The Genom of Homo sapiens.pdf

ONTOLOGIES FOR BIOLOGISTS 229al. 2003). Another way to view this is a progression beyonda vocabulary of fixed phrases toward a combinationof a vocabulary and a grammar for the creation of phrases(see, e.g., WordNet 2003).GO will use a Description Logic (DL) to allow certainGO terms to have properties (i.e., attributes or slots).Properties provide a formalism for dealing with the classificationof finely granular terms by allowing the flexiblecreation of phrases. This approach will move GO awayfrom being merely a word-turned-phrase-based ontologytoward a property-based ontology. This will offer the biologicalcurators a structured means of composing phrasesto use during annotation. A phrase is composed of a primaryterm (e.g., “biosynthesis’) which dictates what othertypes of terms must interact with it in supporting or secondaryroles (e.g., “interleukin-13”). The term “biosynthesis”would include a property for a required term indicatingthe thing-being-synthesized and thus an explicitGO term for “interleukin-13 biosynthesis” would nolonger be necessary. When annotating, if the curator selected“biosynthesis,” she would then be obligated to fillin the “thing-being-synthesized” property, in this case,“interleukin-13” or another identifier from a biochemicalmoietyontology. From the gene product curator’s perspective,the annotations become phrases, composed of aprimary term and some number of other terms that arenecessary to complete the explanation. The number ofproperties required for a term may vary from zero tomany. As an example to illustrate this, consider the term“localization.” A curator would never use this term withoutindicating both the object and a prepositional phrase.Neither “localization,” nor “localization of mRNA,” nor“localization to the cell membrane” suffices; to makesense one must say “localization of mRNA to the cellmembrane.” From the ontology curator’s perspective,defining a term will now include specifying the requiredproperties for that term, to indicate whether or not, andwhat type of, supporting terms are required when a geneproduct is actually being annotated. The use of propertiesappears robust and extensible, and gives both ontology curatorsand biological curators flexibility, while retainingsemantic rigor. It is entirely consistent with both DescriptionLogics and frame-based ontologies, both of which arewidely used in the knowledge representation–artificial intelligence(AI) community. This approach is also verymodular because, as phrases are constructed, they may beused recursively to create more complex phrases.Terms that are modifiers also may be added to thephrase, the difference being that modifying terms are notdemanded by the primary term, but simply provide additionalinformation. So, for the term “biosynthesis,” althoughit is obligatory that a curator supply a term for thething-being-synthesized property, they would not beobliged to fill in the property for a mediator of thisbiosynthesis. An annotation for a protein could then bemade for biosynthesis (primary term) of GPI anchor(mandatory completion property) via N-glycyl-glycosylphosphatidylinositolethanolamine(optional informationalproperty).To illustrate the difference of this approach from thatnow used, this is a current entry from Swiss-Prot for a hutheGO concept “antigen binding activity” in the GOdatabase for utilitarian reasons.All GO concepts have a unique identifier number and adefinition that explicitly states their precise meaning.There is a strict one-to-one correspondence between aGO_identifier and its definition (rather than to the lexicalstring used to refer to the concept). Thus, a change in aconcept’s lexical string (but not its definition) will not alterits GO_identifier; conversely, if there is a change inthe definition that alters the meaning of a GO conceptthen, even if its lexical string remains identical, this willresult in the new concept having a new GO_identifier. Ifa GO concept is found to be incorrect or irrelevant it is notsimply discarded, it is marked with the attribute is_obsoleteand it, its GO_identifier, and definition remain in thedatabase.Finally, GO concepts may well be equivalent or have aclose relationship with a term in some other database. Forexample, a GO concept describing the activity of an enzymecan be considered to have a close relationship withthe name of that enzyme within the Enzyme Commission’sdatabase. This relationship may be expressed as adatabase cross-reference. GO also maintains a number oftables that represent semantic mappings of GO conceptsto concepts in other databases; e.g., between GO conceptsand Swiss-Prot keywords and between GO concepts andthe functional catalog of MIPS (MIPS 2003).LIMITATIONS OF THE GO MODELDuring the early stages of the development of GO, adecision was made to restrict GO to using only the relationshipsof subsumption and meronomy. This decisionwas made for pragmatic reasons, and it was accepted thatat some time in the future the implicit limitation in the expressivepower of GO might become limiting. That futurehas arrived: GO, and similar artifacts, need to be property-basedontologies to be sufficiently expressive.Currently, GO includes a large number of terms thatcan be exemplified by “biosynthesis of interleukin-13.” Itis obvious that this is a compound term, composed of anoun describing the action “biosynthesis” and a prepositionalphrase “of interleukin-13,” describing what is beingbiosynthesized. The same idea could also be expressedusing the verb “biosynthesizes” and the directobject “interleukin-13.” In either expression there is acentral topic, “biosynthesis” or “biosynthesizes,” that requiresan additional supporting term, “interleukin-13,” tocompletely convey the entire concept.At present, GO terms are indivisible; there is no notionof the decomposition of phrases into either individualwords or concepts and properties. If GO continues to introducecompound terms, the system will become progressivelymore redundant, less flexible, and increasinglydifficult to manage. This is apparent in the parts of GOthat involve implicit cross-products (e.g., between a processlike “transport” and an ontology of chemical compounds).We must therefore provide a solution that bothdisambiguates fundamental concepts and also enables accurateannotation. One approach is to migrate to a property-basedontology (a Description Logic; see Baader et