Research on model organisms is crucial for discovering the function of genes and DNA elements and for understanding the phenotypic effects of mutations on these genes, which is leading to a better understanding of the pathobiology of human disease ^{1, 2} . The amount of phenotypic information derived from targeted mutations and hypothesis-driven studies is increasing rapidly, and is now being further augmented by high-throughput international efforts to systematically analyse the effects of genomic variation on model organism phenotypes. For example, the International Mouse Phenotyping Consortium (IMPC ³ ), is undertaking systematic phenotyping studies of the knockouts generated by the International Knockout Mouse Consortium (IKMC ⁴ ). This means that there will soon be structured phenotype data for loss-of-function mutants for every protein-coding gene in the mouse. Similar approaches are being taken in zebrafish ( Danio rerio) by the Zebrafish Mutation Project (ZMP, http://www.sanger.ac.uk/Projects/D_rerio/zmp/) and the data is being made available through the Zebrafish Model Organism Database (ZFIN ⁵ ).

Model organism phenotype/genotype datasets are extremely valuable as they can provide clues to human gene functions and involvement in disease processes where no data is available for the human ortholog. At the time of writing, 2,358 human genes are associated with Mendelian phenotypes, but more importantly there are 5,492 human genes with no such phenotype associations, where an orthologous mouse or zebrafish gene does have phenotype data (Data obtained by analysing the file HSgenes_crossSpeciesPhenoAnnotation.txt from http://purl.obolibrary.org/obo/hp/uberpheno/). We have previously demonstrated the power of this approach in determining likely pathogenicity of genes within the intervals of recurrent copy number variation (CNV) diseases ⁶ and it can be applied much more widely in, for example, prioritizing candidate genes identified through human genome wide association studies (GWAS) ^{7, 8} . Historically, a major problem has been the lack of common semantics across databases, with each project using some combination of free-text descriptions or in-house vocabularies. Thus, phenotype information is not easily integrated across different species. This inhibits comparisons based on phenotype alone, and where orthology is useful phenotypic comparisons cannot be used to their full potential. This is made even more complicated by different conceptualizations of phenotypes in different species and the impact of species-specific anatomies. As the ability of investigators to mobilise this growing collection of model organism data has become more important, it is crucial to develop appropriate ontologies and computational strategies to describe phenotypes such that phenotype descriptions can be objectively related to each other, both within and between species. This becomes even more important as the divergence between the number of human genes with phenotype information and the amount of systematically phenotyped model organism genes is expected to increase in the near future due to high throughput-screens ¹ .

The application of controlled vocabularies and ontologies has accelerated over recent years; the Gene Ontology (GO ⁹ ) being probably the most successful example in the field of biomedical ontologies. Many other ontologies exist, each of which has been developed for a specific domain in biomedicine. Now a major goal is to increase semantic and syntactic interoperability between those ontologies (e.g. the Open Biomedical Ontologies (OBO) Foundry ¹⁰ ). One approach is to develop ontologies by defining complex ("pre-composed") classes in terms of other more elementary (atomic) classes (building blocks) that are species-agnostic. If several ontologies make use of shared building block ontologies, interoperability can be facilitated across a larger domain. For example ontologies that contain classes concerned with DNA-replication in different organisms or cells should refer to a shared class representing DNA-replication-process, enabling computers to detect that the same class is referenced.

We have previously shown how phenotype information can be linked and used in cross-species phenotype analyses ^{11– 15} . A crucial part of this strategy is the use of logical definitions to render ontology terms in a way that is computable. Recently, logical definitions of terms representing classes of phenotypic deviations have been developed by several groups. Developers of OBO Foundry ontologies, such as the GO ¹⁶ , the Mammalian Phenotype Ontology (MPO ¹⁷ ), the Human Phenotype Ontology (HPO ^{18, 19} ), the Worm Phenotype Ontology ²⁰ , and also the Cell Ontology ²¹ , are now creating logical definitions of their ontology-classes using terms from other building block ontologies. In this effort the Phenotype, Attribute and Trait Ontology (PATO), an ontology of phenotypic qualities, is a key tool ^{19, 22} . Examples for building block ontologies that are used for the representation of classes of phenotypic abnormalities are given in the upper part of Table 1.

Given that logical definitions exist for most classes of an ontology, automatic reasoners can be applied. These implement algorithms for computing the logical consequences that can be inferred from a set of asserted axioms. An example can be seen in Figure 1a), where logical definitions are used to automatically infer that Hypoglycemia is a subclass of Decreased aldohexose concentration (blood) based on the asserted subclass relationship between 'glucose' and 'aldohexose' in ChEBI. This means that reasoners are able to use computable, logical definitions to infer the positions of classes in a subsumption hierarchy. Thus, those definitions can be helpful tools for the development and maintenance of ontologies ^{16, 23} .

Although several methods, ideas, and applications on cross-species phenotype integration have been presented before ^{11, 12, 16, 24, 25} , accessing such data resources has been complicated by the lack of consistent documentation and distribution of data across heterogenous resources. For example, some ontologies are provided in the Web Ontology Language (OWL ²⁶ ) and others in the Open Biomedical Ontologies (OBO) format. Although the OBO-format focuses especially on human readability and ease of parsing, OWL is often needed to enable complex reasoning tasks. Unfortunately, the power and complexity of OWL may discourage some researchers.

For example, the OWLSim package ( http://owlsim.org) provides the ability to execute a number of standard semantic similarity techniques. Although access to the results of OWLSim in phenotype analyses is available ( ²⁵ , http://www.mousemodels.org), there is at the moment no single set of gene annotations linked to a single integrated ontology.

The Uberpheno-ontology is similar to the "phene.owl" ontology distributed as part of the phenomeblast-project ( http://code.google.eom/p/phenomeblast/) and generated as part of a phenotype data analysis executed within PhenomeNET ²⁴ . These two ontologies differ in a number of characteristics. The first characteristic is the underlying OWL model, and the set of external ontologies that are brought in to enrich the ontology - it is not yet clear how far the OWL model or some of these external ontologies affect the resulting structure of the ontology. Also it is likely that both Uberpheno and "phene.owl" will converge on the same model and a standard set of imported ontologies. The second characteristic is the breadth of species covered, with "phene.owl" including fly, worm and yeast; in contrast, Uberpheno focuses on human, mouse and zebrafish, yielding a smaller more focused ontology. Further investigations are required to determine the extent to which the adding of more distant organisms help or hinder analyses. Another difference is that Uberpheno is intended for a wide range of biomedical researchers, some of who may be unfamiliar with OWL or OWL reasoning.

Our objective here is to provide an OBO-format ontology ( Uberpheno), which we update at regular intervals and which can easily be used for downstream analysis, e.g. by applying semantic similarity measures ²⁷ or gene set enrichment analyses ²⁸ . Of similar importance are the data that link into such an ontology by means of the annotation relation. To the best of our knowledge, no single integrated cross-species ontology together with annotation of all genes in human and model organisms (here mouse and zebrafish) has been made easily available for researchers and kept up-to-date on a regular basis.

All of the above described methods are integrated into a single pipeline. This pipeline automatically downloads required files, preprocesses the data and applies a reasoning procedure to the obtained set of ontology classes. The ontologies used to construct Uberpheno are summarized in Table 1.

The construction pipeline is set up as a job in our continuous integration system accessible at http://compbio.charite.de/hudson, which is already used for data related to the HPO ⁴² . The job (called hpo.ontology.uberpheno) is configured to run once a week, ensuring that the most recent version of all ontologies and annotation files are used. Only stable releases of the generated files are made available to the users and errors are immediately forwarded to us via email. The generated build artifacts are available at http://purl.obolibrary.org/obo/hp/uberpheno/, whereas the file crossSpeciesPheno.obo contains the cross-species phenotype ontology in OBO-format. The resulting ontology has a light footprint (3.5 MB) and can easily be explored by using tools such as example OBO-Edit ⁴³ . Note that only phenotype classes are present in the ontology and classes from the referenced building block ontologies are filtered out. Each build also generates the file HSgenes_crossSpeciesPhenoAnnotation.txt, which contains the annotation of all human genes to terms of HPO, MPO, and ZP. A summary of the data contained in the two files is given in Table 4.

An excerpt of the Uberpheno ontology is shown in ( Figure 1b), demonstrating how the phenotype descriptions from different ontologies are combined and automatically organised into a single, integrated hierarchy. For instance, the fact that the mouse term posterior microphthalmia is inferred to be a subclass of the human term Bilateral microphthalmos can be used to transfer the information that the mouse gene PRSS56 is known to cause Bilateral microphthalmos. This implies that querying the cross-species ontology for genes related to Bilateral microphthalmos will return the human gene TCOF1, the mouse gene PRSS56 and the zebrafish gene tcf7l1a.

In total, the annotation file contains approx. 235,000 annotations of human genes with phenotype classes (see Table 4). For example the human gene TCF7L1 is associated with the zebrafish phenotype abnormal(ly) hypoplastic eye because the ortholog zebrafish gene ( tcf7l1a, ZDB-GENE-980605-30) is annotated with this phenotype. Thus, the generated file HSgenes_crossSpeciesPhenoAnnotation.txt contains the line:

83439;TCF7L1;abnormal(ly) hypoplastic eye (ZP:0003395);tcf7l1a (ZDB-GENE-980605-30/ZEBRAF)

The phenotype resources for mouse, zebrafish, and human are used by several research projects ^{44– 46} .

The problem of comparing phenotypes between species can be overcome by using formal logical definitions that make use of species agnostic ontologies together with a multi-species anatomy ontology, Uberon. The approach to implementing the paradigm that we report in this paper constructs a single, integrated, cross-species phenotype ontology, Uberpheno, based on the logical definitions of human and the main model species, mouse and zebrafish. The resulting construct is continuously updated and automatically constructed as the constituent ontologies are updated and augmented, making it a dynamic and current resource available to the community.

Increasingly model organism data are being used for gene set enrichment, pathogenicity prediction and semantic similarity analyses ²⁷ and the high throughput phenotyping projects newly underway promise rich genome-wide phenotypic coverage within a decade. This will complement the new initiatives to systematically gather high precision, formally coded, phenotype data from clinical studies ⁴⁷ . The promise that all this data holds can only be realized if the informatics tools are available to handle and analyse this rich resource and we believe that Uberpheno is an accessible and widely applicable resource with which this may be achieved.