Bringing research data artifacts to life through federated curation
Public research data sources such as the International Nucleotide Sequence Database Collaboration (INSDC) [1] are indispensable substrates for secondary research uses including meta-analysis and the assembly of machine-learning training corpora. However, their reuse value is constrained by metadata deposited largely as free text or represented in source-specific structures [2]. In plant transcriptomics metadata, for example, even basic concepts such as a sample being a control can be encoded in dozens of lexical forms, with analogous variation across tissue, treatment, ecotype, developmental stage, growth condition, and sample title.
Submission-side improvements such as checklists, controlled vocabularies, validation rules, and minimal information standards improve incoming records but do not repair decades of legacy deposits. Secondary analyses therefore repeat the same harmonization work without a mechanism to aggregate curation effort. LLM-assisted curation can support first-pass extraction but becomes reusable scientific curation only within defined target structures, domain constraints, provenance capture, and review mechanisms.
We address these problems by separating authoritative deposition from federated curation: we model curation itself as a version-controlled FAIR Digital Object (FDO) that can create or update a curated metadata record as a second, derived FDO. This enables federated curation: contributors can improve curations without modifying the authoritative source record, producing reproducible, reviewable, and citable consensus artifacts. Our approach builds upon the Annotated Research Context (ARC), using its established metadata structures and annotation principles as a general target model rather than introducing a project-specific curation format. It combines structured curation-process metadata with mapping tables represented according to the Simple Standard for Sharing Ontological Mappings (SSSOM) [3], which capture semantic mappings together with predicates, confidence, justification, and provenance. This makes curation work observable and measurable through metrics such as ontology coverage, mapping-confidence distributions, and inter-curator agreement. Existing curation outputs can also be translated into explicit curation artifacts within this framework.
Curated consensus records can be published through established ARC publication pipelines, providing citable outputs for downstream reuse [4]. Representing curated records as ARCs also exposes them through a unified process graph, allowing study, sample, assay, data, and computational steps to be queried with consistent semantics across projects. Since ARCs are representable as ISA-compliant RO-Crates, curated metadata can enter existing FAIR data publication and exchange workflows without requiring a new downstream data model. As these RO-Crates expose experimental, computational, and curation provenance through the same process graph, the approach also creates a growing AI-ready dataset for downstream discovery, reasoning, and model training. It aligns with the ARC concept of ‘immutable yet evolving’ FDOs: each released state or commit is immutable and reproducible, while the FDO evolves through versioned, provenance-preserving updates [5]. We further provide a proof-of-concept implementation of these concepts, termed OverARC.
In accordance with the underlying principles of the ARC, metadata curation is modelled as a process with the authoritative source artifact or record ensemble as input and the curated consensus ARC as output. In our demonstrative use case, the INSDC ensemble comprises the related project, study, sample, experiment, and run records that together describe a sequencing dataset. Rather than treating curation as an external annotation layer, we represent each curatorial decision as part of the provenance of the resulting ARC. The curated record is therefore not only a cleaned version of the original metadata, but also an observable and reproducible account of how that interpretation was produced.
We identify three main branches of curation. The first is structural mapping, where entities and fields in source records are mapped to the ARC process graph. In the INSDC use case, for example, a BioProject usually contributes to the ARC’s administrative metadata, BioSamples map to sample entities, and experiment or run records correspond to steps of the data generation process. By defining these mappings against structural descriptions of both record systems, many baseline curations can be generated automatically while remaining explicit and reviewable. Even this baseline adds value by placing source-specific records into a consistent process-centric representation.
The second branch is semantic mapping, where free-text annotations or existing terms are linked to ontology terms. These mappings are represented using SSSOM tables and include mapping predicates, confidence values, curator identity, provenance, and justification. This supports manual curation as well as automated first-pass extraction, including LLM-assisted suggestions, without losing provenance or reviewability. The SSSOM tables also provide the primary assessment layer for each curated record. For each curation ARC, progress and quality can be described through field-level ontology coverage, mapping-confidence distributions, and the balance of exact, close, broad, or narrow matches. Aggregated across curation ARCs or consensus ARCs, the same metrics characterize the overall curation effort and enable comparison of curation campaigns, including automated or LLM-assisted suggestions evaluated against reviewed mappings.
The third branch is structured extraction, where information encoded in titles, descriptions, protocols, publications, or supplemental sources is decomposed into typed ARC elements that cannot be represented by term-level mappings alone. Examples include extracting treatment parameters, growth conditions, or protocol steps from unstructured descriptions. The source information and any transformations required to create these elements are recorded as part of the curation provenance.
The resulting curation decisions are then compiled into a consensus ARC. In the INSDC use case, curators decide how the INSDC record ensemble should be represented as an experimental process. Administrative metadata usually originate from project or study records; sample-level metadata often describe biological material, growth conditions, or treatments; and experiment or run records describe how sequencing data were produced. Sensible defaults can therefore be provided, while the model remains flexible enough for curator intervention where the original metadata are ambiguous, incomplete, or project-specific.
Curation efforts do not need to depend on undirected crowdsourcing of complete public archives. They may originate from a concrete meta-analysis, a focused collection, a research consortium, or a domain-specific curation campaign. Representing their outputs as compatible, provenance-bearing artifacts allows independently motivated efforts to be reviewed, extended, and combined rather than remaining isolated preprocessing steps.
We position OverARC relative to three lines of work. First, domain-specific metadata harmonization resources such as MetaSRA [6] and ESPERANTO [7] normalize public sequence-archive metadata into more reusable forms. MetaSRA provides automated ontology-based re-annotation of SRA sample metadata, originally focused on human samples, while ESPERANTO offers a semi-supervised harmonization framework for toxicogenomics metadata. The Arabidopsis RNA-seq Database (ARS) [8] similarly combines uniform processing with application-specific classification of public A. thaliana RNA-seq libraries. These approaches demonstrate the value of recovering and harmonizing legacy metadata for concrete downstream applications. OverARC is demonstrated on the same general problem using INSDC records but focuses on representing the curation process and its individual decisions as independently reviewable and extensible research objects.
Second, provenance-aware metadata curation is also relevant beyond sequence archives. OpenCitations Meta [9] aggregates bibliographic metadata from multiple sources and performs deduplication, error correction, and enrichment while tracking the provenance and change history of the resulting entities. This demonstrates the value of retaining curation provenance alongside a curated metadata collection. OpenCitations Meta implements this for bibliographic metadata according to the OpenCitations Data Model, whereas OverARC proposes a generally applicable, process-centric pattern based on the ARC ecosystem. Source evidence, semantic mappings, structured transformations, and resulting research metadata can thereby be represented within the same versioned research object framework.
Third, general metadata-authoring environments such as the CEDAR Workbench [10] and the FAIR Data Station [11] support the creation of structured, ontology-backed metadata using reusable templates or ISA-compatible representations. This aligns with our use of established target structures and annotation principles rather than project-specific metadata formats. Their primary focus, however, is prospective metadata authoring, validation, and submission support, while OverARC focuses on the provenance-bearing interpretation and transformation of existing authoritative source artifacts into curated derived FDOs.
The proposed system also builds on two concepts previously contributed by us to the Integrative Bioinformatics Conference in 2025: fragment-level provenance annotation [12], and the fusion of experimental and computational provenance in RO-Crates via ARCs [13]. OverARC applies these concepts to metadata curation, demonstrated here using INSDC records, by turning harmonization from a private preprocessing step into a shared, observable, reviewable, and citable curation workflow.
Our contribution is fourfold. First, we propose a process-centric model for metadata curation in which the authoritative source record remains unchanged, while curated interpretations are represented as versioned ARCs. INSDC metadata serve as the demonstrative use case rather than the conceptual boundary of the approach. Second, we show how using the established metadata structures and annotation principles of the ARC ecosystem makes curation processes reusable and observable: source information, mappings, transformations, confidence assessments, and resulting metadata elements remain traceable and amendable. Third, we show how this model enables federated curation in the DataHUB framework [4, 14], where independently motivated contributions can be reviewed, attributed, merged, and re-rendered into consensus artifacts. Fourth, we provide a proof-of-concept collection of manually curated A. thaliana ARCs and an initial OverARC implementation for representing source evidence, SSSOM mappings, transformations, and resulting ARC elements without requiring curators to interact directly with the underlying representation.
Together, these contributions turn single-use private harmonization work into a federated curation process that produces reusable, AI-ready FDOs.