GeneReL: a large language model-powered platform for gene regulatory relationship extraction with community curation

Abstract

Gene regulatory networks provide fundamental insights into plant biology, yet extracting structured interaction data from scientific literature remains a significant bottleneck. Traditional manual curation cannot scale to meet the demands of modern research, while automated text mining approaches struggle with the complexity of gene nomenclature and relationship classification. Large language models (LLMs) offer promising capabilities for information extraction, but integrated platforms combining LLM extraction with community validation for plant regulatory databases remain scarce.

We developed GeneReL, an integrated platform combining LLM-based extraction with community-driven curation for gene regulatory networks in Arabidopsis thaliana. The system employs a tiered pipeline using Claude Haiku 4.5 for screening, Claude Sonnet 4 for extraction, and Claude Opus 4 for verification, along with a novel five-step gene normalization pipeline incorporating paper-text search and LLM-based disambiguation with UniProt annotations. The database contains 13,710 curated interactions across 51 relationship types, with 90.2% classified as high confidence based on linguistic certainty markers in source text. Comparison with IntAct reveals 86.0% of interactions are unique to our literature-derived database, demonstrating complementary coverage to existing resources. The web platform provides card-based browsing with voting capabilities, interactive network visualization using Cytoscape.js with locus-ID-based node consolidation, and administrative interfaces for curator review of ambiguous gene mappings.

GeneReL is freely accessible at https://generel.newgenes.me.

Introduction

Gene regulatory networks (GRNs) represent the complex web of molecular interactions governing cellular processes in plants. In Arabidopsis thaliana, decades of research have elucidated numerous regulatory pathways controlling flowering time, organ identity, hormone signaling, and stress responses [1]. The importance of comprehensive GRN databases has driven development of resources such as AGRIS and specialized tools like AGENT for exploring published regulatory networks [2,3]. However, significant challenges remain in capturing the rapidly growing body of regulatory information scattered across thousands of publications.

The information extraction challenge

The biomedical literature expands at an unprecedented rate, with gene-gene interactions typically described in natural language text. Traditional manual curation approaches cannot scale proportionally with literature growth—a systematic analysis of TAIR curation workflows estimated that only ∼30% of relevant Arabidopsis literature receives curator attention [4]. Community curation has been explored as a strategy, with the CACAO project demonstrating that undergraduates can contribute high-quality functional annotations [5]. However, even successful community initiatives require substantial coordination and expertise.

Large language models (LLMs) represent a paradigm shift in text processing capabilities, with growing applications in plant biology [6]. Recent evaluations have benchmarked LLMs including GPT-4 and domain-specific models on biomedical tasks such as named entity recognition and relation extraction [7,8]. A comprehensive evaluation of 21 LLMs on gene regulatory relation extraction found that GPT-4 and Claude-Pro achieved F1 scores of 0.4448 and 0.4386, respectively, demonstrating both potential and remaining challenges for this task [9]. These studies reveal that LLMs can achieve competitive performance compared to traditional methods, particularly for complex information extraction requiring contextual understanding [10]. Recent work has also explored LLMs for gene regulatory network inference from expression data (LLM4GRN), though literature-based extraction remains underexplored [11]. The GIX framework demonstrated a three-stage approach for automated extraction of genetic interactions from literature using language models [12], while TRRUST pioneered sentence-based text mining for TF-target interaction curation in human and mouse [13].

Distinction from existing resources

Several databases provide regulatory interaction data for plants. AGRIS (Arabidopsis Gene Regulatory Information Server) and its successor AtRegNet represent the primary Arabidopsis-specific regulatory network resources, containing transcription factor binding site information and curated TF-target interactions [14]. PlantRegMap extends regulatory mapping across 165 plant species with curated TF-target interactions derived from ChIP-seq and other high-throughput experiments [15]. PlantTFDB provides comprehensive transcription factor annotations with regulatory relationship predictions [16].

GeneReL complements these resources by (i) extracting relationships beyond TF-target interactions to include post-translational modifications, physical interactions, and miRNA targeting (51 distinct relationship types vs. 3–4 in AGRIS); (ii) capturing regulatory evidence from narrative text rather than high-throughput data alone; and (iii) implementing community validation through voting mechanisms for continuous quality improvement.

RegNetwork 2025 recently demonstrated large-scale literature-based regulatory network extraction for mammalian systems, integrating 31 databases and over 20,000 published papers to compile more than 11 million regulatory relationships for human and mouse [17]. GeneReL addresses the distinct needs of plant biology with Arabidopsis-specific gene nomenclature handling and a focus on semantic relationship types rather than association-based scoring.

Among broader molecular interaction databases, STRING integrates known and predicted protein-protein interactions across thousands of organisms. The 2025 update (STRING v12.0) now includes a regulatory interaction mode with directionality information, marking a significant enhancement [18]. IntAct provides expertly curated molecular interaction data derived from literature curation and high-throughput experiments, serving as a primary reference for protein interaction validation [19]. For Arabidopsis, STRING v12 contains 27,462 proteins with over 12.5 million putative interactions, though only 5.3% meet the high-confidence threshold (score ≥ 700).

Objectives

This study presents a platform integrating LLM-based extraction with community-driven curation for gene-gene interaction discovery. We address three objectives: (i) develop an automated pipeline leveraging the Claude API to extract structured regulatory relationships with gene identifier normalization; (ii) implement a web-based platform for browsing, filtering, and visualizing interactions; and (iii) establish community curation through voting mechanisms.

Materials and methods

System architecture

The platform comprises four integrated components: (i) data source management, (ii) an automated extraction pipeline, (iii) a Firebase backend for data storage and user management, and (iv) a web frontend for visualization and community curation (Supplementary Fig. S1).

Literature retrieval

Full-text articles were retrieved from PubMed Central using the BioPython Entrez module. Publication date filters restricted retrieval to 2010–2024. PMC articles in JATS XML format were parsed using lxml (https://lxml.de/), with section-aware chunking employing overlapping segments of 3,000 words with 200-word overlap for articles exceeding the LLM context window.

LLM-based extraction

The extraction pipeline employs a tiered multi-model approach optimizing for both accuracy and cost-efficiency. We selected Claude models based on preliminary evaluations demonstrating stable extraction performance across varied scientific text scenarios, consistent with recent findings on LLM stability in chemical-disease extraction tasks [20]. Abstract screening uses Claude Haiku 4.5 (claude-haiku-4-5-20251001) to filter irrelevant papers before full extraction. Primary extraction uses Claude Sonnet 4 (claude-sonnet-4-20250514) based on its 200K token context window and strong performance on scientific text comprehension benchmarks. Verification uses Claude Opus 4 (claude-opus-4-20250514) to validate extracted interactions, confirming that both entities are genes (not phenotypes or processes), the evidence sentence supports the stated relationship, and the relationship type is correctly classified. The extraction prompt comprises task description, output schema definition, relationship taxonomy, and few-shot examples; the exact prompts used at each pipeline stage are provided in Supplementary Table S1. Malformed JSON responses (∼2.3% of extractions) were re-queried with error feedback; persistent failures were logged and excluded (Supplementary Fig. S2).

Relationship types are constrained to a controlled vocabulary of 51 terms organized into eight categories: transcriptional regulation (e.g., Activates, Represses, Regulates, Modulates), physical interactions (e.g., Binds, Interacts, Recruits), post-translational modifications (e.g., Phosphorylates, Ubiquitinates, Methylates, Acetylates, SUMOylates), protein stability and processing (e.g., Stabilizes, Protects), post-transcriptional regulation (e.g., Targets), transport and localization (e.g., Transports, Translocates), genetic interactions (e.g., IsEpistaticTo, Antagonizes), and other relationships (e.g., Encodes, Synthesizes) (Supplementary Table S2).

Confidence levels are assigned based on linguistic markers: High for definitive experimental demonstration, Medium for indirect evidence, and Low for speculation or predictions. Evidence types classify experimental methodology (Genetic, Biochemical, Transcriptomic, ChIP, Y2H, BiFC, CoIP, Computational, and Degradome).

Gene identifier normalization

Gene names in scientific literature exhibit substantial heterogeneity, and gene symbols are not unique identifiers—one symbol may refer to multiple genes [21,22]. We developed a five-step hierarchical normalization pipeline to convert gene mentions to AGI locus format (Fig. 1; Supplementary Fig. S3):

• Step 1 (Format Check): Verify if the name already conforms to AGI locus ID format (AT#G#####).

• Step 2 (Paper Search for Co-mentioned Locus IDs): Search paper text for locus IDs co-mentioned near gene name occurrences within 50-character windows. This exploits the common practice of authors explicitly including identifiers (e.g., ‘FLC (AT5G10140)’).

• Step 3 (Database Lookup): Query the gene mapping dictionary containing 16,314 genes with alias support (Supplementary Table S3) [23], aggregated from TAIR and curator-validated mappings.

• Step 4 (Paper Search for Candidates): When multiple candidate locus IDs exist, search the paper for mentions of any candidate to resolve ambiguity using paper-specific context.

• Step 5 (LLM Disambiguation): For remaining ambiguous cases, invoke Claude Haiku 4.5 (claude-haiku-4-5-20251001) with UniProt functional annotations [24]. The model evaluates which candidate gene best matches the biological context described in the evidence text.

This paper-first strategy addresses the challenge that TAIR curators face when authors omit AGI identifiers, requiring ‘detective work to infer which identifier should be associated with the new gene symbol’ [23]. Interactions where disambiguation yields Low confidence are automatically flagged for human review.

Web platform implementation

The web application was developed using React 18.2 with TypeScript and Tailwind CSS. Pages include: Home (card-based interaction browser with filtering), Network (Cytoscape.js visualization), My Votes (user voting history), and administrative Review/Suggestions interfaces (for only manager account).

Network visualization uses normalized locus IDs as unique node identifiers, ensuring genes mentioned under different names across papers consolidate into single nodes with correct topology (Fig. 2). Node types include protein-coding genes (circles) and microRNAs (pink diamonds). Edge colors indicate relationship types, while thickness encodes confidence levels where multiple pieces of evidence increase the thickness.

Quality control

GeneReL incorporates a manual curation panel through which interactions with unresolved or ambiguous gene identifiers are routed to expert curators via the administrative Review and Suggestions interfaces, where mappings are confirmed or corrected. In addition, a community voting capability is provided to leverage collective expertise in identifying high-quality versus questionable extractions [25]; Firebase Authentication manages user accounts via Google OAuth 2.0, with vote tracking employing atomic transactions for consistency.

Results

Database composition

The GeneReL database contains 13,710 curated gene-gene interactions from A. thaliana, classified into 51 distinct relationship types (Supplementary Table S2) supported by 9 evidence categories. Source publications span 2010–2024, with the current pipeline having processed 2,051 full-text articles from PubMed Central (Supplementary Fig. S2). Publication years show peak coverage in 2018–2022, accounting for 54% of all extractions, reflecting both literature availability and the PMC open access corpus composition. Transcriptional regulation represents the predominant category (58.6%), dominated by Regulates (including Modulates and Controls; 24.4%), Activates (including Induces, Upregulates, Promotes, Enhances, and Enables; 18.7%), and Represses (including Suppresses, Inhibits, Downregulates, Reduces, Negatively Regulates, Restricts, and Blocks; 15.5%). Physical interactions (Interacts, Binds, and related types) account for 29.6%, while post-translational modifications (phosphorylation, ubiquitination, degradation) comprise 8.0%. Post-transcriptional regulation accounts for 2.8% of extracted interactions (Fig. 3a and Supplementary Table S2). The resulting interaction network is visualized on the web platform’s Network page using Cytoscape.js (Fig. 2). Users can search for one or more genes and toggle between 1-step and 2-step network depth. Edge types can be filtered by relationship category. Clicking nodes displays all interactions for that gene, while clicking edges reveals supporting evidence sentences with links to source publications (Supplementary Fig. S4).

Beyond pairwise relationships, the Network page reconstructs multilayer regulatory subnetworks: by entering multiple query genes and selecting 2-step depth, users can recover multi-gene regulatory modules in which intermediate regulators connect the queried genes. This allows canonical pathways—such as the FLC–FT–CO–SOC1–SVP flowering-time module—to be examined as integrated, multilayer networks rather than as isolated edges, with locus-ID-based node consolidation merging evidence for the same gene across multiple papers into a single node (Fig. 5).

Confidence and evidence distribution

Of the 13,710 extracted relationships, 90.2% (12,368) received high confidence ratings, with Medium confidence interactions comprising 9.5% (1,304) and low confidence relationships 0.3% (38) (Fig. 3b). It is important to clarify what these ratings measure: confidence here reflects the linguistic certainty with which the source text asserts an interaction—definitive experimental demonstration versus qualified, indirect, or speculative phrasing—and the correct parsing of that claim, not an independent judgment of whether the interaction is biologically real. A confidently stated author claim is not, by itself, evidence that the interaction is correct. Accordingly, the High-confidence fraction should be read as a measure of source-text assertiveness and extraction fidelity rather than of biological validity; estimating how many interactions are genuine biological relationships versus extraction or interpretation errors requires expert human assessment, which we begin to address through manual validation and community voting. Evidence type distribution: Biochemical (38.4%), Genetic (32.9%), Transcriptomic (18.2%), ChIP (4.0%), Y2H (3.5%), BiFC (1.2%), Computational (0.8%), CoIP (0.6%), and Degradome (<0.1%) (Fig. 3c).

Gene normalization performance

The five-step pipeline achieves 50.7% successful gene normalization counted per interaction (Supplementary Fig. S5). Step-by-step performance: Step 1 (Format Check) identified 3/13,710 (0.0%) interactions where genes were already in AGI locus format. Steps 2–4 (Dictionary & Paper Search, combining paper search for co-mentioned locus IDs, database lookup with alias support, and paper search for candidate locus IDs when ambiguous) resolved 6,260 (45.7%) of total interactions. miRNA Recognition resolved an additional 238 (1.7%). Step 5 (LLM-Assisted Resolution using UniProt functional annotations) resolved 445 (3.2%). Overall, 6,946/13,710 interactions (50.7%) were successfully normalized; 6,764 (49.3%) were flagged for human review or remained unresolved. Analysis of human review cases (n = 7,223 interactions flagged with ‘needsReview’; unresolved + resolved but need to review) shows that normalization failure (not in common name-to-AGI locus format dictionary) accounts for 92.0%, LLM suggestions needing verification for 6.8%, ambiguous genes with multiple candidates for 0.7%, and other review needed for 0.5%.

Comparison with existing DB; STRING and IntAct

Of 13,710 GeneReL interactions, 6,569 (47.9%) have both genes resolved to AGI locus IDs and are thus matchable against external databases; the remaining interactions have one (28.9%) or neither (23.2%) gene resolved and cannot be meaningfully compared (Fig. 4a). Among these 6,569 matchable interactions, 4,656 (70.9%) have corresponding gene pairs in STRING v12.5 considering interactions at all confidence levels, while 1,913 (29.1%) are unique to GeneReL (Fig. 4b and Supplementary Table S4) [18]. To be explicit about the denominator, this 70.9% is the proportion of GeneReL’s matchable gene pairs (n = 6,569) that are present in the A. thaliana interaction set of STRING v12.5 (any combined score, across the 27,462 A. thaliana proteins). It is therefore an overlap (concordance) measure of GeneReL against STRING, not a recall of the full STRING network. Because STRING’s all-confidence set includes weak, predicted associations, we additionally report the overlap against STRING’s high-confidence interactions (combined score ≥ 700): 44.7% of the matchable GeneReL pairs are present in this more stringent set, indicating that a large fraction of the agreement is supported by high-confidence STRING evidence. We report these overlaps to gauge agreement with an existing resource rather than to claim comprehensive coverage of STRING. GeneReL’s distinct contribution lies in its semantic richness: among the 1,913 interactions unique to GeneReL, regulatory relationships with directionality (Regulates: 516; Activates: 363; Represses: 296) go beyond STRING’s association-based scoring. Post-translational modifications (Phosphorylates: 39; Degrades: 20) and miRNA targeting (Targets: 42) capture regulatory mechanisms absent from STRING.

To further characterize the unique contribution of literature-based extraction, we compared GeneReL interactions against the IntAct molecular interaction database [19], a primary curated resource for experimentally validated protein interactions. Gene pairs were matched bidirectionally (A-B equivalent to B-A; Supplementary Table S4). Of 6,569 matchable interactions, 5,651 (86.0%) represent relationships not present in IntAct, while 918 (14.0%) have corresponding gene pairs (Fig. 4c). We note that this high proportion of relationships absent from IntAct reflects complementary coverage, not validated correctness: absence from IntAct alone does not establish that these interactions are biologically real, and we therefore assessed the precision of the not-in-IntAct subset directly (Section ‘Validation’; Supplementary Table S5). The difference in coverage reflects fundamentally distinct data sources: IntAct primarily captures physical associations from high-throughput screens (Y2H, mass spectrometry, co-immunoprecipitation), while GeneReL extracts semantically rich regulatory relationships from targeted experimental studies described in narrative text. Relationship types with no IntAct equivalent include Regulates (1,352 unique), Interacts (1,180 unique), Activates (1,137 unique), Represses (958 unique), and miRNA Targets (118 unique).

We note that STRING and IntAct are not independent references: STRING’s combined score aggregates evidence channels, including experiments and databases that import curated records from the IMEx consortium, of which IntAct is a principal member. No filtering was applied to remove IntAct-derived records from the STRING set. The two comparisons therefore serve different purposes rather than corroborating one another: concordance against a broad resource aggregating evidence of mixed strength (STRING), and the increment relative to an expert-curated, experimentally supported set (IntAct). Both are matched on gene pairs rather than on typed relationships, so the shared fractions are upper bounds.

Validation

Manual validation of 100 high-confidence extractions, randomly sampled from the database as a whole, was performed. Each extraction was evaluated on three criteria: (i) correct gene identification—whether both gene names appear in the supporting evidence sentence; (ii) appropriate relationship type—whether the evidence sentence contains language consistent with the assigned interaction type; and (iii) evidence relevance—whether the supporting sentence is a well-formed, meaningful text passage containing at least one gene name. Results confirm approximately 90% of extractions accurately reflect source literature claims (Supplementary Table S5). We emphasize that this criterion assesses faithfulness of the extraction to the source sentence (i.e., whether GeneReL correctly captured what the authors stated), and not whether the underlying biological interaction is independently true.

Because the comparison with IntAct centers on the interactions absent from that database, we additionally validated a second random sample of 100 high-confidence interactions drawn specifically from the subset whose gene pairs are not present in IntAct, using the identical three-criteria convention (Supplementary Table S5). Relationship-type assignment remained accurate in 90% of cases (versus 98% in the whole-database sample) and evidence relevance in 100%; the overall accuracy was lower (69% versus 90%) almost entirely because of gene-identifier normalization (gene-identification check 75% versus 91%) rather than relationship errors. Thus, the interactions not present in IntAct are, in the large majority, correctly typed and evidence-supported relationships, with the principal residual weakness being identifier mapping—the limitation addressed by the manual curation panel.

Several cases of validation failure resulted from sentence parsing errors while the gene-to-gene relationship itself was correct (e.g., ‘APRF1 as a Repressor of FLC Expression’ extracted correctly but with truncated evidence text). Some unpredictable parsing errors produce incomplete sentences in the evidence field; however, as the pipeline continuously adds evidence from new publications, multiple supporting sentences accumulate for each interaction edge. This redundancy allows users to cross-validate relationships and identify problematic extractions through the voting mechanism, flagging incorrect interaction cards for future curator correction. Comparison with well-characterized pathways confirms capture of canonical relationships including FLC repression of FT, CO activation of FT, and SVP repression of SOC1.

Network hub analysis reveals coverage biases reflecting publication patterns in the source literature. Well-studied regulatory genes (e.g., FLC, CO, AG, AP1, ABI3) are substantially over-represented: the top 50 most-connected nodes account for 24% of total network degree (Supplementary Table S6). This reflects both genuine biological importance of these regulatory hubs and systematic under-representation of less-studied genes.

Discussion

Principal findings

This work demonstrates successful integration of LLM-based extraction with community curation for biological database construction. The LLM extraction recognizes diverse relationship types across multiple regulatory layers and captures them in a form amenable to browsing, filtering, and network visualization.

Ablation analysis confirms the value of multi-step normalization: format check alone (Step 1) resolved only a bit (∼0.0%) of interactions, while Dictionary & Paper Search (Steps 2–4) contributed the largest gain at 45.7%. miRNA Recognition added 1.7%, and LLM-Assisted Resolution (Step 5) contributed an additional 3.2%, bringing the full pipeline to 50.7% successful normalization. The remaining 49.3% of interactions were flagged for human review, reflecting the inherent complexity of gene nomenclature disambiguation. This performance is consistent with advances in gene normalization systems such as GNorm2 [26], though direct comparison is complicated by different evaluation corpora.

Comparison with IntAct shows that 86.0% of matchable GeneReL interactions are gene pairs not present in that curated resource. We interpret this as evidence of complementary coverage rather than of superiority: absence from IntAct indicates that GeneReL surfaces relationships from narrative text that high-throughput-derived resources do not capture, but it does not by itself establish that those relationships are biologically correct. Importantly, our validation against the supporting evidence sentences (Supplementary Table S5) indicates that the extracted relationships are reliable—90% of sampled high-confidence extractions faithfully reflect their source text, with the relationship type correct in 98% of cases—and a matched validation drawn specifically from the not-in-IntAct subset (Supplementary Table S5) shows that relationship-type accuracy remains high there too (90%), so the principal residual source of error is gene-identifier normalization rather than relationship extraction. The difference in coverage primarily reflects distinct data sources—IntAct aggregates high-throughput physical-association data, whereas GeneReL extracts semantically rich regulatory relationships from descriptions of targeted experiments.

LLM Advantages and Limitations

LLMs demonstrate advantages over earlier text mining methodologies. The extensive pre-training provides robust handling of nomenclature heterogeneity, while contextual understanding enables extraction across linguistic variation without exhaustive rule specification. Zero-shot to few-shot learning proves valuable for specialized domains with limited annotated corpora. However, LLMs occasionally generate plausible but incorrect extractions—hallucinations that cannot occur with rule-based systems. The human-in-the-loop correction strategy addresses this by routing Low-confidence disambiguations to curator review.

Relatedly, the confidence levels reported here (Section ‘‘Confidence and Evidence Distribution’’) should not be over-interpreted: they grade the linguistic certainty of the source claim and the fidelity of its extraction, not the biological truth of the interaction. High confidence therefore indicates that an interaction was asserted definitively and parsed correctly, leaving the orthogonal question of biological validity to expert curators and to the accumulating cross-paper evidence and community votes.

Curation and community validation

GeneReL provides two complementary human-in-the-loop mechanisms. First, a manual curation panel: interactions whose gene identifiers cannot be confidently normalized are automatically flagged and routed to expert curators through the administrative Review interface, where curators confirm or correct the AGI locus mapping and may accept the suggested mappings. This panel is operational and progressively resolves the flagged backlog, so that mapping coverage improves over time rather than remaining fixed at the automated baseline. Second, a community voting capability for distributed quality assessment, following successful crowdsourcing precedents [5,25]; transparent display of positive and negative vote counts provides agreement information while preserving access to all extracted data. We present the voting feature as an infrastructural capability with the potential to scale quality control as the user base grows, rather than as a mechanism with demonstrated large-scale uptake at this early stage.

Limitations

The extraction operates on individual passages independently, lacking mechanisms to integrate evidence across papers or resolve conflicting claims. Gene normalization remains imperfect for recently characterized genes lacking UniProt entries. Importantly, the normalization step currently resolves both interactors to AGI locus IDs for only 50.7% of extracted interactions, leaving 49.3% flagged for human review; this constrains the recall of GeneReL as a fully automated resource and means that, in its present form, the system functions as much as a curation aid that accelerates expert review as a finished, self-contained dataset. Normalization performance for Arabidopsis nomenclature also lags that typically reported for human gene normalization, reflecting sparser alias resources, the prevalence of author-defined symbols, and frequent omission of AGI identifiers in source text. We are actively reducing this backlog through the manual curation panel described in Section ‘‘Curation and Community Validation’’: expert curators review flagged interactions and apply gene-identifier corrections, so that mapping coverage improves over time. Because our evidence-based validation indicates that the underlying relationships are extracted accurately (Supplementary Table S5), this residual normalization gap is a tractable curation problem rather than a limitation of the extracted biology. The current schema captures relationship existence but loses contextual detail, including tissue-specificity and developmental stage dependence. Coverage biases inherent to literature-derived databases mean users should interpret network centrality metrics with awareness of publication patterns rather than assuming uniform genome-wide coverage.

Conclusion

GeneReL demonstrates successful integration of modern language models with community curation for biological database construction. By combining LLM extraction providing scalable processing with distributed validation ensuring quality control, the system addresses longstanding challenges in knowledge resource development. The platform provides the Arabidopsis research community with an accessible, continuously updated repository of regulatory relationships with transparent evidence provenance, advancing efforts to transform biomedical literature into structured knowledge resources.

Author contributions

Joo-Seok Park: Conceptualization, Methodology, Software, Writing—Original Draft, Visualization. Suyeon Ha: Data Curation, Validation. Yejin Lee: Software (Cloud Infrastructure). Yang Jae Kang: Supervision, Project Administration, Funding Acquisition, Writing, Review & Editing.

Conflicts of interest

The authors declare no competing interests.

Funding

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. RS-2024-00336161) and the fund of research for the next generation of academics of Gyeongsang National University in 2024.

Data availability

The GeneReL database and web platform are freely accessible at https://generel.newgenes.me. Data can be downloaded in CSV and JSON formats.

Notes

Biographical Note: Yang Jae Kang is a professor at Gyeongsang National University, Republic of Korea, specializing in plant genomics, gene regulatory networks, and bio-bigdata. His research focuses on integrating computational approaches with biological data curation.

References