1.1 Motivation: the Gene Expression Atlas

The Gene Expression Atlas (Kapusheskyet al.,2010) provides summaries of gene expression across multiple experimental conditions, called ‘experimental factors’. It also provides a gene level view of experimental data acquired from ArrayExpress (Parkinsonet al.,2009). This data is manually curated to provide an explicit, consistent and homogenous description across a wide range of sample attributes, such as species, developmental stage, disease and tissue type. Protocol parameters related to the processing of samples, such as application of chemical compounds and sampling times, are also needed. As of November 2009, there are ∼40 000 unique annotations of sample or assay properties covering 330 species in datasets suitable for the Gene Expression Atlas.

Given the diverse nature of the annotations, there is a need to support complex queries that contain semantic information. For example, the query, ‘which genes are under-expressed in brain cancer samples in human or mouse’, requires the querying mechanism to understand the term ‘cancer’. Annotations made at the experimental level are necessarily granular in nature; an experiment where the sample is of adenocarcinoma will be annotated with ‘adenocarcinoma’ rather than more generally ‘cancer’. A database query requiring ‘cancer’ would therefore not return annotations to adenocarcinoma since this requires additional knowledge. An alternative solution would be to annotate this sample with adenocarcinoma and cancer and any other intermediate classifications such as ‘carcinoma’; however, this has a number of disadvantages. First, this requires curation, a labor-intensive process. Second, it embeds the semantics within the database, tightly coupling the data with the domain knowledge. This makes the approach fragile, since a change or extension to domain knowledge may require a large database update. It also limits reuse of the knowledge within other resources.

A better solution to this problem is to annotate data using ontologies. This enables the separation of the formal description of domain knowledge, allowing reuse of these resources and improving interoperability with other data with similar semantic representations. To annotate the diverse data in the Gene Expression Atlas, classes are required from multiple existing ontologies to capture the cross-domain nature of the data.

Initially, we limited scope to data generated to 12 species including: human, mouse, rat,Arabidopsis, budding yeast, fission yeast,Drosophila melanogaster,Caenorhabditis elegans and zebra fish. These species have ontologies that describe anatomy and developmental stages, though the limitations of gene expression technology mean that only a subset of tissues or other variables are typically analyzed. An important use case is the comparability between experiments, for example, where the same tissue, cell type, disease and developmental stage was studied across experiments and species and the data can be potentially combined. Finally, name value pairs that could be mapped to existing domain ontologies were prioritized as these also cover the most common queries e.g. disease state, cell line, cell type developmental stage, etc. Data in the gene expression domain are typically not mapped to an ontology at the point of submission, and neither Gene Expression Omnibus nor ArrayExpress use species-specific ontologies in their submission tools. Requiring use of ontologies at this point is a barrier to data deposition, therefore, the majority of ontology mapping occurs after submission and is based on user-supplied name value pairs e.g. ‘DiseaseState=breast cancer’. An important use case is text mining of data prior to its inclusion in ArrayExpress.

Exploratory analyses of the data prior to the construction of EFO revealed that many terms appear at high frequencies and there is a ‘long tail’ on the data distribution (Maloneet al., 2009). For example, in the ArrayExpress archive ∼1350 samples have the annotation ‘heart’, 65 ‘ventricle’, 14 ‘myocardium’ and a single annotation for ‘pericardium’. Compare this with the representation of the human heart from the Foundational Model of Anatomy (FMA) (Rosse and Mejino,2003) where there are >20 terms describing the various parts of the heart. It is clear that comparatively few terms are needed to describe the data in the Gene Expression domain and that the complexity in FMA is not needed. For both text mining and query purposes across free text in the data, there is a requirement for synonyms. This includes ‘local synonyms’, e.g. ‘whole brain’, to detect user-defined annotation or to deal with alternate spellings. Our approach for the gene expression domain therefore is analogous to that of the GO (Blake and Harris,2008), which was initially developed to describe gene products for model organism databases; it has a data-driven motivation, with ontological principles such as use of an upper level ontology applied to provide robustness and to allow interoperability with other ontologies.

2.1 Detecting external ontology changes

Ontologies that are used within biology evolve rapidly due to scientific advances and because the associated computational technologies are themselves rapidly evolving (Smithet al.,2007). Because we consume from multiple ontologies, class information must be maintained and updated. This problem is no more severe than if we mapped data annotations to each ontology separately, rather than to EFO that maps to external reference ontologies.

Here we list the changes in external ontologies which affect EFO:

1. An axiom is added to an existing named class.

2. An axiom is removed from an existing named class.

3. A new named class is added to the ontology.

4. A named class is made obsolete.

5. An annotation property is edited on a named class.

The OWL-API (Horridgeet al.,2009) provides a Java-based interface which allows manipulation of OWL ontologies at the axiom level. Therefore, comparing two different versions of OWL ontologies in an axiom-based approach, as seen in the OWL-API, can be achieved using a set difference operation. In set theory this is given by a relative complement. Formally, for setsA andB the relative complement ofA inB, that is, the set of elements in B, but not in A, is given as:

(1)

Given two sets of axioms,A andB, and axioma_n:

For a set of axioms which are equal:

(2)

(3)

We can use this information to deduce that no changes have occurred between ontologies and moreover to infer that the classesA andB are logically equivalent. We have designed a freely available tool, Bubastis, to analyze and report on the five major types of ontology changes we enumerate. Specifically, we extract the classes mapped to EFO and check for changes. A log of any changes is created along with relevant time and date stamps and a report generated. Usefully, if there are no changes the tool will automatically report this too.

This approach allows us to computationally manage the imports and mappings we create within EFO, ensuring they are valid and reducing the overhead on ontology curation. It also allows us to manage remapping data annotations to EFO which makes the curation process easier. Importantly, this allows us to maintain a consistent use of external resources ensuring that we do not map to obsolete classes, and erroneous mappings caused by external changes are flagged. There is still an outstanding issue of how ‘correct’ the external resources are. For example, reference ontologies EFO has consumed contain their own mappings which we have further imported to expand interoperability. On scrutiny, some of these were found to be incorrect. For example, mappings to EFO class brain structure derived from synonyms in an external ontology Minimal Anatomical Terminology (MAT) included abnormal brain. Errors of this type are communicated back to the authors of the source ontology. This represents a useful feature of this methodology; we review how reference bio-ontologies map to one another and how correct these mappings are. It is clear that synonyms are used in different ways in different contexts and care must be exercised when using these; we now validate synonyms prior to including these and provide feedback both requesting terms and flagging errors when performing mapping.

2.2 Creating an ontology view

While an upper level framework can provide structure to the ontology, such high-level classes (cf.Fig. 1) can often appear as abstract and confusing for biological users. For example, the Basic Formal Ontology (BFO) (Grenon and Smith,2004) contains the classescontinuant andoccurent. Such classes are useful to organize the ontology and to aid interoperability between ontologies, but are less helpful for a biological user. With this in mind, we use only some of BFO within EFO, and those parts are hidden from users. First, we create an annotation property,ArrayExpress_label, which we use to indicate a preferential label that is displayed in the Atlas browser which replaces any other label on the class, though such labels may also be synonyms and are supported for queries. For example, the BFO classprocessual entity is displayed asprocess in the Atlas user interface for readability. Second, we use a further annotation propertyorganizational_class which is given a value of ‘true’ in any classes we wish to hide from the user (e.g.disposition) which are identified as structural and which are not desired to be visualized in queries. This allows us to show parts of the ontology relevant to the users, while still using an accepted upper level ontology.

Views generated from EFO are used in both the Atlas and ArrayExpress Archive. EFO is used to improve searching across textual experimental descriptions and key value pairs used to annotate samples. When a user enters a keyword that matches an EFO class, synonyms found inalternative_term annotation properties in EFO classes are also used in the search, thereby returning extra matches. We also provide an option to extend searches with classes related to their query viais_a orpart_of ontological relations.

This functionality as deployed in the ArrayExpress Archive is powered by the Apache Lucene as a search engine, and we have packaged the EFO-powered search extension as a separate Java library. The algorithm consists of two parts. First, EFO in OWL format is parsed; the ontology tree is traversed and synonyms, allpart_of oris_a children for all classes in EFO are extracted and a map structure is built for fast lookup. Second, the map structure is used with a rewritten input Lucene Query with additional synonyms and children (if the option is selected and if they exist). For our previous example, query ‘breast carcinoma’ is transformed to (breast carcinoma OR ‘breast cancer’ OR ‘ductal carcinoma in situ’, etc.). This library is available as stand-alone JAR, Maven artifacts and source code fromhttp://github.com/arrayexpress/ae-interface/tree/master/components/efo-query-expand/. The Gene Expression Atlas code base is currently under revision to create a stand-alone install anywhere utility, which will also become publicly available and open source in the near future.

2.3 Supporting the linked data vision

In addition to using EFO within the Gene Expression Atlas and ArrayExpress Archive, we also embrace the ideas of linked data and integration with external resources. In the context of the semantic web, linked data describes a method of creating typed links between data (Bizeret al.,2009). In EFO, we use dereferenceable URIs for all of the classes in the ontology which are assigned EFO URIs. Such classes are assigned a unique identifier, e.g.http://www.ebi.ac.uk/efo/EFO_0000001, with the number fragment incremented for each new class. Since each of these identifiers is dereferenceable via the http protocol, they can be requested from a web server and information about the class returned as user-friendly content. These pages contain information such as the Resource Description Framework Schema (RDFS) class label, parent classes, child classes and annotation properties e.g. text definition. These pages are also machine readable: the source code for each page is actually an EFO Resource Description Framework (RDF) fragment describing a specific class. Computer agents can therefore interact with the EFO class web pages in a way that is analogous to human user interaction. Note, this does not apply to those classes in which URIs are imported from external ontologies; for such ontology URIs, the owner of these ontologies would be responsible for creating dereferenceable URIs.

There are two key elements to linking gene expression data. The first is that parts of the Atlas sample and assay data are annotated with EFO class identifiers. We associate data elements to our explicit definition of what the data represent, by annotating each experiment with EFO classes. The second element is the set of cross-ontology mappings that are maintained within EFO. As EFO is an application ontology, there is an advantage in reusing and importing classes from existing ontologies where possible. Not only does this reduce the effort in adding new classes to EFO, but it also provides interoperability (via cross references) with other resources that use these existing ontologies.

There are a number of challenges associated with this approach. One of the most challenging is deciding upon the appropriate ontology to select when attempting to reuse classes. In the simplest case, where overlap does not exist and there is a clear single authoritative reference ontology, we simply import that class with the original URI maintained, e.g. BFO. However, there are a limited number of examples for where this is the case; for many terms, there may be multiple classes that can fulfill the required definition. For example, consider the termhypertension: as of January 2010, there are 12 exact matches for this class label when querying the NCBO BioPortal and most of these ontologies provide definitions consistent with our data annotation use cases. For this reason, we performed some preliminary data-ontology mapping which allowed us to both assess the matching algorithm and the available reference ontologies for coverage on gene expression data (Maloneet al., 2009). Recently, a tool designed to assist with selecting the most suitable ontologies for a given task has become available, and essentially replicate our early work in an extensible framework (Jonquetet al.,2009). As we require EFO to be cross-referenced to as many external functional genomics datasets as possible, we maximize interoperability and therefore add as many mappings to different ontologies as are valid for our given class and curate these. The decision to create multiple external mappings to EFO classes clearly presents additional overhead to both the initial set of mappings and the subsequent maintenance of these mappings, as both are labor intensive if performed manually. We have therefore developed semi-automatic mapping tools.

Our matching approach uses the Metaphone (Phillips,1990) and Double Metaphone algorithms (Phillips,2000), which were selected following an empirical study of commonly used matching algorithms and their utility in the biomedical domain (Maloneet al.,2008).¹ We were particularly interested in algorithms yielding low false positive rates, as we wished to use the same algorithm for semiautomatic annotation of incoming data to the ArrayExpress Archive as a curator aid. Following the evaluation of several algorithms, a combined strategy was implemented using Metaphone for a first pass and then falling back to Double Metaphone for those terms not matched by Metaphone. This strategy yields the highest overall number of matches with minimal human intervention (required only for multiple matches). Verified matched terms identified by this strategy were included as valid mappings in EFO and added to adefinition_citation annotation property. We have developed a species-specific ranked list of preferred ontologies with known good coverage when mapping to new terms. We prefer to use OBO Foundry candidate ontologies when these provide good matches and use general uncurated resources like Unified Medical Language System (UMLS) only when necessary.

3.1 Querying gene expression data

In order to demonstrate that EFO is fit for purpose, we evaluate it against the competency questions and use cases. One of the primary use cases for EFO was to annotate ArrayExpress data (i.e. providing ontological coverage) and to ask meaningful questions of gene expression data.

In a recent review conducted by Jonquetet al. (2009), EFO was assessed for coverage in annotating biological datasets using an NCBO tool, the Open Biomedical Annotator. Alongside 98 English ontologies in UMLS 2008AA and 92 of the BioPortal ontologies, in total, these resources offer a dictionary of 3 582 434 classes and 7 024 618 textual terms. Following experimentation with three biological datasets, EFO is reported as fourth best in all three tests. EFO performs well in these tests due to the data-driven development, cross-domain method we use. It is also noticeable that the ontologies that finished in the top three were significantly larger than EFO, for example, NCI Thesaurus has ∼35 000 classes compared with ∼2600 classes in EFO.

The real return for the user when using an ontology is in the additional relations used for improving queries. Due to the relations used in EFO, we are able to ask general questions without requiring that every subtype is enumerated in the query. For example, for experiments about cancer, we want all subtypes of cancer, for example prostate carcinoma, without requiring the user to specifically enumerate these subtypes and we want to return only subtypes for which we have data. Similarly, we want a user to be able to ask for ‘forebrain’ and the query to return data that is annotated with forebrain substructures such as hypothalamus. Finally, for mouse we want to return data annotated toMus musculus and substrains thereof.

Figure 4 shows the results of a query for ‘gene that is expressed in craniofacial tissues or sub structures’. Within the ontology, relations are made between classes such as those seen inFigure 4. Specifically here, the query is asking forgenes which are over- or underexpressed in assays that are annotated with an organism part that is craniofacial tissue or a sub-structure.Figure 4 presents the parts of the ontology that satisfy this query at the top of the image. The tree includes classes such aseye (synonymeye structure), which expands to include its substructures such asretina.

As described earlier, EFO has also been used in the ArrayExpress Archive (http://www.ebi.ac.uk/arrayexpress) to enrich querying.Figure 5 illustrates that in addition to keyword ‘breast carcinoma’ (yellow), EFO-enabled search returns experiments matching is-a children, e.g. ‘medullary breast cancer’.

3.2 Linking data through BioPortal

An additional advantage to using an ontology to annotate data in the Atlas is in the use of external ontology tools. The BioPortal resource at NCBO is an open repository of biomedical ontologies that provides access via web services and web browsers to ontologies developed in OWL, RDF and OBO format (Noyet al.,2009). It allows the searching of biomedical data resources such as ArrayExpress, through the annotation and indexing of these resources with ontologies that can be accessed through BioPortal.

Since EFO is used to annotate data in ArrayExpress and also provides multiple mappings to other ontologies, it is possible to query data through BioPortal using ontology class names and return annotated data from multiple resources via the BioPortal's Resources facility, for example, pathway data from Reactome.

1. Blake JA, Harris MA. The Gene Ontology (GO) project: structured vocabularies for molecular biology and their application to genome and expression analysis. Curr. Protoc. Bioinformatics. 2008 doi: 10.1002/0471250953.bi0702s23. Chapter 7, Unit 7.2. [DOI] [PubMed] [Google Scholar]
2. Bizer C, et al. Linked data - the story so far. Int. J. Semant. Web Inf. Syst. 2009;5:1–22. [Google Scholar]
3. Grenon P, Smith B. SNAP and SPAN: towards dynamic spatial ontology. Spat. Cogn. Comput.Interdiscip. J. 2004;4:69–104. [Google Scholar]
4. Gómez-Pérez A, et al. Ontological engineering. New York: Springer-Verlag; 2004. [Google Scholar]
5. Horridge M. The OWL API: A Java API for Working with OWL 2 Ontologies. Proceedings of the OWL: Experiences and Directions 2009, Chantilly, USA, CEUR Workshop Proceedings. 2009 [Google Scholar]
6. Horrocks I, et al. From SHIQ and RDF to OWL: the making of a web ontology language. J. Web Sem. 2003;1:7–26. [Google Scholar]
7. Jonquet C, et al. Proceedings of the 12th Annual Bio-Ontologies: Knowledge in Biology, SIG, ISMBECCB 2009. Sweden: Stockholm; 2009. Prototyping a Biomedical Ontology Recommender Service; pp. 65–68. [Google Scholar]
8. Kapushesky M, et al. Gene Expression Atlas at the European Bioinformatics Institute. Nucleic Acids Res. 2010;38:D690–D698. doi: 10.1093/nar/gkp936. [DOI] [PMC free article] [PubMed] [Google Scholar]
9. Malone J, et al. Proceedings of ISMB 2008 SIG meeting on Bio-ontologies. Toronto, Canada: 2008. Developing an application focused experimental factor ontology: embracing the OBO community. [Google Scholar]
10. Noy NF, et al. BioPortal: ontologies and integrated data resources at the click of a mouse. Nucleic Acids Res. 2009;37:W170–W173. doi: 10.1093/nar/gkp440. [DOI] [PMC free article] [PubMed] [Google Scholar]
11. Osborne JD, et al. Annotating the human genome with Disease Ontology. BMC Genomics. 2009;10:S1–S6. doi: 10.1186/1471-2164-10-S1-S6. [DOI] [PMC free article] [PubMed] [Google Scholar]
12. Parkinson H, et al. ArrayExpress update–from an archive of functional genomics experiments to the atlas of gene expression. Nucleic Acids Res. 2009;37:D868–D872. doi: 10.1093/nar/gkn889. [DOI] [PMC free article] [PubMed] [Google Scholar]
13. Phillips L. Hanging on the Metaphone. Comput. Lang. 1990;7:39–43. [Google Scholar]
14. Phillips L. The Double Metaphone Search Algorithm. C/C++ User's Journal. 2000;18:38–43. [Google Scholar]
15. Rosse C, Mejino J.VL. A reference ontology for biomedical informatics: the Foundational Model of Anatomy. J. Biomed. Inform. 2003;36:478–500. doi: 10.1016/j.jbi.2003.11.007. [DOI] [PubMed] [Google Scholar]
16. Schofield PN, et al. Post-publication sharing of data and tools. Nature. 2009;461:171–173. doi: 10.1038/461171a. [DOI] [PMC free article] [PubMed] [Google Scholar]
17. Sirin E, et al. Pellet: a practical OWL-DL reasoner. J. Web Sem. 2007;5:51–53. [Google Scholar]
18. Smith B, et al. Relations in biomedical ontologies. Genome Biol. 2005;6:R46. doi: 10.1186/gb-2005-6-5-r46. [DOI] [PMC free article] [PubMed] [Google Scholar]
19. Smith B, et al. The OBO Foundry: coordinated evolution of ontologies to support biomedical data integration. Nat. Biotechnol. 2007;25:1251–1255. doi: 10.1038/nbt1346. [DOI] [PMC free article] [PubMed] [Google Scholar]
20. Stevens R, et al. Ontology-based knowledge representation for bioinformatics. Brief. Bioinformatics. 2000;1:398–414. doi: 10.1093/bib/1.4.398. [DOI] [PubMed] [Google Scholar]
21. Uschold M, Grüninger M. Ontologies: principles, methods and applications. Knowl. Eng. Rev. 1996;11:93–115. [Google Scholar]

Supplementary Materials