NutriEpiDB: a database for dietary compounds with epigenetic targets

1 Introduction

Dietary compounds with epigenetic targets are of growing scientific and translational relevance, as nutrition-mediated modulation of epigenetic mechanisms is increasingly implicated in development, health, and disease. Systematic mapping of food-derived or dietary compounds to their molecular epigenetic targets is essential for advancing the mechanistic understanding of “nutri-epigenomics” and for supporting evidence synthesis, prospective research, and responsible innovation in nutritional science (Anderson et al., 2012; Meral et al., 2024; Müller and Kersten, 2003). To support these needs, wide-ranging, machine-actionable datasets with structured metadata, standardized vocabularies, and clear provenance are required, enabling advanced analytics, responsible AI use, and reproducible hypothesis testing.

The dataset described herein, NutriEpiDB, organizes curated evidence linking dietary compounds to specific epigenetic targets and mechanisms. It achieves this by integrating compound–target–mechanism associations extracted from the biomedical literature with multi-level annotation of chemical identity, food source, epigenetic action, and supporting references. The dataset was constructed through systematic retrieval and triage of the primary research literature indexed in PubMed and Web of Science through August 2026, using controlled vocabulary queries to identify articles reporting dietary, beverage, or food-derived compounds with reported effects on epigenetic pathways. Articles lacking a digital object identifier (DOI), title, or abstract, not written in English, or not primary research were excluded. Computational pipelines employing large language model (LLM)-assisted extraction, semantic harmonization, and chemical structure normalization were then used to identify, deduplicate, and link compounds, food sources, and molecular epigenetic targets.

The resulting data resource encompasses structured records describing compound–food–epigenetic mechanism relationships, together with molecular descriptors (such as PubChem compound identification numbers (CIDs), simplified molecular-input line-entry system (SMILES), and InChIKeys), standardized dietary source labels, enriched protein/gene metadata for human and mouse orthologs, and quantified protein–compound binding data where available. All chemical and target entities are normalized and mapped to public database identifiers, supporting cross-database integration. Schema.org and PROV-O-compatible metadata are included at the file and variable levels to facilitate interoperability and traceability. The underlying literature dataset is accompanied by structured metadata on publication type, year, and article-level provenance, further supporting responsible reuse and contextual analysis.

To maximize semantic and technical interoperability, dietary sources are normalized through a controlled semantic clustering process that resolves variable naming and synonymy. Compound identities are reconciled and enriched using the PubChem database, including systematic nomenclature, chemical structure formats, and curated synonyms. Epigenetic targets are annotated with cross-references to the EpiFactors database (Marakulina et al., 2022), UniProt, and Human Genome Nomenclature Committee (HGNC), as relevant. For compounds with available quantitative binding measurements (e.g., IC₅₀, K_i, EC₅₀, or K_d), assay data and binding context from BindingDB are integrated and harmonized via shared reactant and monomer identifiers, with organism and protein chain metadata included. All records are structured to support AI-readiness, enabling direct ingestion by machine learning workflows and ensuring transparency regarding variable definitions, missingness, and unit standards.

This data article characterizes the collection, curation, and structure of the NutriEpiDB resource, following FAIR² principles to ensure findability, accessibility, interoperability, and reusability, with explicit support for AI-readiness and responsible AI applications. Special attention is given to clear variable specification, ontology mapping, and reporting of data transformation protocols, supporting reproducibility and traceable evidence synthesis. NutriEpiDB is intended as a resource for researchers in epigenetics, nutrition, chemical biology, and bioinformatics, as well as for the development and evaluation of predictive and explainable models in nutriepigenomics and responsible AI-based health research.

2 Methods summary

2.1 Study design

The NutriEpiDB dataset was established to systematically catalog dietary compounds with known impacts on epigenetic regulation. To achieve this, study objectives and methodological scope were strictly defined, limiting the dataset to food-derived compounds reported in peer-reviewed literature and supported by evidence of effects on mechanisms such as DNA methylation, histone modification, or chromatin regulation (Morrison and Thakur, 2021). The design mandated the application of inclusion and exclusion criteria to literature sources, with only articles explicitly referencing the influence of dietary compounds on epigenetic pathways being eligible. Publications lacking critical bibliographic data (DOI, title, or abstract), along with review articles and meta-analyses, were programmatically excluded to maintain granularity and relevance.

Further refinement was achieved by restricting eligible publication types to primary research reports, including journal articles, comparative studies, case reports, and clinical trials through randomization or phased intervention. Textual sources were limited to the English language to ensure consistency in data processing and model extraction. Temporal coverage extended to all applicable literature indexed up to August 2026, leveraging PubMed (Kim et al., 2024) and Web of Science for their comprehensive collection of biomedical and chemical research. Data extraction targeted explicit metadata for each publication, encompassing article identifiers, compound and dietary source annotations, and descriptions of epigenetic mechanisms. The identification of unique entities was supported by controlled vocabularies and standardized field definitions, laying a foundation for reproducibility and structured integration at all stages of the dataset lifecycle.

2.2 Literature search and article retrieval

A systematic literature search supported the precise identification of relevant articles forming the core of NutriEpiDB. Comprehensive queries were constructed to encompass all combinations of dietary compounds and epigenetic mechanisms, with terms strategically selected to maximize the capture of eligible records across the source databases. Search strings included controlled vocabulary for foods, nutrients, drinks, supplements, and a wide range of molecular descriptors, intersected with recognized terms for epigenetic processes.

Implementation began by querying PubMed using the Bio.Entrez interface, automating the retrieval of all records available by the cutoff date and actively excluding non-conforming articles. Review literature, systematic reviews, meta-analyses, and entries missing essential metadata were filtered both at the query stage and via post-retrieval programmatic checks. Further selection excluded non-target publication types and non-English texts, preserving fidelity to the established inclusion parameters.

The search string used for both platforms was as follows: (food* OR diet* OR drink* OR supplement* OR nutrition* OR beverage* OR Diet* OR “Dietary Supplements” OR natural) AND (compound* OR nutraceutical* OR nutrient* OR molecule* OR plant-derived) AND (epigenetic* OR nutriepigenetic* OR “DNA methylation” OR histone* OR chromatin OR Epigenetic* OR “DNA Methylation” OR Histone*) NOT (review[pt] OR “systematic review”[pt] OR “meta-analysis”[pt]).

For each article passing these criteria, structured metadata, comprising the title, abstract, DOI, publication date, and PubMed ID, was systematically extracted for downstream integration. A parallel search using matching queries in Web of Science contributed any additional articles, which were then deduplicated against the PubMed corpus to prevent redundancy. The outcome was a precisely curated list of primary studies fully annotated for subsequent compound and mechanism extraction, ensuring a transparent and reproducible literature foundation.

2.3 Compound identification via language model extraction

To support precise identification of dietary compounds exerting epigenetic effects, a controlled AI-based extraction workflow was employed (Elvas et al., 2025). The LangChain Python framework was utilized to generate article-specific prompts that strictly specified task boundaries and data requirements. Prompts were submitted to the claude-haiku-4-5-20251001 model via the Claude API, with the temperature parameter set to zero to control output variability and minimize model hallucinations.

The extraction protocol instructed the language model to assume the analytical stance of a medicinal chemist, annotating all dietary, beverage, or food-derived compounds explicitly mentioned as modulators of DNA methylation, histone modification, or non-coding RNA pathways. The model used explicit tabular field definitions to ensure consistent variable extraction and unambiguous linkage between compounds and their dietary origins. To retain specificity, rules were embedded in the prompt to exclude group designations, such as “vitamins” or “polyphenols,” and focus on identifiable molecular entities. The resulting output comprised structured pairings of compound names with food or drink sources, annotated by underlying articles, thus providing reproducible entity mapping for integrated database construction.

2.4 Dietary source data homogenization

Recognizing the variability in how dietary sources are represented in scientific literature, the extracted terms were subjected to a semantic standardization pipeline. All unique dietary source strings identified in the prior step were first compiled into a comprehensive list. Each term was then embedded as a dense vector using the all-MiniLM-L6-v2 sentence-transformer model, enabled by the sentence_transformers package for Python, to provide a quantitative semantic representation.

Agglomerative hierarchical clustering was performed over these vectors using cosine distance, applying a threshold of 0.3 to cluster lexically and semantically similar terms. From each cluster, the shortest included term was chosen as the canonical label, ensuring brevity and minimizing ambiguity in subsequent analyses. All source terms from the literature were thus mapped to standardized forms, yielding a single categorical dietary source variable with robust terminology control and interoperability across database operations.

2.5 Compound nomenclature standardization and descriptor enrichment

To ensure exact compound identification and interoperability, all extracted dietary compound names underwent programmatic standardization and annotation. PubChemPy, a Python package interfacing with the PubChem database, was used to map each compound name to its corresponding records (Cock et al., 2009). Multiple identifiers and descriptors were retrieved for each successful match, including the PubChem Compound Identifier (CID), International Union of Pure and Applied Chemistry (IUPAC) name, hashed International Chemical Identifier (InChIKey), SMILES representation, and available synonyms. The PubChem compound page title was utilized as the representative name to support uniform nomenclature.

Where a query name returned multiple matches in PubChem, the first valid CID was adopted as the reference to facilitate downstream harmonization. Records missing key descriptors, such as a valid CID, annotated mechanism, or compound name, were systematically excluded to maintain the accuracy and utility of the dataset. Molecular representations, specifically SMILES strings, were further processed using RDKit: salts were removed (retaining only the primary fragment), formal charges were neutralized, tautomers were canonicalized, and all structures were translated into a single canonical SMILES form. This multi-stage validation and cleaning protocol ensured all chemical descriptors were consistent, interoperable, and reproducible for future cheminformatics or biological analyses.

2.6 Binding data integration

To provide mechanistic context for dietary compounds, protein binding interaction data were integrated from external resources. BindingDB (Liu et al., 2024) served as the principal source for binding affinities with protein targets, and comprehensive retrieval was conducted for all standardized compounds in previous steps. Filtering steps ensured that only records detailing interactions with Homo sapiens proteins were retained and that each interaction included a numeric measure of binding activity, namely, K i, I C 50, K d, or E C 50 values, in nanomolar units, along with associated chemical and assay metadata.

Target proteins listed in BindingDB records were systematically cross-referenced with functional annotations available in the EpiFactors database. Only those proteins with defined roles in epigenetic regulation were preserved for integration. Each binding record was further annotated with descriptive target functions, based on EpiFactors categories, such as molecular “writer,” “reader,” or “eraser” of epigenetic marks. This enabled robust mechanistic categorization and supported the assignment of dietary compound interactions to specific epigenetic pathways, providing depth and relevance for downstream analytical applications.

2.7 Final dataset assembly and outputs

Upon the completion of all extraction, cleaning, and annotation steps, curated entries describing compound characteristics, food and beverage sources, mechanistic actions, and binding activities were aggregated into the NutriEpiDB structure. Unique entities, including compounds, dietary sources, and epigenetic targets, were enumerated, and field-level definitions were consistently enforced to facilitate reproducibility and transparency.

Data from individual workflow components, including article metadata, standardized compounds, dietary sources, mechanism annotations, and binding records, were merged by controlled identifiers into relationally coherent datasets. Outputs were framed with explicit field definitions, statistical summaries, and controlled vocabularies to enable reliable independent reuse. The final records met established criteria for quality and completeness, and the dataset was released for public dissemination with comprehensive metadata documentation, thereby supporting transparent and reproducible research in the domain of dietary epigenetic modulation.

3 Data overview

The NutriEpiDB dataset catalogs 653 distinct dietary compounds, including their chemical descriptors, food or beverage sources, and documented influences on epigenetic mechanisms, as reported in the scientific literature through August 2026. Each entry systematically links compounds to their corresponding PubChem identifiers, standardized chemical structures (SMILES, InChIKey, IUPAC name), canonical nomenclature, and curated synonym lists. Compound entries are further annotated with one or more dietary sources, harmonized through semantic clustering to support consistent terminology. Associations between compounds and epigenetic mechanisms are explicitly documented, specifying mechanistic categories such as DNA methylation, histone modification, chromatin remodeling, and non-coding RNA regulation, along with supporting reference DOIs. Integration with curated protein target data enables direct mapping between dietary molecules and protein effectors involved in epigenetic regulation.

The dataset establishes cross-references to human and mouse protein targets using UniProt accession numbers, HGNC and mouse genome informatics (MGI) gene identifiers, and links to molecular function annotations following EpiFactors curation. Detailed binding data for 101 unique compound–target pairs are included (sourced from BindingDB), covering affinity measurements such as IC₅₀, K_i, K_d, and EC₅₀ values, along with experimental context (organism, assay conditions) and source metadata (PMID, DOI, curation authority).

A total of 1,963 primary research articles are indexed, each providing author-supplied abstracts, publication years, metadata fields (DOI, PubMed ID, publication type), and controlled vocabularies for reproducible downstream integration. Food and beverage source data (585 unique, harmonized terms) link back to both compound records and supporting literature, facilitating traceability. Comprehensive field-level definitions and linked identifiers across all tables support transparent data integration, relational consistency, and machine actionability. The dataset is fully exportable in tabular format and aligned to FAIR and AI-Readiness (AIR) standards, with all records annotated to maximize findability, accessibility, and reuse within the epigenetics and dietary research communities.

4 Quantitative summary of the dataset

4.1 Total number of records

The integrated NutriEpiDB dataset contains 1,963 records in the primary article metadata table, corresponding to distinct publications included according to defined eligibility criteria. The compound data table documents 653 entries, each representing a unique instance of a dietary compound mapped to one or more food or beverage sources, mechanistic annotations, and chemical descriptors. Mechanistic associations between compounds and epigenetic mechanisms are detailed in 2,438 records, where each entry links a compound to a specific epigenetic process with reference support and accompanying identifiers. The binding affinity data encompasses 2,877 records, quantifying compound–target binding measurements across human epigenetic proteins. The EpiFactors epigenetic target table represents 80 annotated protein factors, each associated with molecular function, modification, and cross-references to authoritative gene and protein databases. The diet/food source entity table catalogs 585 unique standardized food or dietary source terms, each linked to supporting literature and constituent compound profiles.

4.2 Coverage

The dataset spans a broad array of molecular, biological, and literature-derived categories. The dataset includes 653 distinct compound entries, each explicitly mapped to PubChem CIDs. These compounds link to 373 unique dietary source groups in the compound table, resolved to 585 canonical food or dietary source variables. Mechanistic associations encompass 43 distinct epigenetic mechanisms, from specific processes such as “DNA_methylation” (320 records) and “HDAC_inhibition” (301 records) to broader categories including “epigenetic_modulation” (383 records). Across all datasets, linkage is maintained via foreign keys, including PubChem CIDs for compounds and UniProt accessions for protein targets. The protein target dimension includes 80 EpiFactors entries, each mapped to the Human Genome Nomenclature Committee (HGNC), National Center for Biotechnology Information (NCBI) Gene, and UniProt accessions.

4.3 Temporal coverage

The publication records extend from 1991 to 2026, providing 36 years of temporal representation. Yearly publication counts are highest in recent years, with peaks observed in 2025 (211 articles), followed by 2022 (153 articles), 2024 (150 articles), and 2023 (143 articles). The mean publication year is 2018.5 (SD = 5.44), with a consistent increase in record counts corresponding to growing research interest over the period surveyed.

4.4 Attribute diversity

The dataset offers substantial attribute diversity. Article records include 10 primary metadata variables (e.g., title, abstract, DOI, publication year, PubMed ID, and publication type). The compound data table encodes 15 variables per entry, including chemical structure (SMILES, InChIKey, IUPAC name), mechanistic annotations, dietary source, reference DOIs, and chemical identifiers. Mechanism associations are annotated with compound name, PubChem CID, mechanistic term, supporting DOIs, and reference counts. Protein targets carry 21 attributes, including gene/protein descriptor fields, modification type, specific target, HGNC and NCBI identifiers, and UniProt links. Quantitative binding data include 18 variables such as K_i, IC₅₀, K_d, EC₅₀, measurement conditions (temperature, pH), protein and compound identifiers, and reference source.

4.5 Quantitative attributes

Key numerical and categorical attributes include:

• PubChem Compound Identifier (CID): 653 unique values; range: 72 to 168,013,399.

• Dietary source entries: 373 unique groupings in compound data and 585 unique canonical forms in food sources.

• Mechanism categories: 43 distinct mechanism terms; most frequent: “gene_expression_regulation” (16.4%), “epigenetic_modulation” (15.7%), “histone_acetylation” (13.5%), “DNA_methylation” (13.1%), and “HDAC_inhibition” (12.3%).

• Protein modifications: 15 unique types, with “Histone acetylation” (25.0%) and “Histone methylation” (16.3%) most frequent among target factors.

• Binding affinity measures: IC₅₀ (nM): 2,607 measurements; mean = 21,491.7, SD = 317,629, min = 0.0, max = 12,000,000.

• K_i (nM): 122 values; mean = 7,352.9, SD = 42,547.1, min = 0.1, max = 450,000.

• K_d (nM): 64 entries; mean = 26,664.3, SD = 73,631.0, min = 0.065, max = 445,000.

• EC₅₀ (nM): 85 records; mean = 120,138.5, SD = 652,427.9, min = 0.01, max = 5,780,000.

• Temperature (°C): 76 values, mean = 26.7, SD = 8.84, range 4–37.

• pH: 83 values, mean = 7.73, SD = 0.27, range 7.4–8.1.

4.6 Data completeness

Data completeness varies by record type and attribute. In the article metadata table, all 1,963 records contain a title, abstract, DOI, and publication year, with PubMed ID available for 1,882 records (4.1% missingness for PMID). In the compound data table, all 653 entries have complete coverage for compound name, PubChem CID, and SMILES, with one missing IUPAC name (<0.2% missingness for iupac_name). Mechanism and protein target linkage are complete for all entries. In the compound binding dataset, IC₅₀ values are present for 90.6% of records, while K_i and K_d are less frequently represented (4.2% and 2.2%, respectively); temperature and pH are recorded for less than 3% of entries. Food source and compound name variables are fully populated. At the dataset level, missingness is generally low for identifiers and essential categorical attributes but more common for some experimental condition annotations within the binding dataset.

4.7 Dataset availability

The dataset is FAIR²-certified and publicly available under the Open Data Commons Attribution License (ODC-By v1.0), permitting unrestricted reuse with appropriate attribution. Access is provided via two coordinated components: (1) the FAIR² Data Portal, offering interactive dashboards for visual exploration and quality assessment; and (2) the complete FAIR² Data Package, containing raw and processed data, structured metadata, transformation scripts, and comprehensive methodological documentation. The dataset is indexed in the SenScience Data Portal, archived with a persistent identifier, and linked to relevant data ecosystems. The metadata and methods files include detailed descriptions of experimental procedures and data processing steps. All components are accessible via https://sen.science/doi/10.71728/senscience.sx3s-9110. Licensing terms and usage conditions are specified in the dataset documentation.

4.8 FAIR² compliance introduction

The FAIR² Compliance Certification Table provides a structured, high-level assessment of the dataset’s alignment with findability, accessibility, interoperability, and reusability (FAIR) principles, extended by AI-Readiness (AIR) and Responsible AI (RAI) criteria. This table systematically documents evidence of compliance for each principle and sub-principle, referencing dataset documentation, metadata standards, persistent identifiers, and versioning and access controls. It includes indicators for machine actionability, traceability, and ethical considerations embedded within the dataset. By transparently aligning dataset attributes and curation procedures with established FAIR² standards, the certification table supports compliance verification by external reviewers and end users. This resource enables researchers to assess dataset suitability for integration into advanced analytics, machine learning workflows, and ethically governed AI applications, while facilitating reuse and reproducibility in diverse scientific domains (Table 1).

TABLE 1. FAIR² certification.

4.9 Dataset visualizations

Figure 1 shows the distribution of articles by year. The years 2022, 2024, and 2025 are highlighted in the figure because they have the highest number of articles.

The annual distribution of curated articles documenting dietary compound–epigenetic relationships from 1981 to 2026 demonstrates a marked increase in research activity over time. The number of articles per year remained low and relatively stable through the 1980s and 1990s, with a pronounced rise beginning in the mid-to-late 2000s. This upward trend continues, with publication counts remaining higher through the 2010s. The most recent years, particularly from 2021 to 2025, exhibit the highest annual counts in the dataset, with 2025 representing the peak at 211 articles. Notable secondary peaks are observed in 2022 (153 articles) and 2024 (150 articles). The period from 2021 onward is visually highlighted, indicating a phase of intensified curation and dataset recency. These data reveal both long-term growth and recent surges in the scientific literature addressing dietary compound–epigenetic interactions.

The distribution of publication types among articles in the NutriEpiDB corpus is highly skewed toward original research articles, specifically journal articles, which constitute the vast majority of included records, as shown in Figure 2. Comparative studies, clinical trials, randomized controlled trials, clinical trial phase II studies, and case reports occur far less frequently. This distribution highlights the predominance of primary research outputs, with limited inclusion of experimental and observational study designs directly relevant to nutritional epigenetics.

Figure 3 shows the 20 dietary compounds most frequently cited in the NutriEpiDB literature, highlighting those linked to experimentally validated epigenetic mechanisms. Citation counts reflect research focus and identify primary compounds driving investigation in nutritional epigenetics, with mechanisms distinguished by color.

Figure 3 shows that among dietary compounds cataloged in the NutriEpiDB dataset, folate, choline, L-methionine, resveratrol, and vitamin B12 exhibit the highest citation frequencies, each linked to more than 100 unique literature references. These compounds represent primary targets of research in nutritional epigenetics. Other frequently cited compounds include curcumin, butyrate, bisphenol A, betaine, and sulforaphane. The identified compounds span both micronutrients (e.g., folate, vitamin B6, and vitamin B12), bioactive phytochemicals (e.g., resveratrol, quercetin, curcumin, genistein, and epigallocatechin gallate), and synthetic or environmental agents (e.g., bisphenol A, sodium butyrate, and valproic acid). The distribution strongly emphasizes methyl donors, B vitamins, and polyphenolic compounds, reflecting their established mechanistic links to epigenetic regulation, including DNA methylation, histone modification, and histone deacetylase (HDAC) inhibition. These compounds constitute the most intensively studied dietary agents in the context of experimentally validated epigenetic effects.

Figure 4 highlights foods with exceptional chemical diversity, emphasizing priority candidates for nutritional and functional food research. The visualization demonstrates that certain dietary sources, such as tea, grapes, turmeric, and soy, exhibit the highest number of unique epigenetically active compounds among the food items analyzed, with tea containing the greatest diversity (46 compounds). Multiple plant-based sources (e.g., citrus, olive, broccoli, and Salvia miltiorrhiza) and selected animal-derived foods (beef, milk, egg, and fish oil) are also represented among the top categories. Several micronutrients and supplements (folate, methionine, and fiber) rank within the top 20 sources. The wide range of unique compound counts across food sources highlights pronounced variability in chemical diversity relevant to epigenetic modulation, with plant-derived products generally displaying higher unique compound tallies than animal-derived or supplement categories. Generic food groups are excluded from the visualization and are instead aggregated in a separate annotation panel, as described in the methods, to emphasize specific food items.

Figure 5 shows the distribution of epigenetic modification types by molecular target class in NutriEpiDB. Major modification classes such as histone acetylation and methylation are well-represented among histone targets, while DNA and RNA modifications are less frequent, revealing both coverage and gaps across target molecule categories. Low-frequency associations are labeled to highlight underrepresented areas suitable for future research.

The distribution of epigenetic modifications by molecular target class in NutriEpiDB reveals that histone modifications, particularly acetylation, methylation, and phosphorylation, are predominantly associated with histone targets, with histone acetylation occurring at the highest frequency. DNA-related modifications, such as methylation and demethylation, are less represented and are distributed across both DNA and mixed DNA/histone target classes. Modifications involving RNA, such as mRNA editing, RNA demethylation, RNA methylation, and alternative splicing, exhibit low record counts and are primarily associated with RNA or mixed DNA/RNA targets. Protein-targeted modifications, such as histone phosphorylation, are infrequent. Low-frequency modification–target combinations are annotated, indicating underrepresented areas and highlighting potential gaps for further investigation, particularly among non-histone targets and less prevalent modification types.

Figure 6 shows the distribution of unique epigenetic factors by molecular function, revealing the diversity and prevalence of writer, eraser, reader, and cofactor roles in the dataset. Exact factor counts highlight functional biases and coverage across epigenetic mechanisms. The distribution of unique epigenetic factors by molecular function is highly skewed, with the largest number observed in the “Histone modification write” (29) and “Histone modification erase” (20) categories. This predominance indicates that writer and eraser functions are the most prevalent roles among epigenetic targets in the dataset. Other functions such as “Histone modification write cofactor” (7), “RNA modification” (6), and “Chromatin remodeling” (5) show intermediate counts, while the remaining categories, including “TF,” “Chromatin remodeling cofactor,” and various cofactors or readers, are represented by few factors (ranging from one to four each). The low representation of “Polycomb group (PcG) protein,” “Histone modification cofactor,” and “DNA modification cofactor” (all with a count of one or two) highlights functional biases and possible underrepresentation of these categories. The results reflect a dataset landscape with a clear emphasis on histone-modifying enzymes, particularly those with writer and eraser activities, suggesting potential biases in research coverage or underlying biological diversity of these functional classes.

5 Discussion

5.1 Value and applications of the dataset

Several limitations of the dataset should be noted to contextualize its reuse and support responsible analysis. First, the dataset is constructed from curated sources such as BindingDB, ChEMBL, PubChem, EpiFactors, and published literature. The integration of heterogeneous sources may introduce inconsistencies in nomenclature, annotation granularity, or evidence quality across data fields. Second, certain fields exhibit high degrees of missingness or use placeholder values (e.g., “#”), which may affect automated workflows and the selection of data subsets; for example, not all compound–target pairs have complete binding affinity data (e.g., IC₅₀, K_i, K_d, and EC₅₀), and some protein and compound records lack full ortholog mapping or structural details.

A key limitation is that the dataset is built primarily from article titles and abstracts, which often do not report detailed population descriptors such as ethnicity, genetic background, or country of recruitment. As a result, systematic contextualization by population characteristics is not yet possible. In addition, the representation of epigenetic mechanisms depends heavily on the scope and focus of available source databases and publications, potentially introducing selection bias toward well-studied targets (e.g., histone deacetylases) and model organisms, primarily Homo sapiens and Mus musculus. The data are also limited to dietary compounds with established PubChem CIDs, which may underrepresent metabolites without well-defined structures or less-studied phytonutrients.

The long-term maintenance strategy for the database will include semi-automated updates, whereby the PubMed and Web of Science search pipeline will be rerun at regular intervals to identify newly published articles, followed by LLM-assisted extraction and PubChem/EpiFactors harmonization under manual curation review. In addition, the database will continue to integrate external resources, including BindingDB, EpiFactors, PubChem, and UniProt, to preserve interoperability as these repositories evolve.

5.2 Limitations of the dataset

Several limitations should be noted when reusing the dataset. It is compiled from curated sources including BindingDB, ChEMBL, PubChem, EpiFactors, and published literature, so heterogeneous reporting may introduce inconsistencies in nomenclature, annotation granularity, and evidence quality. Missing values and placeholder entries may also affect automated analyses, and some compound–target pairs lack complete affinity, ortholog, or structural information.

Population descriptors are often unavailable because the database is built mainly from titles and abstracts, limiting contextualization by ethnicity, genetic background, or country of recruitment. The dataset may also reflect selection and reporting bias toward well-studied targets, model organisms, and positive findings, while most binding data are derived from in vitro assays.

In addition, the compounds and food sources listed in NutriEpiDB should not be interpreted as dietary or therapeutic recommendations; safety, tolerability, allergic risk, food intolerance, medication interactions, and dose-dependent effects must be assessed individually. Although the database includes persistent identifiers and provenance metadata, some external links may become outdated as source resources evolve.

5.3 Opportunities for future research

This dataset provides structured resources for investigating the relationships between dietary compounds, epigenetic mechanisms, and molecular targets. Integration of compound–mechanism, binding affinity, and dietary source data enables interrogation of compound activities relevant to epigenetic regulation. Future developments could include the extension of compound coverage to reflect recently reported dietary components, expansion of curated epigenetic targets to incorporate emerging modifiers such as non-coding RNAs and readers, and enrichment of mechanistic annotations with more granular epigenomic and transcriptomic context. Incorporating time-resolved or dose-response data linked to experimental conditions would further facilitate mechanistic modeling and benchmarking of AI-driven inference methods. Improvements in annotation of food matrix components, standardization of compound names and identifiers, and mapping to ontologies would enhance interoperability and reusability for nutrition, chemical biology, and toxicology research. Integration with additional external reference resources, such as large chemical-protein interaction databases or multi-omics platforms, may broaden application to systems-level studies. Adoption of expanded provenance metadata, including detailed reporting of experimental protocols and curation decisions, could further strengthen reproducibility and support responsible AI requirements. Overall, the dataset’s structure and data lineage facilitate adaptation to evolving standards in the epigenetics and nutrition fields, supporting cross-domain analyses, data harmonization, and machine-actionable research workflows.

6 Conclusion

The NutriEpiDB dataset establishes a rigorously structured and systematically documented resource that enables reproducible exploration of dietary compounds and their reported epigenetic targets. Each entity, including articles, compounds, dietary sources, mechanisms, protein targets, and binding interactions, has been defined according to explicit field-level standards, with comprehensive provenance tracked through a detailed linkage to the underlying methodology. The workflow integrates literature curation, AI-assisted compound extraction, nomenclature harmonization, descriptor enrichment, protein binding annotation, and domain-specific data normalization, resulting in a resource characterized by transparency in field origin and processing.

All variables are mapped to their corresponding data generation steps, supporting the full traceability required for accurate interpretation, rigorous analysis, and robust reuse. Adherence to FAIR² principles, including AI-readiness and machine actionability via MLCommons Croissant and interoperable metadata schemes, ensures that the dataset is not only findable and accessible but also technically aligned for integration into automated or algorithmic analyses. Extensive metadata, standardized identifiers, and harmonized terminologies support consistent cross-linking to external databases, maximizing interoperability and data cohesion.

Quality control is maintained throughout the data lifecycle: exclusion criteria, identifier resolutions, structural standardization, and quantitative binding measures are applied according to defined protocols, ensuring that all records are eligible for reproducible analysis. Ambiguities in nomenclature, entity associations, and mechanistic annotations are minimized via enforced vocabulary controls and normalization strategies. The dataset, therefore, provides an actionable foundation for the systematic study of diet-derived molecular influences on epigenetic mechanisms.

By providing precise documentation of variable definitions, processing workflows, and data structures, NutriEpiDB is positioned to facilitate a broad range of scientific inquiries, from hypothesis-driven investigation to machine learning model development. The resource is immediately applicable to questions of compound–target specificity, dietary source impact, and mechanism mapping, and supports scalable analytical workflows requiring structured, high-integrity data. The combination of domain-specific detail, methodological rigor, and machine-actionable standards ensures that NutriEpiDB can be responsibly extended, validated, and repurposed by diverse research communities. Through its transparent construction and explicit linkage to methods, the dataset advances reproducible, responsible data use in computational nutrition, epigenetics, and bioinformatics.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. This document and the accompanying data package were prepared by the author(s) with assistance from the SENSCIENCE FAIR² Data Publishing platform (v0.11β), which uses generative AI technology.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Giuseppe Felice Mangiatordi, National Research Council (CNR), Italy

Reviewed by: Ruby Srivastava, Indian Institute of Technology Bombay, India

Mirko Treccani, University of Parma, Italy