The Human Phenotype Ontology: a tool for annotating and analyzing human hereditary disease

Abstract

There are many thousands of hereditary diseases in humans, each of which has a specific combination of phenotypic features, but computational analysis of phenotypic data has been hampered by lack of adequate computational data structures. Therefore, we have developed a Human Phenotype Ontology (HPO) with over 8000 terms representing individual phenotypic anomalies and have annotated all clinical entries in Online Mendelian Inheritance in Man with the terms of the HPO. We show that the HPO is able to capture phenotypic similarities between diseases in a useful and highly significant fashion.

Figure 1

The Human Phenotype Ontology The HPO term Bilateral congenital hip dislocation and all paths to the root that emanate from this term are shown. Some of the annotated disease entries from OMIM, as well as the total number of annotated diseases, are shown next to the terms. Note that because of the true-path rule, a disease that is directly annotated to a specific term is also implicitly annotated to all ancestors of that term. For instance, Ehlers Danlos syndrome type VII is directly annotated to Bilateral congenital hip dislocation and is thereby implicitly annotated to Abnormality of the hips, Dislocations, and the other terms shown in the figure.

Figure 2

Applications of the HPO (A) Visualization of the human phenome. Each of 727 diseases listed in OMIM for which a disorder class was defined is shown as a node in the graph and is colored according to membership in a set of 21 predefined disorder classes, defined on the basis of the physiological system. The organic layout algorithm of Cytoscape was used for showing the clustered structure of the phenotypic network. Connections between nodes are shown starting from a similarity score of 4.5, whereby the thickness of the connection reflects the degree of phenotypic similarity. Abbreviations are as follows: CV, cardiovascular; derma, dermatological; endo, endocrinological; heme, hematological; immuno, immunological; metab, metabolic; neuro, neurological; ophth, ophthalmological. (B) Analysis of randomized phenotypic networks. In order to estimate the probability that this result could be due to chance, we created 10,000 random networks in which edges were randomly rewired 2000 times. The mean network score of the random nets was 0.182 ± 0.0098. Thus, the actual score of 0.645 was 47.2 standard deviations above the mean random score. (C) Searching the HPO. All 2116 diseases listed in OMIM with at least six HPO annotations were identified and included in the analysis. For each disease, between 1 and 6 of the most specific terms to which the disease is annotated in the HPO were used for the search (“Annotated terms”). The set of clinical features determined in this way was then used for querying the entire set of OMIM diseases for the best match. The average rank of the original disease among the diseases identified by the search for different proportions of removed terms is shown. In separate experiments, each of these terms was mapped to a parent term or 50% unrelated (“noise”) terms were added (rounded up for odd numbers of terms; i.e., one noise term was added to searches with one term, two noise terms to searches with three terms, and three noise terms to searches with five terms).