Logical Gene Ontology Annotations (GOAL): exploring gene ontology annotations with OWL

In this paper, we explore the use of the Web Ontology Language (OWL) [1] as a practical means of delivering sophisticated querying of mouse gene product annotations such as the Gene Ontology Annotations (GOA) [2], the Mammalian Phenotype Ontology (MPO) [3] and the Human Disease Ontology (HDO) [4]. To achieve this goal, we present the Logical Gene Ontology Annotations (GOAL) tool as a practical way to query across ontologies and explore mouse gene products.

The creation of the Gene Ontology (GO) [5, 6] has had a major impact on the description and communication of the major functionalities of gene products for many species. At the time of writing, GO has more than 35 000 terms for annotating gene products; it is used in more than 40 species-specific model organism databases and in cross-species databases such as UniProt and InterPro [2]. It is widely used for querying such databases, making cross-species comparisons or in data analyses, such as over-expression analysis in microarray data [7, 8].

Similarly, the MPO and HDO have been used, with a lesser coverage, to indicate the associations of gene products with phenotype and disease [4, 9]. The MPO is used in the Mouse Genome Informatics (MGI) database [10] to characterise the phenotypic outcome of mutagenesis experiments in mouse, and serves as a vocabulary that is being applied in the International Mouse Phenotyping Consortium [11] to semantically annotate phenotypes of knockout mice. Both the MPO and the HDO afford a rich potential for querying and analysing a gene products’ biology.

The GO and other bio-ontologies are mainly used as a controlled vocabulary to ensure genes are consistently annotated using standard terminology across many data resources; this alone offers many benefits for data integration and analysis. Bio-ontologies are, however, much more than just vocabularies; they also provide additional information about how the entities they describe are related to each other. In well-formed ontologies, these relationships have a well-defined semantics that bring added value to the ontologies [12]. For example, the hierarchical relationships allow for all kinds of a particular entity to be retrieved, as well as those with an annotation to the entity itself. These and other relationships provide support for navigation, as well as making explicit the relationship between the entities being described. It is these relationships and their associated semantics that enable communication of knowledge and the analysis of the data arising from many experiments.

For the GO, software tools such as the AmiGO browser [13], DynGO [14] or QuickGO [15]) provide interfaces to exploit the hierarchical structure of the GO and to support query expansion. For example, when searching AmiGO for receptor activity genes, the results returned also include genes involved in GPCR activity, because GPCR activity is a subclass of receptor activity. The ontologies’ hierarchical structure is also useful for data mining tasks [16]. Enrichment analysis over the GO [8] is a common technique used in the analysis of high-throughput gene expression data; sets of interesting genes can be grouped or clustered based on common GO annotations (see [17] for more GO tools).

While GO is relatively rich with query-based tools, other ontologies and their annotations are less well endowed. Only a few query tools are able to take the information in several ontologies into account when retrieving annotated data [18–20]. In particular, in the context of complex diseases and syndromes, it can be important to retrieve data based on explicit and implicit relations that hold between classes in several biomedical ontologies. For example, diseases may be related to gene products annotated with a particular molecular function and biological process, so that they produce a distinct phenotype when defective or deactivated. Relations between different kinds of entities (functions, processes, phenotypes, diseases) can be exploited both to perform expressive queries and to add background knowledge in an ontologies’ class definitions. For example, diseases can be characterized based on their phenotype, and phenotypes in turn can be characterized based on molecular functions and biological processes in an organism. These relations can be exploited to suggest causal genes for diseases, identify genes participating in the same pathways and group orthologous genes together [21, 22].

Bioinformatics is, of course, replete with systems for querying data [23, 24]. Cross-resource querying has long been a goal, and to achieve this goal, integration of data and associated meta-data is required. Ontologies are proposed as a means for such integration—a common schema and vocabulary across resources will ease integration, querying and subsequent data analysis [25]. TAMBIS [26] was an early attempt to use ontologies to integrate and query across multiple resources. Latterly, the adoption of ontologies across many resources has eased cross-resource query answering [5, 6]. The rise of Semantic Web [27] technologies, such as the Resource description Framework (RDF) and OWL, has eased some features of integration with resources such as Bio2RDF [28] bringing together many resources in a common format and semantics. Other RDF stores have used ontologies to a greater or lesser extent [25, 29]. Yet, little work has been done to bring ontologies together with the data they annotate in order to use automated reasoning to query those data using the knowledge expressed in bio-medical ontologies. Achieving this goal demands the use of complex query languages as well as highly efficient and optimised automated reasoners.

Whilst highly useful, many of the GO-orientated tools fail to exploit the full potential of the GO’s representation for reasoning and querying over gene annotations. In particular, most of the GO tools that we investigated do not facilitate rich querying that takes into account the semantics of the GO. For example, it was difficult to ask for all proteins that are located in a membrane, or part of a membrane, that are receptor proteins involved in a metabolic process. Extending the queries to include associations of gene product functional attributes, location with phenotype and disease phenomena, such as linking together proteolysis, insulin secretion, plasma membrane, increased glucose concentration and diabetes, is not yet possible. To answer such a query correctly, some form of reasoning over the ontologies is required. The ability to perform such rich queries would enable more precise and flexible exploration of the annotations with GO, MPO and HDO, as well as other ontologies used to annotate gene products.

OWL [1] and the OBO Flat File Format [30] have a strict semantics that makes it possible to use automated reasoners to help build and use knowledge captured in an ontology. In order to explore the potential of reasoning over the various gene product annotations, we need to describe the relationships between the genes and their annotation within a framework that can also exploit the semantics encoded into the ontologies. Our approach uses OWL, for which a mapping from the OBO Flat File Format has been created [30], to represent the GO, MPO and HDO based annotations and the axioms in the ontologies so that query systems can exploit both the ontology and its annotations. In doing this, we will also investigate the ability of OWL to scale to such a task and how to deliver the querying facilities on offer to users in a way that is reasonable to use without having to write complex queries.

As an ontology of biological processes, molecular functions and cellular components, GO itself does not explicitly contain classes for gene products; GO annotations are attached to gene products in databases or flat-files (See http://www.geneontology.org/GO.annotation.shtml). We can use these annotations to create explicit OWL descriptions of the relationships between gene products and their annotations. For example, the mouse gene Taar4 [MGI:2685072] has several GO annotations including ‘integral to membrane’ [GO:0016021], ‘G-protein coupled receptor activity’ [GO:0004930] and ‘signal transduction’ [GO:0007165]. We can create an OWL class that captures the annotations using the following Manchester OWL syntax [31] (note that an axiom annotation is used to assert the evidence code for each annotation):

Class: MGI_2685072

oboInOwl: hasDefinition ”trace amine–associated receptor 4”,

rdfs: label ”Taar4”

SubClassOf:

Annotations: oboInOwl: evidenceCode ”IEA”

ro: located_in some GO: GO_0016021, (integral to membrane)

Annotations: oboInOwl: evidenceCode ”IEA”

GOAL: is_capable_of_function some GO: GO_0004930, (G–protein coupled receptor activity)

Annotations: oboInOwl: evidenceCode ”IEA”

GOAL: is_capable_of_process some GO: GO_0007165 (signal transduction)

rdfs: label ’G–protein coupled receptor activity gene product’

Using the compositional approach to ontology building [32], we can create an ontology from these annotations that explicitly relates gene products to GO, HDO and MPO and then add defined classes to impose a hierarchy. To construct the GOAL ontology we take the existing GO, HDO and MPO classes and create a set of defined classes that enable us to query for gene products. For example, for the GO class G-protein coupled receptor activity we would create a new class that queries for the gene product using the following Manchester OWL syntax:

Class: GOAL: GO_0004930

EquivalentTo:

GOAL: is_capable_of_function some GO: GO_0004930 (G–protein coupled receptor activity)

This defined class will recognize any class of gene product that has these attributes, or children of these attributes, and subsume it within the hierarchy of gene products. In this standard use of OWL subsumption querying and automated reasoning, we can add more of such defined classes to build an arbitrarily complex polyhierarchy for querying and navigation of entities annotated with the GO. Figure 1 shows such an inferred polyhierarchy centered on annotations for the TAAR4 gene product.

The following ontologies and annotations were downloaded on 4 October 2011 and processed as described:

We created the following ‘singleton’ defined classes to act as the ‘building blocks’ for GOAL queries:

The numbers of these classes simply reflect the numbers of classes in each of the supporting ontologies. The total number of defined classes in the GOAL ontology is 37943.

After importing the three branches of GO, the HDO, the MPO, the GOAL ontology of named gene product classes plus the gene product annotations, the ontology contains 108226 OWL classes.

Classification of the GOAL was performed on a 2.2GHz i7 Mac Book Pro requiring around 3GB of memory. Table 1 shows the performance times for each reasoner. Figure 2 shows the classification times in a chart. The ELK reasoner significantly outperforms the other three reasoners when classifying the GOAL ontology. Given that the GOAL ontology is within the OWL 2 EL profile and that ELK is specifically optimised for classifying ontologies in this profile, this result is not surprising.

Reasoner classification times
Reasoner	t1 (ms)	t2 (ms)	t3 (ms)	Mean time (ms)
ELK	2806	2611	2729	2715
CB	31292	31261	33988	32180
Pellet	40519	40801	41009	40776
HermiT	61461	62855	62677	62331

Reasoning times (milliseconds) for the three OWL automated reasoners used to classify the GOAL ontology.

The following queries illustrate the capabilities of the generated ontology that span the annotations from all five imported ontologies. The first query class, written in Manchester OWL syntax, returns subclasses of gene product that are annotated with immune system disease from the HDO. In addition, we know that cytokine genes are secreted by numerous cells of the immune system, so we extended the description to include genes that have a phenotype associated with abnormal cytokine secretion from the MPO. Using the GO annotations, we then reduced the result set further by filtering on genes that have the function of ion binding, participate in an inflammatory response and are located in intracellular membrane-bounded organelle s.

Class: ImmuneSystemDiseaseGeneProduct

EquivalentTo:

GOAL: GO_0006954 (’intracellular membrane–bounded organelle gene product’)

and GOAL: GO_0043167 (’inflammatory response gene product’)

and GOAL: GO_0043231 (’ion binding gene product’)

and MouseGOAL–MP: MP_0003009 (’abnormal cytokine secretion gene product’)

and MouseGOAL–HD: DOID_2914 (’immune system disease gene product’)

Using the GOAL browser, we constructed the intersection of the gene product classes and asked for all subclasses that are gene products. The following code appeared in the DL query box of the GOAL browser:

’intracellular membrane–bounded organelle gene product’

and ’inflammatory response gene product’

and ’ion binding gene product’

and ’abnormal cytokine secretion gene product’

and ’immune system disease gene product’

Running this query returned one gene, Mefυ (MGI:1859396). The Mefv gene (Mediterranean fever gene) is known to play a role in the inflammation response and in fighting infection [33].

Although this is a relatively simple query, some reasoning is required in order for it to return the correct answer. To understand what reasoning took place to answer this query, we used the explanation facility offered by Protégé. We asked the reasoner for an explanation for the Mefv gene’s being a subclass of our ImmuneSystemDiseaseGeneProduct class. Explanations provide the minimal set of axioms asserted in the ontology that are required for this subclass entailment to hold [34]. Figure 3 shows the explanation, in terms of the asserted axioms, for this entailment as it is shown in Protégé. We see from this explanation that the assertions on Mefv involved classes that are deep within the GOAL ontology’s hierarchy. The query class has restrictions on more general terms that the ones used to describe Mefv and it is only by inference up the hierarchy that the query can be answered. This example shows both how the reasoner can be used to answer questions about gene products, and how explanation technology can be used to provide details of how a particular query is being answered.

We can continue to construct such queries and thus begin to explore relationships between GO annotation, observed phenotypes and diseases. The ability to exploit the ontological hierarchies enables us to ask generalized queries, such as all metabolic disease genes and then start filtering down to genes with an abnormal cholesterol level phenotype, and finally down to genes that are involved in a particular biological process in the cell such as regulation of cholesterol storage. The OWL representation and reasoning make such querying and exploration trivial and there are currently no tools available that can facilitate such querying and reasoning over these sets of annotations. The results from the ELK reasoner in particular demonstrate that such interactive querying, supported through automated reasoning, is supportable for ontologies of this size that are expressed in the OWL 2 EL profile.

There are a host of example queries for users to browse via the GOAL web interface. The GOAL interface provides a mechanism to explore these annotations and develop new queries. Its current functionality demonstrates how OWL and reasoning technology can be used to deliver novel search capabilities to biomedical applications. Through GOAL, users can browse the gene product hierarchy generated for each class from the imported ontologies. Selecting a class shows the full OWL description rendered in Manchester OWL syntax. Any selected class can be added to the DL query box to create an intersecting query of named classes. Each query returns all subclasses that are gene products and links through to the gene accession page at the MGI. Users are also free to type in any arbitrary DL query using the Manchester OWL syntax.

Although the queries demonstrated here are relatively simple, they serve to illustrate the potential of a pure OWL approach to querying GOA, MPO and HDO annotations. By combining the five ontologies in one place, we can make queries that connect gene products with basic functional attributes, their disease and phenotype. It is easy to imagine adding further ontologies such as cell type, gross anatomical location and so on, to provide even richer queries. We can already see queries that are not possible in other popular browsing based tools.

By keeping to the OWL 2 EL profile, we can keep the query responsiveness to reasonable limits that enable real-time automated reasoning and access to reasoning through a web interface. Moving outside this profile makes querying based on automated reasoning significantly slower, to the extent of making it currently unfeasible to use in real-time applications that rely on user interactions.

The key, on top of these features, is to provide a user interface to this querying facility. The GOAL user interface, built using GWT, on top of the OWL API means that we can offer access to the GOAL ontology and its query facilities without resorting to clumsy syntax. We have provided a straight-forward means of building reasonably complex queries over a range of ontologies. We have done this by providing what we have called “singleton” defined classes underneath our gene product classes for each of the supporting ontologies. We then do subsumption querying by making anonymous classes of conjunctions of these singleton defined classes. The classifier places these query classes in the GOAL hierarchy, providing results consisting of the primitive classes that represent the actual gene products. We see this mechanism as a reasonably straightforward means of providing sophisticated queries to users.

Using similar patterns, we can also begin to imagine more complex class descriptions that utilise additional expressivity in OWL, such as the use of complement classes to query for genes that is_capable_of_function some (not ReceptorActivity) and is_capable_of_process some SignalTransduction, which would find those genes that have a function other than receptor activity and are involved in signal transduction. (Note that the semantics of OWL implies that such genes can have a receptor activity, but must have some activity that is proven to be disjoint from any receptor activity.) We can also forsee queries that involve disjunctions—’show all gene products that participate in either Sensory perception of sound or Sensory perception of smell or both’. Such expressivity falls outside the OWL 2 EL profile that we have used and the performance of automated reasoning would be significantly worse when these constructs are used: the GOAL ontologies classifies, but response times in the user interface are currently slow enough to detract from usability (data not shown).

Adding further semantics into the ontology will afford further opportunities; adding disjointness axioms to GO may help us uncover mis-annotations and we have yet to fully exploit property characteristics such as transitivity and functionality. We can also explore ways of flexibly incorporating annotations with differing degrees of confidence through use of the GO evidence codes and programmatically generating the defined classes that form the polyhierarchy of genes.

The announcement of the GO cross-products extension to the GO [35, 36] will provide rich formal definitions for some GO classes. These definitions will enable more expressive OWL queries over the GO annotations and the potential to infer more annotations on existing GOA genes [37, 38]. These types of extension should also address the issue of coupling functions, processes and locations together in the GO annotations—something that will improve the computational aspects, including querying, of these annotations. While more precision in the formal representation of the ontology will generally lead to the possibility for more powerful queries, this precision and expressiveness must be balanced with the computational performance of the resulting representation; to do practical work, compromises on representation must be made when appropriate.

In this work, we have made a straight-forward use of OWL and automated reasoning to deliver a flexible way to query all aspects of GO annotations. The polyhierarchy formed also provides similarly rich navigation in a gene product orientated setting. Finally, we provide a flexible framework for exploring and manipulating GO and other valuable annotations developed by the community. By taking care with the OWL profile used and the means of issuing queries, we have shown it is possible to deliver real-time query systems for ontologies via the direct use of automated reasoners.

SJ constructed the GOAL ontology and developed the GOAL web application. RH constructed the MPO and HDO annotation sets and ontologies. SJ, RS, and RH developed the example query classes. SJ and RS drafted the manuscript whilst all authors contributed and approved the final manuscript.