Phenotype ontologies and cross-species analysis for translational research

Abstract

The use of model organisms as tools for the investigation of human genetic variation has significantly and rapidly advanced our understanding of the aetiologies underlying hereditary traits. However, while equivalences in the DNA sequence of two species may be readily inferred through evolutionary models, the identification of equivalence in the phenotypic consequences resulting from comparable genetic variation is far from straightforward, limiting the value of the modelling paradigm. In this review, we provide an overview of the emerging statistical and computational approaches to objectively identify phenotypic equivalence between human and model organisms with examples from the vertebrate models, mouse and zebrafish. Firstly, we discuss enrichment approaches, which deem the most frequent phenotype among the orthologues of a set of genes associated with a common human phenotype as the orthologous phenotype, or phenolog, in the model species. Secondly, we introduce and discuss computational reasoning approaches to identify phenotypic equivalences made possible through the development of intra- and interspecies ontologies. Finally, we consider the particular challenges involved in modelling neuropsychiatric disorders, which illustrate many of the remaining difficulties in developing comprehensive and unequivocal interspecies phenotype mappings.

Figure 1. Interspecies phenotype mapping strategies.

This review highlights three major methodologies to identify phenotypes in the mouse that are relevant to a human disease. (A) Classical approach. A mouse model is made or identified that possesses a genotype equivalent to a penetrant mutation that in human underlies the disease of interest (termed construct validity). The mouse model is examined for phenotypes that resemble those that define the human disorder (face validity). (B) Phenolog mapping. A group is formed containing candidate genes for a disease of interest. The respective mouse models for the orthologues of these genes are then examined for any unusually overrepresented phenotypes among them and these phenotypes (termed phenologs) are deemed relevant to the disease. (C) Direct phenotype mapping. Given the phenotype(s) that describe a human disease, the corresponding phenotypes in mouse are inferred by means of computational reasoning using interspecies phenotype ontology analysis. In the example shown, the HPO term Aortic stenosis is defined on the basis of the PATO term constricted and aortic valve (term from the Foundational Model of Anatomy ontology of human anatomy [35]). Similarly, the MPO term aortic valve stenosis is defined using the same PATO term constricted and aortic valve (term from the Mouse Anatomy ontology [77]). Since both the Mouse Anatomy and FMA terms for aortic valve are children of the cross-species anatomy ontology (Uberon [40]) term for aortic valve, automatic reasoning places the HPO term Aortic stenosis and the MPO term aortic valve stenosis in the direct vicinity of one another in a cross-species phenotype ontology . Therefore, these terms display a high semantic similarity to one another.

Figure 2. Phenotype ontologies.

Phenotype ontologies (an excerpt from the Human Phenotype Ontology is shown here) consist of thousands of terms describing phenotypes arranged in a hierarchical system of subclasses and superclasses. The structure of an ontology enables annotation propagation whereby more specific phenotypic terms are also described by more general parent terms, and thus all ancestral terms. The terms are related to one another by subclass (“is a”) relations, such that the ontology can be represented as a so-called directed acyclic graph. The terms themselves do not describe any specific disease. Instead, annotations to terms are used to state that a certain disease is characterised by a certain phenotypic feature.

Figure 3. Predicting human genotype-phenotype relations from functional genomics data.

The mouse phenotypes associated with the orthologues of human genes are a better predictor of genes that share human phenotypes than other popular gene annotations of the same genes, such as GO or KEGG. As both GO and KEGG include information derived from multiple sources, including annotations from the mouse, the success of the mouse phenotypes is likely due both to the genetic relevance of the mouse models and the fact that human and mouse phenotypic annotations both describe abnormalities (see Figure 1C). Resnik's measure, together with the GraSM approach , was used to calculate the similarity of terms organised in these hierarchical ontologies, defining the semantic similarity between any two terms as the average information content of their disjunct common-ancestor terms. Gene pairs were ordered by their semantic similarity scores based on either the human KEGG pathway annotations (pink circles), human GO biological process (grey circles), or MPO annotations to genes (blue circles). For each of KEGG, GO, and MPO annotations, gene pairs were ordered in decreasing annotation similarity and grouped into bins of 2,000, and then the median semantic similarity score between gene pairs' Human Phenotype Ontology annotations was calculated. The dashed line marks the degree of similarity expected from pairs of random genes.