The Gene Ontology (GO) consists of almost 40,000 terms and has been used to annotate millions of gene products to record their subcellular location (e.g., lysosome), their molecular function (e.g., kinase activity) and their wider role in cellular, developmental and physiological processes (e.g., signal transduction) [1]. In its original form, the GO was conceived of as a directed acyclic graph in which terms referring to classes are nodes and edges record relationships between classes including classification (is a) and partonomy (part of). This graph structure is commonly used to group genes annotated with related terms in user-facing tools such as QuickGO [2] and AmiGO [3] and to generate gene sets for enrichment (over- representation) analyses [4] to interpret the results of genomics experiments. For example, an experiment to assess the how treatment of liver cells with a particular drug effects the transcriptome (genome-wide gene expression profile) in liver cells may result in a list of all genes whose expression is increased by the drug treatment. The GO graph structure can be used to group all genes in the relevant genome into sets sharing function or location. One can then ask which gene sets are statistically over or under-represented in the gene list compared to the expected number of of genes from that set in an equivalent length list generated by random sampling from the set of all genes in the genome.

In recent years, the GO has developed into a richly axiomatised formal ontology specified using Web Ontology Language (OWL) [5, 6] and defining GO terms with reference to terms from other ontologies. For example, the GO now records the chemical participants in over 12,000 processes and functions via axioms referencing chemical entities defined by the Chemical Entities of Biological Interest (ChEBI) ontology [7, 8]. Over 8000 GO classes have some direct or indirect logical link to a term from the Cell Ontology (CL) [9] or the Uber anatomy ontology (Uberon) [10]. These record, for example, the location of cellular components (the acrosome and its parts are present only in sperm), cell types that are the sole location of some process (‘natural killer cell degranulation’ only occurs in natural killer cells), and the products of developmental processes (bone is a product of ‘bone morphogen- esis’). There are also over 2500 logical axioms recording the functions of cellular components via links to molecular function and biological process terms.

When combined with standard OWL reasoning technologies, this improved axiomatisation opens up new possibilities for grouping terms and their xannotations in biologically meaningful and semantically precise ways that are potentially useful in enrichment analyses. For example, we can use OWL reasoning queries to group processes occurring in T-cells or in the pancreas, or processes involving nitric oxide or collagen fibers.

The results of enrichment analyses using gene sets for all GO terms can be difficult and slow to interpret due to high levels of overlap between gene sets. There are a number of sources of overlap: grouping via class and part hierarchies means that gene sets derived from annotation to a class subsumes the gene sets of its subclasses and subparts; one GO class can sit in multiple branches of the hierarchy; a single gene product may be annotated to terms in multiple branches. For this reason, many enrichment analyses rely on a more limited number of gene sets, corresponding to grouping under a limited number of high or intermediate level GO terms commonly referred to as a slim.

Rather than use a slim of GO terms, F. Hoffmann-La Roche Ltd. (“Roche”), maintains an internal controlled vocabulary (referred to hereafter as the RCV) for use in enrichment analyses. The RCV consists of around 360 terms, each of which is mapped to a set of terms from the GO. It is tailored to the research interests of Roche, and its terms were chosen with the aim of achieving gene set composition descriptive and broad enough to allow robust and statistically significant results though not so broad and redundant in composition that it prevents easy interpretation of results. Detecting enrichment to gene products involved in anatomy, organ or cell-specific processes or components can be critical for pharmacological research, especially when working with complex tissues where there is a need to tease apart events occurring in specific tissue compartments or cell types. To support this, many RCV terms group GO terms in ways that are out of scope for classes in the GO, including groupings of GO terms related to specific cell, tissue or molecule types.

Here we describe the development and testing of a dynamic, computable mapping between RCV terms and the GO that makes use of OWL reasoning. We show that RCV groupings of GO terms related to specific cell, and molecule types can be used to identify the transcriptomes of those cell types via enrichment analyses.

As the RCV is a flat list and includes classifications that are orthogonal to the classification schemes used by the GO, it is not amenable to mapping via ontology alignment techniques that use ontology structure [11]. Given the small size of the RCV, it is viable to manually map each RCV term to an OWL class expression (DL query), which can then be used in conjunction with an OWL reasoner to generate lists of GO terms. The RCV does not include textual definitions to clarify meaning, so for each RCV term we attempted to find a class expression (a mapping query) that reflected the intended meaning of the RCV term, as judged by the RCV term name, manual mappings and discussion with RCV developers.

Mapping of RCV terms to set of GO terms is now fully dynamic, allowing RCV to be automatically kept up to date as the GO. Where new terms follow mapping query patterns that are already used, they can be added simply by specifying an additional line in the mapping file.

48% of mapped RCV terms have 10 or fewer manual mappings. We are reviewing all of these cases to decide whether to drop manual mappings or whether complete automation might be achieved by a different query strategy. In some cases, a more complete mapping could be achieved by a disjunctive query. For example, all RCV terms referring to metabolism of some specified chemical are mapped to GO terms referring to metabolic and transport processes in which the specified chemical is a participant. (This is consistent with some medical use of the term metabolism.) A more complete mapping could be achieved using a disjunctive query with an OWL2 DL reasoner such as HermiT. This approach was found to be prohibitively slow but new generation reasoners such as MoRE [32] which combine ELK with DL reasoners such as HermiT may turn out to be useful in this approach. A simple, if potentially incomplete alternative would be to simply run two EL mappings using ELK and generate a union of the results.

56 terms were not mapped. Some were rejected from the pipeline as they were judged to be too close in meaning to other RCV terms. The rest were rejected as currently un-mappable due to the lack of suitable terms or axiomatisation within the GO at this time. For example, GO currently has no formal way to group aerobic or anaerobic metabolic processes, although it does reflect the aerobic or anaerobic nature of many metabolic processes in their names and textual definitions.

Our work demonstrates how the logical structure of the GO can be used to achieve biologically meaningful mappings between concepts in external controlled vocabu- laries and corresponding sets of GO terms, even when there is no concept in the GO that is directly equivalent the term to be mapped. This is possible as long as the concept can be mapped to an OWL class expression referencing classes and relations in the full version of the GO. These classes may come from the GO, or from ontologies from which the GO imports classes such as ChEBI, CL and Uberon. The resulting mapping is dynamic and so can easily be kept up to date as the GO evolves.

While OWL 2 DL profile queries could be used for these mappings, this would make mapping software slow to run, resource intensive, and may not be sustainable as the GO becomes still larger and more complex [6]. The mapping system we describe uses class expressions restricted to the OWL 2 EL profile to ensure that mapping is fast and scaleable. It also demonstrates how OWL property chains and property hierarchy can be used to partially overcome the absence of disjunction (OR) in OWL 2 EL.

The RCV includes many terms that group GO terms in novel ways by their rela- tionship to some type of cell, molecule, tissue or cell component. One possible usage of these terms is to provide a mechanism for detecting the signatures of particular cell or tissue types in transcriptomic data. We demonstrate the effectiveness of this by showing how gene sets derived from RCV terms for cell types can be used to identify specific immune cell types and how gene sets derived from RCV terms for cell, molecular and cell component types can be used to identify tissue types. In some cases, gene sets derived directly from transcriptomic data may be used for the same purpose (see Figs 3 and 4). In these cases the RCV term sets provide an alternative method using gene sets with very little overlap to those derived from differential expression (Fig. 3). Unlike gene sets derived from transcriptomic data, those derived from the GO and its annotations are not limited by the ability to experimentally isolate suitable biological samples and can include broad groupings of cell or tissue types that are unlikely to be isolated together (e.g. all glial cells (Fig. 4), or all epithelia).