Introduction
In bioinformatics, phenotypes—observable characteristics such as "high blood calcium"—have traditionally been represented as "pre-composed" terms within ontologies like the Human Phenotype Ontology (HPO). While these labels are useful for human researchers, they act as "black boxes" for computational systems. A machine does not inherently understand that "Hypercalcemia" involves the chemical entity "Calcium" or the anatomical location "Blood." To overcome this, the field is shifting toward post-composed descriptions.
The Shift to Post-Composition
Post-composition defines a phenotype as a combination of entities from multiple specialized ontologies. For example, Hypercalcemia can be decomposed into a chemical (from ChEBI), a quality (from PATO), and a location (from the FMA).
While this approach enables more granular data analysis, it introduces significant logical complexity. When independent ontologies—developed by different teams with varying assumptions—are merged, the resulting logic can break, leading to "logical leaks" or unintended inferences.
Methodology: Modularization and Reasoning
The authors developed a pipeline to assess these integrations using Description Logic (DL). A primary challenge is the sheer size of ontologies like the Foundational Model of Anatomy (FMA), which contains approximately 80,000 concepts. To make reasoning computationally feasible, the team employed locality-based modularization to extract only the relevant "slices" of external ontologies.
The integrated knowledge base, $HPO_{ALL}$, was then processed using the HermiT reasoner. To identify changes, the authors calculated the "logical difference" between the original HPO and the merged version, focusing on subsumption ("is-a") and "part-of" relationships.
Key Results and Logical "Bugs"
The automated reasoning process uncovered 759 new "is-a" relationships within the HPO. However, it also revealed several logical "bugs" or unintended consequences of the integration:
- Unintended Equivalence: The reasoner incorrectly concluded that certain distinct terms were logically identical. For example, Edema was equated with Generalized Edema because of definitions provided by the FMA that collapsed the hierarchy between them.
- Cross-Ontology Entailments: The system identified 80 new relationships between concepts in different domain ontologies (e.g., a Gene Ontology process being classified under an FMA anatomical entity). These were largely attributed to overly broad range axioms in the FMA.
Conclusion
This study demonstrates that logic-based assessment is a scalable and necessary part of the ontology development lifecycle. By using modularization, the researchers reduced a potentially hours-long reasoning task to just 45 seconds. Future work will focus on developing tools to help human experts "repair" these unintended logical consequences and incorporating existing cross-ontology mappings to further refine the integration process.
Summary updated