Our understanding of gene function is often informed by comparing the phenotypic consequences of mutation with the canonical “wild-type” in a single organism, as well as between mutants of orthologous genes in different organisms. In particular, model organisms have provided great insight into gene function in humans. The importance and need for automating these cross-species comparisons has become imperative as large-scale mutagenesis screens are conducted in model organisms. A fundamental roadblock for analysis is, however, the lack of a computationally tractable method for describing phenotypes that is applicable across multiple domains of biological knowledge and species (for example, see 
). Not only does each model organism have its own vocabulary for describing the phenotypic consequences of mutation, but these vocabularies are usually tied to the particular anatomies or physiologies of the organism. Often these descriptions are recorded as free text, and although wonderfully expressive, free text remains difficult to reliably compare with computational methods. For example, a computer program would not be able to recognize the fact that there is a significant similarity between the PAX6
mutations that result in “small eyed” mice, “opaque cornea” in humans, a “malformed retina” in zebrafish, and “eyeless” Drosophila
Current methodologies traditionally identify animal models on the basis of sequence orthology between the mutant animal model and a human gene. For example, Schuhmacher et al. recently developed a mouse model of human Costello syndrome (OMIM: #218040), which is a neuro-cardio-facio-cutaneous developmental syndrome resulting from mutations in the H-RAS
. The mouse H-Ras
gene was mutated in the orthologous position as in Costello patients, and the resulting phenotype recapitulates the disease. Occasionally, spontaneous models can be identified by the observation of symptoms reminiscent of human disease, for example the fat aussie
mouse develops obesity, type 2 diabetes, and male infertility. This phenotype is similar to human Alström syndrome, which is caused by mutation in the ALMS1
. Sequencing and further characterization of fat aussie
revealed a mutation in Alms1
, and fat aussie
is emerging as a good animal model for understanding Alström syndrome and the function of cilia-localized Alms1 
. These examples for identifying animal models of disease relied on knowledge of the genetic basis of the human disease, but there are many human diseases for which it is not yet known. If a researcher could compare human model organism, and even ancestral phenotypes directly, they would have a mechanism to more rapidly identify candidate genes and models of disease.
Model organism communities benefit from centralized collections of curated research, where a scientist can search for extensively cross-referenced gene expression, phenotype, and genomic data, referred to as “model organism databases” (MODs). Research in the field of human biology suffers because there is no equivalent resource for the human research community, and linking these diverse datasets requires searching many detached resources. There are, however, several valuable data resources for human phenotypic data, including the Online Mendelian Inheritance in Man
published by the National Center for Biotechnology Information (NCBI). OMIM contains more than 19,000 records, divided between genes and phenotypes/diseases. Approximately 53% of the gene records have detailed allelic variant descriptions and/or general clinical synopses, while 43% of phenotype/disease records have a known molecular basis. OMIM is a text-based resource, and retrieval of information suffers from this fact, as the Entrez searches in show. For an individual researcher wanting to know which human mutations may result in an increase in bone size, or a computer script mining OMIM data, free text annotations do not provide the rigor necessary for querying. While successful mining of the literature to relate genes to phenotypes has been shown 
, it does not provide a mechanism to compare phenotypes directly.
One of the most revolutionary tools for the biologist has been the ability to compare sequences using algorithms such as BLAST 
, which allows one to quantitatively assess similarity between one or more sequences. However, the genetic basis of a disease is often unknown, and in this case a sequence-comparison tool is of no use to identify sequence mutations. If descriptions of phenotypes were based on a common controlled vocabulary—an ontology
—they would be structured such that algorithms could be written to compare phenotypes computationally. One of the benefits of using ontologies is the ability to use general-purpose logical inference tools called reasoners (for example, see 
). Reasoners can assist in query answering and analysis. As an example, consider two different queries, one to find genes expressed in the ZFA:gut
, and the other to find genes expressed in the ZFA:epithelium
(we write ontology terms prefixed with the name of the ontology; see Materials and Methods
for further explanation). We would expect both of these searches to return annotations to the ZFA:intestinal epithelium
, because the intestines are a part_of
the gut, and the intestinal epithelium is_a
type of epithelium (). Analogous to the nucleic and amino acid alphabets and distance matrices used in the BLAST algorithm, ontology terms and their relationships to one another can be used to group and compare phenotypic and gene expression data and can be utilized for cross-species phenotype analysis.
A phenotype can be defined as the outcome of a given genotype in a particular environment (for review see 
) and can be described using ontologies to facilitate comparisons. A description of an individual phenotypic character can be recorded using a bipartite “EQ” (Entity + Quality) method, where a bearer entity (such as an anatomical part, cellular process, etc.) is described by a quality (such as small, increased temperature, round, reduced length, etc.). The EQ method is sufficient for the description of many phenotypes, provided the source ontologies are rich enough. The entity terms may be structures from any anatomy ontology, or biological processes, cellular components, or molecular functions from the Gene Ontology (GO) 
. The quality terms come from the Phenotype and Trait Ontology (PATO), which is designed to be used in combination with species-specific anatomical ontologies or other cross-species entity ontologies (see, for example, 
). For instance, a Drosophila
“redness of eye” phenotype could be described using the terms “red” from PATO and “eye” from the Fly Anatomy ontology (FBbt) into the EQ statement EQ = FBbt:eye + PATO:red
. The EQ method has been extended to include related qualities and additional entities, and with a post-composition approach to describe more granular entities. Many MODs already utilize community-specific anatomy ontologies, in addition to GO, for annotation of gene expression and/or phenotype data 
, and these methods are described in detail elsewhere 
. Ontological reasoning can also be applied to EQ descriptions, just as for a single ontology, because they too represent nodes in a graph structure. For example, queries for cranial cartilage position
should return genotypes that have the phenotype ZFA: ceratohyal + PATO:mislocalised_ventrally
. Similarly, queries for superstructures of the ceratohyal cartilage, such as cranial cartilage, should also return these genotypes ().
Subsumption reasoning EQ descriptions.
Any EQ description can be combined with other EQ descriptions and data, such as genotype, environment, and stage identifiers from other databases or ontologies, to fully express the phenotypic state of an individual or group. For example, one could record the zebrafish phenotype EQ
= ZFA:median fin fold + PATO:attenuate
at the embryonic stage ZFS:26-somite
with genotype fbn2bgw1/gw1
(AB) (defined in the Zebrafish Information Network, ZFIN). With this method, phenotypes can be recorded using multiple ontologies in a highly expressive and finely detailed manner while maintaining correct logic and computability.
Existing computational tools are inadequate to store and analyze this ontology-based phenotype annotation data in a generic, species-neutral way. In particular, there is a lack of tools for the cross-species comparisons needed to identify gene candidates and animal models of disease. Many existing algorithms have been developed and tested using the GO to measure the semantic similarity of annotations and provide a good starting point for analysis (for example, see 
). It was unclear how well these algorithms would work for analyzing datasets using a combination of ontologies. Additionally, cross-species comparisons would not be possible because there were no links between the various anatomical ontologies. Schlicker and Albrecht 
suggest an information content (IC)–based approach to analyzing phenotypic profiles made with multiple ontologies, although they only tested their results with annotations made with the species-neutral GO. Their FunSimMat tool requires a specific list of proteins to compare and therefore does not provide a means to comprehensively search for phenotypically similar genes. PhenomicDB 
is a cross-species resource that has pulled together annotations from diverse resources and mined free-text phenotypes to provide “phenoclusters” of phenotype-related genes. However, their analysis did not make use of the relationships in the source ontologies. Although known interacting proteins were clustered together, they note that their resulting “phenoclusters” tended to be species-specific due in large part to the community-specific terminologies that were used in the annotations, and not necessarily due to the underlying biology. These existing methods were insufficient for our needs because they were either free-text based or used a limited set of ontologies for annotation, and because they lacked a framework to integrate and compare anatomical entities between organisms. They also lacked metrics for determining significance in similarity calculations. Lastly, apart from the querying aspect, none included a species-neutral method for recording phenotypes de novo.
By annotating phenotypes using this EQ method, together with appropriate computational analysis tools, we have a unique opportunity to standardize and query phenotypic data in a rigorous and illuminating manner. In this study, we tested the hypothesis that EQ annotation of disease phenotypes will facilitate the discovery of new genotype-phenotype relationships within and across species. We EQ-annotated 11 human disease genes from free-text OMIM descriptions with Phenote software 
to provide a dataset for cross-species comparison. We compared these annotations to annotations of the mouse and zebrafish orthologs, which required the development of a cross-species unifying ontology (UBERON) to provide a bridge between different anatomy ontologies. We also developed new, and extended existing, metrics for measuring the phenotypic similarity between genes. We assessed their relative performance through analysis of known signaling pathways and genetic interactions and show that these data can be queried and compared by phenotype alone
to identify biologically meaningful similarities. Furthermore, these annotations provide a resource for a better understanding of existing disease phenotypes. We conclude that this method can facilitate the discovery of new genotype-phenotype associations within and between species.