.78 0.81 0.3 0.4 0.5 0.6 0.7 0.40 0.42 0.44 0.46 0.20 0.25 0.30 0.35 0.40 0.45 0.72 0.75 0.78 0.81 0.3 0.4 0.5 0.6 0.7 Temperature ChatGPTDeepSeek (b) Decoding temperature effect Figure 4: Effect of prompting input, decoding temperature and “Unknown” label option on LlmBdc performance across AD and FXS datasets. Gene Ontology continues to evolve—with new terms added, exist- ing definitions refined, and obsolete terms removed—LlmBdc al- lows researchers to incorporate the most current, or any version of the ontology, into their analyses. This flexibility reduces reliance on static annotations and ensures that downstream biological in- terpretations remain aligned with the latest community standards. LlmBdc : LLM for Gene Ontology ClusteringBioKDD’26, August 09–13, 2026, Jeju, Korea Scalability,Reproducibility,andModelSelection.Importantly, LlmBdc also improves scalability and reproducibility. Manual group- ing of GO terms is labor-intensive and prone to subjectivity, often leading to inconsistency between annotators. By applying consis- tent semantic reasoning through LLMs, LlmBdc reduces inter-user variability and allows for reproducible annotation at scale. This feature is particularly valuable in large collaborative projects or automated bioinformatics pipelines where standardization is cru- cial. We also explored how the choice of underlying LLM affects LlmBdc’s performance. Using an identical prompt structure, we compared annotations generated by different fundamental models (e.g., gpt-4o-mini, DeepSeek-Reasoning). Interestingly, DeepSeek Reasoning achieved higher annotation accuracy and closer align- ment with human-labeled BioDomains. However, this improvement came at a cost: DeepSeek consumed more tokens, required longer runtime, and incurred higher computational expense than gpt-4o- mini. These differences highlight that the performance gap is not necessarily due to one model being universally superior, but may instead reflect distinctions in their reasoning depth, prompt inter- pretation, or architectural design. Reasoning-oriented models like DeepSeek may be particularly well-suited for tasks involving nu- anced semantic ranking, but their higher cost and slower through- put may limit scalability in certain use cases. Therefore, users should weigh the benefits in annotation quality against practical trade-offs such as speed, cost, and computational resources when selecting an LLM backend for LlmBdc. LimitationsandFutureDirections.While LlmBdc shows strong performance and versatility, there are limitations to consider. LLM outputs can still be influenced by prompt phrasing, model-specific biases, or limited knowledge of rare or emerging GO terms. More- over, although LlmBdc allows optional contextual input, its cur- rent implementation does not yet fully leverage structured biologi- cal knowledge graphs [ 6] or curated databases such as KEGG [11], Reactome [18], or STRING [24]. Future extensions of this frame- work could explore hybrid strategies that integrate LLM reason- ing with structured ontologies, graph-based embeddings, or path- way co-membership information. Additionally, as larger and more specialized LLMs become available, benchmarking across domains (e.g., immunology, developmental biology, or cancer) will be im- portant for optimizing annotation performance. Ethical Considerations.LlmBdc relies on large language models that may inherit biases from their training data, potentially lead- ing to annotation inaccuracies for underrepresented or emerging biological processes. Users should exercise caution when interpret- ing outputs for poorly characterized GO terms or novel domains where the LLM lacks sufficient knowledge. All data used in this study (e.g., GO term annotations and curated BioDomain labels) are derived from published, fully de‑identified sources and comply with HIPAA standards, as no human subject or personally identi- fiable information was involved. Additionally, while LlmBdc im- proves reproducibility by automating annotation, it does not elim- inate the need for expert review, especially in high‑stakes settings such as clinical target discovery or drug development. Transparent reporting of the underlying LLM version, prompt design, and any optional context is essential to ensure result interpretability and cross‑study comparability. 6 Conclusion We introducedLlmBdc, a flexible, zero-shot framework that lever- ages large language models to translate Gene Ontology (GO) terms into biologically meaningful, user-defined BioDomains. By fram- ing GO-to-BioDomain annotation as a semantic ranking problem, LlmBdc avoids costly training and manual curation while enabling context-aware, interpretable grouping of GO terms. We evaluated LlmBdc on two biologically distinct, manually curated datasets, which differed in GO term counts and BioDomain definitions. Across both settings, LlmBdc consistently outperformed traditional GO- based methods and alternative models like SapBERT, achieving higher accuracy and semantic alignment with human annotations. These results demonstrate strong generalizability across diverse bi- ological systems and research contexts. By reducing redundancy, improving reproducibility, and adapting seamlessly to user-defined annotation schemes, LlmBdc bridges the gap between statistical enrichment outputs and higher-level biological interpretation. It empowers researchers to extract clearer insights from large-scale gene sets, accelerate hypothesis generation, and stay aligned with evolving GO standards. These capabilities make the proposed LlmBdc a practical, interpretable tool for modern functional genomics anal- ysis. References [1]Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Flo- rencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shya- mal Anadkat, et al. 2023. Gpt-4 technical report.arXiv preprint arXiv:2303.08774 (2023). [2]Michael Ashburner, Catherine A Ball, Judith A Blake, David Botstein, Heather Butler, J Michael Cherry, Allan P Davis, Kara Dolinski, Selina S Dwight, Janan T Eppig, et al. 2000. Gene ontology: tool for the unification of biology.Nature genetics25, 1 (2000), 25–29. [3]Gregory A. Cary, Jesse C. Wiley, Jake Gockley, Stephen Keegan, Sai Sruthi Amirtha Ganesh, Laura Heath, Robert R. Butler, Lara M. Mangravite, Ben- jamin A. Logsdon, Frank M. Longo, Allan Levey, Anna K. Greenwood, and Gre- gory W. Carter. 2024. Genetic and Multi-Omic Risk Assessment of Alzheimer’s Disease Implicates Core Associated Biological Domains(Alzheimer’s & Demen- tia: Translational Research & Clinical Interventions, Vol. 10, No. 2). e12461. doi:10. 1002/trc2.12461 [4]Edward Y. Chen, Christopher M. Tan, Yan Kou, Qiaonan Duan, Zichen Wang, Gabriela V. Meirelles, Neil R. Clark, and Avi Ma'ayan. 2013. Enrichr: interactive and collaborative HTML5 gene list enrichment analysis tool.BMC Bioinformat- ics14 (2013), 128. doi:10.1186/1471-2105-14-128 [5]Li Chen, Peng Jin, and Zhaohui S Qin. 2016. DIVAN: accurate identification of non-coding disease-specific risk variants using multi-omics profiles.Genome biology17, 1 (2016), 252. [6]Hejie Cui, Jiaying Lu, Ran Xu, Shiyu Wang, Wenjing Ma, Yue Yu, Shaojun Yu, Xuan Kan, Chen Ling, Liang Zhao, Zhaohui S. Qin, Joyce Ho, Tianfan Fu, Jing Ma, Mengdi Huai, Fei Wang, and Carl Yang. 2025. A review on knowledge graphs for healthcare: Resources, applications, and promises.Journal of Biomedical In- formatics(July 2025). doi:10.1016/j.jbi.2025.104861 [7]Yu Gu, Robert Tinn, Hao Cheng, Marie Lucas, Naoto Usuyama, Xiaodong Liu, Tristan Naumann, Jianfeng Gao, and Hoifung Poon. 2022. PubMedBERT: A Pre- trained Biomedical Language Model for Biomedical Text Mining.Bioinformatics 38, 20 (2022), 5506–5513. doi:10.1093/bioinformatics/btac682 [8]Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. 2025. Deepseek-r1: Incen- tivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948(2025). [9]Sonja Hänzelmann, Robert Castelo, and Justin Guinney. 2013. GSVA: gene set variation analysis for microarray and RNA-seq data.BMC bioinformatics14, 1 (2013), 7. [10]Mengzhou Hu, Sahar Alkhairy, Ingoo Lee, Rudolf T. Pillich, Dylan Fong, Kevin Smith, Robin Bachelder, Trey Ideker, and Dexter Pratt. 2025. Evaluation of large language models for discovery of gene set function.Nature Methods22 (2025), 82–91.doi:10.1038/s41592-024-02120-3 [11]Minoru Kanehisa, Miho Furumichi, Yoko Sato, Masayuki Kawashima, and Mari Ishiguro-Watanabe. 2023. KEGG for taxonomy-based analysis of pathways and BioKDD’26, August 09–13, 2026, Jeju, KoreaRan et al. genomes.Nucleic acids research51, D1 (2023), D587–D592. [12]Fangyu Liu, Ehsan Shareghi, Zaiqiao Meng, Marco Basaldella, and Nigel Collier. 2020. Self-alignment pretraining for biomedical entity representations.arXiv preprint arXiv:2010.11784(2020). [13]Yujun Liu and Jichun Xie. 2020. Cauchy combination test: a powerful test with analytic p-value calculation under arbitrary dependency structures.J. Amer. Statist. Assoc.115, 529 (2020), 393–402. doi:10.1080/01621459.2018.1554485 [14]Jiaying Lu, Jiaming Shen, Bo Xiong, Wengjing Ma, Staab Steffen, and Carl Yang. 2023. HiPrompt: Few-Shot Biomedical Knowledge Fusion via Hierarchy- Oriented Prompting. In46th International ACM SIGIR Conference on Research and Development in Information Retrieval ((SIGIR)).doi:10.1145/3539618.3591997 [15]Rui Luo, Lianhui Sun, Yuxuan Xia, Bing Qin, and Ting Liu. 2022. BioGPT: Gener- ative Pre-trained Transformer for Biomedical Text Generation and Mining.Brief- ings in Bioinformatics(2022). [16]Yuwei Miao, Yuzhi Guo, Hehuan Ma, Jingquan Yan, Feng Jiang, Weizhi An, Jean Gao, and Junzhou Huang. 2024. UniEntrezDB: Large-scale Gene Ontology Anno- tation Dataset and Evaluation Benchmarks with Unified Entrez Gene Identifiers. arXiv:2412.12688[cs.DB]https://arxiv.org/abs/2412.12688 [17]Yuwei Miao, Yuzhi Guo, Hehuan Ma, Jingquan Yan, Feng Jiang, Rui Liao, and Junzhou Huang. 2025. GoBERT: Gene Ontology Graph Informed BERT for Uni- versal Gene Function Prediction. arXiv:2501.01930[cs.LG]https://arxiv.org/ abs/2501.01930 [18]Marija Milacic, Deidre Beavers, Patrick Conley, Chuqiao Gong, Marc Gillespie, Johannes Griss, Robin Haw, Bijay Jassal, Lisa Matthews, Bruce May, et al. 2024. The reactome pathway knowledgebase 2024.Nucleic acids research52, D1 (2024), D672–D678. [19]National Institute on Aging and Alzheimer’s Association. 2024. Common Alzheimer’s and Related Dementias Research Ontology (CADRO).https://iadrp. nia.nih.gov/about/cadroAccessed: July 2025. [20]Eric Sven Ristad and Peter N Yianilos. 2002. Learning string-edit distance.IEEE Transactions on Pattern Analysis and Machine Intelligence20, 5 (2002), 522–532. [21]Emily B Sessa, Rishi R Masalia, N Arrigo, Michael S Barker, and John A Pelosi. 2023. GOgetter: A pipeline for summarizing and visualizing GO slim annotations for plant genetic data.Applications in Plant Sciences11, 4 (2023), e11536.doi:10. 1002/aps3.11536 [22]Brad T. Sherman, Ming Hao, Ju Qiu, Xiaoli Jiao, Michael W. Baseler, H. Clifford Lane, Tomozumi Imamichi, and Weizhong Chang. 2022. DAVID: a web server for functional enrichment analysis and functional annotation of gene lists (2021 update).Nucleic Acids Research50, W1 (2022), W216–W221. doi:10.1093/nar/ gkac194 [23]Fran Supek, Maja Bošnjak, Nives Škunca, and Tomislav Šmuc. 2011. REVIGO summarizes and visualizes long lists of gene ontology terms.PLoS One6, 7 (2011), e21800. doi:10.1371/journal.pone.0021800 [24]Damian Szklarczyk, Rebecca Kirsch, Mikaela Koutrouli, Katerina Nastou, Far- rokh Mehryary, Radja Hachilif, Annika L Gable, Tao Fang, Nadezhda T Doncheva, Sampo Pyysalo, et al. 2023. The STRING database in 2023: protein– protein association networks and functional enrichment analyses for any se- quenced genome of interest.Nucleic acids research51, D1 (2023), D638–D646. [25]Paul D Thomas, Dustin Ebert, Anushya Muruganujan, Tremayne Mushayahama, Laurent-Philippe Albou, and Huaiyu Mi. 2022. PANTHER: Making genome-scale phylogenetics accessible to all.Protein Science31, 1 (2022), 8–22. [26]Zhizheng Wang, Qiao Jin, Chih-Hsuan Wei, Shubo Tian, Po-Ting Lai, Qingqing Zhu, Chi-Ping Day, Christina Ross, Robert Leaman, and Zhiyong Lu. 2025. Ge- neAgent: self-verification language agent for gene-set analysis using domain databases.Nature Methods(2025). doi:10.1038/s41592-024-02111-4 [27]Zhexing Wen, Ha Nam Nguyen, Ziyuan Guo, Matthew A Lalli, Xinyuan Wang, Yijing Su, Nam-Shik Kim, Ki-Jun Yoon, Jaehoon Shin, Ce Zhang, et al. 2014. Synaptic dysregulation in a human iPS cell model of mental disorders.Nature 515, 7527 (2014), 414–418. [28]Tian Wu, Eryong Hu, Shuqing Xu, Meng Chen, Peizhi Guo, Zhitang Dai, Ting Feng, Lin Zhou, Wanyi Tang, Li Zhan, Xi Fu, Sheng Liu, Xiaowen Bo, and Guangchuang Yu. 2021. clusterProfiler 4.0: A universal enrichment tool for in- terpreting omics data.The Innovation2, 3 (2021), 100141.doi:10.1016/j.xinn.2021. 100141 [29]Yuzhang Xie, Jiaying Lu, Joyce Ho, Fadi Nahab, Xiao Hu, and Carl Yang. 2024. PromptLink: Leveraging Large Language Models for Cross-Source Biomedical Concept Linking. In47th International ACM SIGIR Conference on Research and Development in Information Retrieval. [30]An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. 2025. Qwen3 technical report.arXiv preprint arXiv:2505.09388(2025). [31]Guangchuang Yu, Fei Li, Yide Qin, Xiaochen Bo, Yibo Wu, and Shengqi Wang. 2010. GOSemSim: an R package for measuring semantic similarity among GO terms and gene products.Bioinformatics26, 7 (2010), 976–978.doi:10.1093/ bioinformatics/btq064 [32]Yuyang Zhou, Beiqi Zhou, Lars Pache, Miao Chang, Amir H Khodabakhshi, Olga Tanaseichuk, Christopher Benner, and Sumit K Chanda. 2019. Metascape pro- vides a biologist-oriented resource for the analysis of systems-level datasets.Na- ture communications10, 1 (2019), 1523.doi:10.1038/s41467-019-09234-6

Instant research discovery