3. Introduction to terminologies and ontologies¶
3.1. Main objectives¶
The aim of this recipe is to provide a compact introduction about
controlled terminologies and
ontologies, why these resources are central to the preservation of knowledge and data mining and how such resources are developed.
3.2. Controlled terminology or ontology: what’s the difference?¶
The need for
controlled vocabulary often arises in situations where validation of textual information is necessary for operational requirements.
The main initial driver for data entry harmonization is to increase query recall. In its most basic form,
keywords may be used to perform indexation. However, if relying on user input alone, the chances of typographic errors increases with the number of users. These unavoidable events accumulate over time and end up hurting the accuracy of search results and this is the reason for offering sets of predefined values. It reduces the noise.
However this can come at the cost of precision, as the predefined terms may not cover the exact thing users may need to describe. Furthermore, term mis-selection by the user is not eliminated and introduces another type of error.
controlled terminology is a
normative collection of terms, the spelling of which is fixed and for which additional information may be provided such as a
definition, a set of
version, as well as a
license determining the condition of use. The set of information about a specific controlled terminology term is designated as
term metadata. In a controlled terminology, terms appear as a
flat list, meaning that no relationship between any of the entities the controlled terminology represents is captured in any formal way.
This is the main drawback and limitation of
controlled terminologies, which are often developed to support a data model or an application.
ontology on the other hand, is a
formal representation of a domain knowledge where concepts are organized hierarchically. The qualifier
formal refers to a set of axioms and rules based on logic (e.g.
first order logic) to structure, organize and check the consistency of the term hierarchy. As one can sense right away, ontologies are often a more sophisticated artefact, supported by more advanced theoretical frameworks and dedicated tools to develop them (e.g. Protégé, TopBraid Composer, OBO foundry INCAtools or Robot tool).
3.3. How are they built and maintained and why does it matter?¶
In order to improve over simple
controlled terminologies, a huge area of research has developed to provide
frameworks supporting the representations of relationships between entities. The field is known as
formal semantics in knowledge representation circles. One of the most immediately available examples of
entity relationships found in ontologies, and their potential for improving searches, is the
is_a relationship, which aims to cover the Parent/Child relationship that holds between two entities. For instance:
-Vertebrate --Mammal ---Dolphin --Bird ---Pigeon
In this representation,
classes are directly asserted (placed) under a parent class if and only if the rule
new class is a child of the parent Class holds. ‘Orchid’, which in this hierarchy, would not be nested under ‘Vertebrate’.
While working on small structured vocabularies, it is still possible to detect potential errors but this approach does not scale to support real life semantic artefacts which support complex biological and biomedical information systems.
Languages such as RDF, SKOS, and OWL, exist to provide the expressivity required to establish relations between entities.
In turn, building on these formal rules, automatic classifiers, known as a
reasoner, can inspect semantic artefacts to detect inconsistencies and suggest parent classes.
This is a step known as
inference, where new knowledge is produced by the software agent rather than direct assertion by humans.
This provides a significant support, even if far from supporting all the subtleties of actual knowledge.
six important features to consider when selecting a semantic artefact for making FAIR datasets:
What format does it come in?
Is it well maintained, i.e. frequent release, term requests handling, versioning and deprecation policies clarified?
Are there stable persistent resolvable identifiers for all terms?
Who use it and what resources are being annotated with it?
Is it well documented? There should be enough metadata for each class in the artefact and enough metadata about the artefact itself.
3.4. Why are they useful?¶
As outlined in the introduction, the most immediate use for a controlled terminology is to ensure consistency in data entry. Controlled terminologies are important tools to improve data indexing and query recall. The usefulness of ontologies and controlled vocabularies goes beyond this initial use. The main purpose of biomedical ontologies is to structure knowledge so that it can be operated on by software agents.
One needs to also understand that the two processes coexist and operate in parallel. As more experiments are performed, new discoveries are made. This new knowledge needs to be represented in the domain ontology so the new notions can be used to annotate the results of earlier experiments in the context of retrospective analysis.
For example, he Gene Ontology (GO) is a widely used resource to describe
Biological Functions and
The Gene Ontology Consortium maintains the controlled vocabulary and also releases of Genome Wide Gene Ontology Annotations.
These are resources which associate genes and genomic features found in those genomes with GO terms. These are very important resources especially in the context of genome wide analysis such as transcriptomics profiling analysis.
A particular type of analysis,
enrichment analysis, relies on the availability of such annotations to detect departures from the expected probability distribution in an expression profile and which biological processes are most affected in specific conditions.
The applications are plentiful. The importance of ontologies for structuring information will only grow with the need to obtain Machine Learning ready datasets and speed up the readiness of datasets. This is what FAIR is all about.
So ontologies are of particularly help for the following tasks:
Improve query recall: Given a ‘search string’, having a resource which holds synonyms can be used by a search index to retrieve data annotated with a synonym.
Enable query expansion: Owing to the hierarchical (parent/child) structure of ontologies, a search index exploiting this information can retrieve all datasets annotated with a child term of items matching the input search string. For example, searching with the string “breast cancer” againts an ontology aware search index could return records annotated with
ductal carcinoma in-situ (DCIS), both of which are types of mammary gland malignancies.
Build knowledge graphs: Ontology languages can be used to represent domain knowledge and build reference terminologies but the same technologies constitute powerful tools for modelling instance datasets as nodes in a graph and linking resources together.
3.5. Are all ontologies compatible with each other?¶
There is not a simple answer to that question as it depends heavily on the type of tasks data scientists have in mind. If the purpose is simply to improve query recall on a limited set of fields, a curation policy could be devised to mix and match resources to meet the needs at hands, possibly by building an application ontology, i.e. an ontology specifically for designed for the use case with terms drawn from existing ontologies.
However, in a more integrated framework, it is important to be aware of some of the development choices made by the maintainers of the semantic artefacts.
In the context of basic research and model organism based research, the
OBO foundryis an organization which coordinates the development of interoperable resources. GO, mentioned earlier is one of them. The establishment of domain specific reference ontologies sharing the same underlying rules means that some level of compositional development can be done. This means that axioms can be built connecting classes from compatible resources. This point becomes particularly important when considering the role of the
reasonerwhen assessing and checking the consistency of artefacts themselves but also when analysing instance datasets themselves.
In the context of observation studies, the Observational Medical Outcomes Partnership (OMOP) model also relies on controled terminologies such as SNOMED-CT, RxNORM for drugs and LOINC for clinical and laboratory test descriptions.
In the context of Clinical Data collections, the Clinical Data Interchange Standards Consortium (CDISC) models work tightly with CDISC Terminology, National Cancer Institute’s Enterprise Vocabulary Services (EVS) and also recommend use of SNOMED-CT and terminologies such as LOINC, both of which come with specific licensing terms users need to be familiar with.
3.5.1. Use cases and iterative approach¶
The use and implementation of common terminologies will enable a normalization/harmonization of variable labels (data label) and allowed values (data term) when querying a database. Implementing the use of common terminologies in the curation workflow will ensure consistency of the annotation across all studies.
3.5.2. Selection criteria¶
A set of widely accepted criteria for selecting terminologies (or other reporting standards) do not exists. However, the initial work by the Clinical and Translational Science Awards’ (CTSA) Omics Data Standards Working Group and FAIRSharing (http://jamia.bmj.com/content/early/2013/10/03/amiajnl-2013-002066.long) has been used as starting point to define possible criteria for excluding and/or including a terminology resource.
🔸 Absence of term definitions
🔸 Absence of sufficient class metadata (indicator of quality)
🔸 Absence of sustainability indicators (absence of funding records)
🔰 Scope and coverage meets the requirements of the concept identified
🔰 Unique URI, textual definition and IDs for each term
🔰 Resource releases are versioned
🔰 Size of resource (indicator of coverage)
🔰 Number of classes and subclasses (indicator of depth)
🔰 Number of terms having definitions and synonyms (indicator of richness)
🔰 Presence of a help desk and contact point (indicator of community support)
🔰 Presence of term submission tracker/issue tracker (indicator of resource agility and capability to grow upon request)
🔰 Potential integrative nature of the resource (as indicator of translational application potential)
🔰 Licensing information available (as indicator of freedom to use)
🔰 Use of a top level ontology (as indicator of a resource built for generic use)
🔰 Pragmatism (as indicator of actual, current real life practice)
🔰 Possibility of collaborating: the resource accepts complaints/remarks that aim to fix or improve the terminology, while the resource organisation commits to fix or improve the terminology in brief delays (one month after receipt?)
These criteria are simply indicative and need to be modulated depending on the
contexts described in the introduction, as specific constraints (e.g. regulatory requirements) may take precedence over some of the criteria listed here.
Choosing an ontology and semantic resources is a complex issue, which requires careful consideration, taking into account the research context of the data production workflow and regulatory requirements that may apply. The choices made affect the integrative potential of a dataset as they influence the level of
Clearly, declaring the semantic resources used to annotate a dataset also influences the
reusability and it is good practice to do so as it allows potential users to gauge the amount of mapping work that may be required to combine two datasets.
OBO Foundry. http://obofoundry.org/
CDISC Controlled Terminology. https://www.cdisc.org/standards/terminology
Gene Ontology. http://geneontology.org/
Topbraid composer. https://www.topquadrant.com/products/topbraid-composer/
This page is released under the Creative Commons 4.0 BY license.