TL;DR
This paper reveals that a significant portion of Gene Ontology (GO) terms are built compositionally, with many terms containing other GO terms as substrings. These substring relationships often correspond to meaningful "derivational phrases" that encode specific semantic relationships, offering new ways to computationally access and enrich GO's knowledge base.
What Problem Does This Paper Address?
The Gene Ontology (GO) is a crucial resource for describing genes and gene products, organizing concepts into a hierarchical structure. However, the semantic relationships explicitly encoded in GO (like "isA" and "partOf") don't always capture the full richness of information present in the natural language terms themselves. For instance, a term like "positive regulation of cell migration" implies more than just a hierarchical relationship; it signifies a specific type of regulation and its direction. This implicit information within the terms is largely inaccessible to computational analysis. The paper aims to explore the linguistic structure of GO terms to uncover these hidden semantic relationships and make GO more computationally useful.
Methods
The authors analyzed the June 2003 release of the Gene Ontology, which contained 13,361 concepts, 13,361 names, and 3090 synonyms, totaling 16,451 terms. They systematically examined all terms for the occurrence of other GO terms as substrings. For each such occurrence, they classified the relationship between the two terms based on GO's hierarchical edges (isA, partOf, synonymy) and dominance relations. They also analyzed the "complements" – the phrases added to a substring term to form a larger term. Two key aspects of these complements were examined: their frequency of occurrence and the consistency of their usage within specific ontological contexts (e.g., always appearing with "isA" relationships or within particular subtrees).
Datasets
The primary dataset used was the XML-formatted version of the June 2003 release of the Gene Ontology. This dataset comprised 13,361 concepts, 13,361 names, and 3090 synonyms. No new benchmarks were introduced; the study focused on analyzing the existing GO structure.
Key Insights and Ideas
- Ubiquitous Substring Relationships: A surprisingly large percentage (65.3%) of GO terms contain another GO term as a proper substring. This suggests a highly compositional nature to GO term construction.
- Derivational Phrases as Semantic Clues: These substring relationships often coincide with "derivational phrases" (e.g., "regulation of," "positive regulation of"). These phrases are not GO terms themselves but frequently recur and appear to encode specific semantic relationships that complement the explicit GO hierarchy.
- Complements Encode Meaning: Many of these recurring phrases act as "complements," adding semantic content to the base term. For example, "regulation of" in "regulation of cell proliferation" specifies the type of relationship.
- Consistency of Complements: The study found that these complements tend to be used consistently with particular GO relationships (isA, partOf) and often occur within specific subtrees of the ontology. This consistency suggests they are not arbitrary but carry meaningful semantic weight.
- Potential for Knowledge Enrichment: The identified derivational phrases can be used to infer new semantic relationships, suggest new GO concepts, and identify potential synonyms, thereby enriching the GO knowledge base.
Key Results
- 65.3% of GO terms contain another GO term as a proper substring. This accounts for 72.2% of GO concepts.
- 52.5% of GO terms are inferior nodes containing their superior nodes via an "isA" relation.
- "regulation of" is a highly frequent complement, appearing 1,053 times in GO terms, often directly modifying another GO term.
- Complements show consistency with specific relation types: For instance, 88.8% of complements found in "isA, A << B, A ⊂ B" relationships were specific to that type.
- Analysis of complements can suggest new GO terms and relationships: The authors identified potential new GO terms and relationships by analyzing frequently occurring complements and their contexts. For example, they suggested "REM sleep" as a potential GO concept and a new "undergoesProcess" edge between "nucleosome" and "nucleosome disassembly."
Limitations and Future Work
- Distinguishing Trivial vs. Non-Trivial Relations: The paper acknowledges the challenge of differentiating between substring relationships that are semantically meaningful and those that are merely incidental.
- Computational Expense of Linguistic Analysis: While GO terms offer implicit information, extracting it computationally can be complex. The authors propose their substring analysis as a more efficient way to find linguistic boundaries compared to extensive morphological or syntactic analysis.
- Future Work: The authors suggest several applications for their findings:
- Assisting GO evaluation and curation: Identifying frequent complements could inform the creation of new GO concepts or relationship types.
- Enriching GO's conceptual representations: Derivational phrases can suggest new relationships and make the GO knowledge base more interconnected.
- Natural Language Processing (NLP): The compositional structure of GO terms can aid in recognizing variant forms of terms in natural language texts, particularly in handling coordinated phrases.
- Text Data Mining: The identified relationships can be valuable for mining molecular biology literature.
Summary updated