Ontology Development Kit (ODK)

• Ontology Development Kit (ODK) is a comprehensive toolkit designed for building, maintaining, and distributing ontologies with reproducibility and quality control.
• It integrates Docker-packaged tools and standardized Makefile workflows to automate release preparation, quality control, and dependency management.
• ODK supports CI/CD integration, LLM-assisted curation, and adherence to FAIR metadata standards, driving efficiency in biomedical and AI ontology projects.

The Ontology Development Kit (ODK) is a comprehensive, standardized toolkit for building, maintaining, and distributing ontologies. It packages all required tools and workflows within a Docker image, facilitating reproducible, automated, and quality-controlled ontology engineering. ODK is widely adopted in the biomedical domain and has demonstrated versatility in the construction and evolution of AI ontologies, including large-scale, LLM-assisted projects such as the Artificial Intelligence Ontology (AIO) (Joachimiak et al., 2024, Matentzoglu et al., 2022).

1. Architecture and Component Overview

ODK integrates two principal architectural layers: a Docker-based collection of ontology and software engineering tools, and a set of orchestrated workflows embedded as Makefiles and wrapper scripts. The Docker images (e.g., obolibrary/odklite and obolibrary/odkfull) encapsulate tools such as ROBOT, OWLTools, DOSDP-tools, ELK, and additional utilities for validation, as well as standard Unix tools and Python toolkits. All dependencies are version-pinned for full reproducibility. Workflows are exposed primarily through Makefile targets, which define the dependency structure for building, testing, and releasing ontology artifacts. Repository templates are distributed with standardized directory structures (e.g., src/, imports/, release/). This layered design abstracts platform- and user-specific differences, enabling one-command execution of complex ontology lifecycle operations.

2. Standardized Workflows

ODK establishes a set of executable, standardized workflows covering all phases of ontology management. These workflows are directly accessible either via Makefile targets or via Docker-invoked commands:

• Release Preparation and Versioning: Artifact production (OWL, OBO, JSON, TTL) through automated merging of ontology source, imports, and templates. The process includes logical classification, version IRI assignment, multiple variant serializations (base, full, simplified), and comprehensive quality control. Semantic versioning is enforced for reproducible release management.
• Continuous Quality Control and Validation: Automated via make test or odk make quality-control, employing ROBOT for syntactic and semantic checks (e.g., missing labels, licence verification), SPARQL anti-pattern scripts, and OWL reasoner validation for logical coherence. Execution can be configured to skip specific checks per project requirements.
• Dependency Management: Explicitly managed via import lists (e.g., imports/import_list.txt). ODK handles extraction (SLME method), modularization (base/full imports), and automated updating of external ontology terms as part of the release or as a dedicated workflow.

Pseudocode and algorithmic representations of these processes are included in the ODK documentation to illustrate the formal workflow logic.

3. Metadata, FAIR Standards, and Interoperability

ODK enforces ontology-level metadata driven by explicit configuration files (e.g., metadata/ontology.yml) and prescribed annotation schema. Required fields include ontology IRI, version IRI, title, description, license, contributors (ORCIDs), created and modified dates, and declared imports. ODK’s metadata validation uses SPARQL checks for completeness and consistency. This strict metadata regime advances semantic interoperability and compliance with FAIR (Findable, Accessible, Interoperable, Reusable) principles across domain ontologies.

A summary table of standard metadata fields enforced by ODK:

Field	OWL Property	Cardinality
Ontology IRI	rdf:about / owl:Ontology	1
Version IRI	owl:versionIRI	1
Title	dc:title, dcterms:title	1
Description	dc:description	1
License	dcterms:license	1
Contributors	dcterms:contributor	*
Created Date	dcterms:created	1
Modified Date	dcterms:modified	1
Imports	owl:imports	*

This practice is directly inherited in downstream projects such as AIO (Joachimiak et al., 2024, Matentzoglu et al., 2022).

4. Integration with Curatorial and Computational Pipelines

ODK is designed for extensibility and seamless integration with both manual and automated curation paradigms. In the development of AIO, ODK orchestrates a modular, branch-specific template system (TSV files per domain branch) to enable parallelizable content development. These templates serve as the canonical source of ontology terms and are dynamically updated via both domain experts and LLM prompts.

Human-in-the-loop curation is secured through manual review checkpoints in the update pipeline, particularly when leveraging LLMs for branch extension or term suggestion (e.g., via Python scripts interfacing with GPT-4). The process is fully automatable within the ODK+GitHub Actions CI context, allowing continuous merging, validation, and release of new content. Custom scripts, such as those for semantically mining publications (fetch_new_candidates.py) and label/ID validation (validate_term_uniqueness.sh), supplement the standard ODK pipeline for the broader discovery and integration of novel terminology.

5. Continuous Integration, Release Automation, and Dynamic Updating

ODK exposes direct integration points for CI/CD via platforms such as GitHub Actions, where common triggers include pushes, pull requests, and periodic scheduled jobs. Complete build-and-test, LLM-driven curation, and release automation (including artifact assembly and BioPortal updates) are built into the workflow. Artifacts and metrics, such as detailed statistics via ROBOT, are published alongside releases, allowing traceability and quantitative ontology growth analysis.

Dynamic updating pipelines monitor literature sources (e.g., Papers-with-Code annotated via OAK), aggregate candidate terms, and feed curated TSVs back into the ODK release workflow. All steps are reproducible within Dockerized environments, with version pinning strongly recommended to prevent toolchain drift.

6. Best Practices, Impact, and Adaptability

Adoption of ODK has resulted in lower curation effort, higher quality ontologies, and rapid onboarding of new contributors, as documented in deployed biomedical ontologies (e.g., Human Phenotype Ontology) and AIO (Joachimiak et al., 2024, Matentzoglu et al., 2022). Key best practices include per-branch modularization, explicit versioning strategy, comprehensive template use, and automated statistics reporting in each release. The selection of the EL profile and ELK reasoner supports efficient, sub-second reasoning feedback suitable for integration in tight CI/CD feedback loops.

A plausible implication is that the combination of human-in-the-loop LLM curation and strict pipeline automation will further increase the scalability of ontology development, especially in domains characterized by fast-evolving technical vocabularies. ODK’s open extensibility provides a blueprint for integration of additional downstream analytics, documentation automation, and provenance tracking.

7. Extensions and Future Directions

Planned extensions to ODK encompass advanced validation schemes (e.g., SHACL, LinkML design pattern enforcement), fine-grained import mechanisms (beyond subclass traversals), and enhanced integration with third-party ontology aggregation services. There is community-driven interest in expanding the tooling ecosystem, incorporating more text mining and advanced reasoning frameworks, and providing pre-packaged GitHub Actions primitives for projects seeking no-local-install workflows.

The ODK framework’s application to AI ontology projects demonstrates its generality beyond the biomedical space, supporting cross-disciplinary, LLM-assisted, and rapidly-adaptive ontology engineering (Joachimiak et al., 2024, Matentzoglu et al., 2022).