The Gene Ontology (GO) consists of around 40,000 terms refering to classes of biological process, cell component and gene product activity. It has been used to annotate the functions and locations of several million gene products. Much pharmacological research focuses on understanding how disease conditions di er from physiological conditions in molecular terms with the aim of nding new drug targets for therapy. Gene set enrichment analysis using the GO and its annotations provides a powerful way to assess those di erences. Roche has developed a bespoke controlled vocabulary (RCV) to support enrichment analysis. Each term is manually mapped to a list of Gene Ontology (GO) terms. The groupings are tailored to the research aims of Roche and as a result, many groupings are out-of-scope for GO classes. For example, many RCV terms group process and cell parts according to the cell type they occur in. The manual mapping strategy is labour intensive and hard to sustain as the GO evolves. We have automated mappings between RCV and the GO via OWL-EL queries. This is made possible by extensive axiomatisation linking the GO to ontologies of cells, anatomical entites and chemicals. We can fully automate mapping for approximately one third of the terms in the RCV, with another 40% having 10 or fewer GO terms requiring manual mapping. Automated mapping uncovers many missing mappings. GSEA using the resulting, semi-automated mapping of RCV to GO detects enrichment to gene sets missed with the manual-only mapping. The OWL query approach we describe can be used as the basis of new ways to query the GO, group annotations and carry out GSEA. Importantly, it allows the classi cations used in enrichment analysis to be much more closely tailored to the needs of researchers and industry than was previously possible.
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) [
Much pharmacological research focuses on understanding the molecular differences between disease conditions and physiological conditions, with the aim of nding new drug targets for therapy. Di erential expression experiments analysing disease models or pathological tissue samples are an important source of data contributing to our understanding of this. GSEA using GO derived gene sets is an e cient way to nd functionally coherent gene sets that are statistically over or under-represented in gene lists derived from these experiments. GSEA results using the full GO and large numbers of genes can be di cult 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 heirarchies 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 heirarchy; a single gene product may be annotated to terms in multiple branches.
One way to reduce overlap is to use a at list of high or intermediate level GO terms, commonly referred to as a slim. But for this to provide useful results, the terms in the slim need to be su ciently descriptive to t the experimental use cases. Rather than use a slim of GO terms, F. Ho mann-La Roche Ltd. (\Roche"), maintains an internal controlled vocabulary (referred to hereafter as RCV) for use in GSEA. The RCV consists of around 360 terms, each of which is mapped to a set of terms from GO, just as a term in a GO slim maps to a set of subclasses and subparts. 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 signi cant 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-speci c 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 speci c 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, such as grouping processes solely by where they occur.
To date, Roche has manually maintained mappings between RCV and GO.
Keeping this mapping up-to-date and complete has become impractical given
the evolution of the GO. Recent developments in the GO make it possible to
automate mappings between the RCV and the GO. The GO has switched its
underlying formalization to Web Ontology Language (OWL2) [
As the RCV is a at list and includes classi cations that are orthogonal to the
classi cation schemes used by the GO, it is not amenable to mapping via ontology
alignment techniques that use ontology structure [
To ensure speed and scalability, we chose to restrict mapping queries to the
EL pro le of OWL2, allowing us to use ELK, a fast, scaleable EL reasoner,
to run queries [
These new, high-level object properties are di cult to name in a way that communicates the meanings of mapping queries clearly. In order to compensate for this, we used scripting to generate human readable descriptions for each mapping query. Compare, for example, the mapping query for the RCV term cannabinoid with its description:
noid
Description : \A process in which a cannabinoid participates, or that
regulates a process in which a cannabinoid participates."
Mapping queries were run using the ELK OWL reasoner [
Mapping queries were selected, tested and the results reviewed against manual mappings to decide which patterns were most appropriate. Once a mapping query was chosen, corrections and/or additions to the GO were made where results were wrong or incomplete. At this point, any clear errors in the manual mapping were blacklisted. Review of automated mappings was then passed to Roche who approved or blacklisted individual classes (see table 1 for an example). When satis ed with the results, the corresponding GitHub ticket was closed, thereby indicating the mapping as approved. Results approved by Roche were combined to produce a new RCV mapping table1. 2.3
GSEA was performed using an open dataset comparing gene expression in adult
liver and embryonic cells of mice [
We developed mapping queries for 308/364 RCV terms. Over a third (104) of the mapping queries were su cient - meaning that no manual maintenance is required. A further 40% of the mappings (148) had 10 or fewer additional manual mappings ( gure 1A) and most of these (114) had fewer than 5.
Mapping queries identi ed many GO terms that were not in the manual mapping ( gure 1B). In some cases (e.g., leukocyte activation), over 1000 new mappings were found. Only 8 automated mappings had blacklisted terms, reecting minor di erences between the meaning of the mapping query and the intended meaning of the RCV term. 56 terms were not mapped. Some were were judged to be semantically equivalent to other RCV terms. The rest were rejected as currently not mappable due to the lack of suitable terms or axiomatisation within the GO. For example, RCV has terms for aerobic and anaerobic metabolic 1 Available from: https://github.com/GO-ROCHE-COLLAB/Roche_CV_mapping/blob/ master/mapping_tables/results/combined_results.tsv 2 Available from http://geneontology.org/page/download-annotations processes, but GO currently has no formal way to group these and no sustainable mechanisms for grouping them manually. Further formalisation of the GO is likely to improve the number of RCV terms that can be mapped. 3.2
We tested the the revised, semi-automated RCV mapping by performing GSEA comparing the transcriptome of adult mouse liver against embryonic mouse cells using a standard GO slim (Figure 2A), the original, manually mapped RCV (RCV man; Figure 2B) and the new partially automated RCV GO mapping (RCV auto; Figure 2C). Pro les of gene enrichment in liver compared to undi erentiated cells are potentially useful benchmarks for the development and testing of stem cell derived liver in vitro systems increasingly employed in toxicology testing. Results were loosely grouped by experts at Roche to provide a preliminary, biologically plausible interpretation. All approaches detected enrichment to gene sets involved in cell division and gene expression in the embryonic sample, but there were dramatic di erences in detection of enrichment in the liver sample.
GSEA with the standard GO slim detects enrichment in the liver to a few sets of genes involved in metabolic processes that are known to be up-regulated in the liver. GSEA with RCV man also detects enrichment to many more metabolic processes that are speci c to or upregulated in the liver, and to a much ner level of detail. It also detects enrichment of genes involved in immune cell related process, consistent with detection of the resident immune system in the liver. The results of enrichment with RCV auto are similar, but provide much more detail. For example, GSEA with RCV auto detects enrichment to gene sets involved in detoxi cation (important for toxicology use cases) and a wider range of immune cell processes. There is also increased overlap between enriched gene sets compared to RCV man, but at a level that is potentially informative (see edges between nodes in 2C). For example, there is overlap between sets of genes involved in both chemotaxis and processes involving types of immune cell that are known to be capable of chemotaxis. 3.3
While GO has extensive axiomatisation linking processes to cells, anatomical structures and chemicals, this is not always complete. In mapping from the RCV to GO we found and corrected over 200 omissions in the axiomatisation. This included missing links from processes to participant cell types, anatomical structures, chemicals, cell components and transcript types. We also found and corrected a number of errors, including errors in axiomatisation of developmental processes that led to incorrect inferences for RCV anatomy terms. 4
This work demonstrates how the logical structure of the GO can be used to
achieve biologically meaningful mappings between GO terms and terms from
external controlled vocabularies de ned with reference to ryps of cells,
chemicals or anatomical structures. Mappings are straightforward to specify and the
reasoning system used is fast and scalable [
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
Fig. 2. GSEA comparing expression between embryo and adult liver using gene sets derived from the generic GO slim (Panel A), manually mapped RCV (panel B), and semi-automated RCV (panel C). Red nodes indicate gene sets enriched in liver compared to embryos. Blue nodes gene sets enriched in embryos compared to liver. The size of the node is proportional to the size of the gene set. Connecting edge thickness is a measure of the number of enriched genes in common between two gene sets. complete automation might be achieved by a di erent query strategy. In some cases, a more complete mapping could be achieved by combining the results of multiple mapping queries. For example, RCV terms for chemical metabolism are all manually mapped to GO terms for both metabolism and transport. A more complete mapping could be achieved by combining the results of separate OWL mapping queries for GO transport and GO metabolic process 3. 4.2
The OWL axioms used to automate RCV mapping to GO can also be used to provide alternative views of the GO and its annotations. This is already re ected in some of the newer functionalities of the GO browsing tool AMIGO, which now displays inferred annotations to cell-types based on axioms in GO recording where processes occur4. 4.3
The system described here was designed to be lightweight and exible, allowing maximum interaction between the designers of RCV at Roche and GO editors with minimal development overhead. Where new terms following mapping query patterns already used, they can be added via the same mechanism.
The system described bears some relationship to TermGenie [