PhenomeNET is a system for disease gene prioritization that includes as one of its components an ontology designed to integrate phenotype ontologies. While not applicable to matching arbitrary ontologies, PhenomeNET can be used to identify related phenotypes in di erent species, including human, mouse, zebra sh, nematode worm, fruit y, and yeast. Here, we apply the PhenomeNET to identify related classes from four phenotype and disease ontologies using automated reasoning. We demonstrate that we can identify a large number of mappings, some of which require automated reasoning and cannot easily be identi ed through lexical approaches alone.
1.1
PhenomeNET [1] was built in 2011 as a system for disease gene discovery and prioritization. PhenomeNET consists of an ontology integrating speciesspeci c phenotype ontologies based on the PATO ontology [2] and relations between anatomical structures and physiological processes, a database of gene-tophenotype associations, and a measure of similarity between sets of phenotypes. Within PhenomeNET, species-speci c phenotype ontologies are combined so that phenotypes observed in di erent species can be compared directly. The main application of PhenomeNET is the prioritization of candidate genes for human diseases by comparing human disease phenotypes to existing gene-phenotype associations derived from model organisms. In particular, human phenotypes associated with a disease can be compared to phenotypes observed in mouse or other model organisms using the integrated PhenomeNET ontology, and similarity between phenotypes can then be used to indicate the genetic basis of a disease. PhenomeNET has been successfully used to nd candidate genes for diseases [1, 3], identify novel pathways [4], and repurpose drugs using mouse model phenotypes [5, 6].
Here, we use the PhenomeNET ontology to identify alignments between phenotypes in di erent species. We present three versions of the PhenomeNET ontology; the rst version consists of the plain ontology using only the axioms provided in the Human Phenotype Ontology (HPO) [7] and the Mammalian Phenotype Ontology (MP) [8]; the second version uses additional lexical mappings and represents them as equivalent class axioms in the ontology; the third version further uses mappings generated by the AgreementMakerLight [9] to generate equivalent class axioms between classes in the PhenomeNET ontology and the Disease Ontology (DO) [10] and the Orphanet Rare Disease Ontology (ORDO) [11]. 1.2
Phenotype classes in the HP and MP ontologies are formally de ned using the Entity-Quality (EQ) pattern [2, 12]. Based on the EQ patterns, a phenotype is decomposed into an a ected entity and a quality that speci es how the entity is a ected. The Entity will usually be a class taken either from an anatomy ontology or a physiology ontology. For example, the phenotype class macroglossia (HP:0000158) describes an anatomical abnormality and is de ned as equivalent to 'has part' some ('increased size' and ('inheres in' some tongue) and ('has modifier' some abnormal)), relying on the entity tongue (from the UBERON anatomy ontology) and the quality increased size (from PATO) in its de nition. The class abnormality of salivation (HP:0100755) is a physiological abnormality and is de ned as equivalent to 'has part' some (quality and ('inheres in' some 'saliva secretion') and ('has modifier' some abnormal)), where saliva secretion is a class from the biological process branch of the GO.
The general pattern for de ning a phenotype class in both the HP and MP ontologies, given Entity E and Quality Q, is to declare them equivalent to 'has part' some (Q and 'inheres in' some E). In some cases, the Entity E is further constrained, e.g., by a location in which a certain process may happen. The \E" classes are generally taken either from the UBERON cross-species anatomy ontology [13] or from the GO. As the use of anatomy and physiology ontologies (UBERON and GO) is shared between MP and HP, it should be possible to integrate both ontologies directly, based on the axiom patterns used to constrain their classes. However, the type of axiom pattern used in both ontologies results in a classi cation that is primarily based on the PATO ontology, as the Quality Q is the main feature that distinguishes di erent classes.
In the PhenomeNET ontology, we rewrite all axioms in HP and MP using a pattern-based approach that allows us to utilize axioms from anatomy and physiology ontologies and enrich the classi cation of phenotype classes [14]. In general, we declare phenotype classes de ned using an Entity E and Quality Q as equivalent to 'has part' some (E and has-quality some Q) and we further add grouping classes that are de ned as equivalent to 'has part' some (('part of' some E) and has-quality some Q). The aim of rewriting the axioms is to base the classi cation of phenotype classes primarily on anatomical or physiological entities instead of the quality, and to utilize the axioms involving parthood in anatomy and physiology ontologies. Crucially, all axioms we generate fall in the OWL 2 EL pro le [15]. The rst version of the PhenomeNET ontology (PhenomeNET-Plain) consists only of these axioms and no additional mappings.
In addition to this knowledge-based approach to linking the HP and MP ontologies, we also add lexical mappings, mappings derived from cross-references in the ontologies [3], and mappings between HP and MP from BioPortal [16]. Each mapping is added as a single equivalent classes axiom to the rst version of the ontology (PhenomeNET-Plain) to generate a version of the PhenomeNET ontology with mappings (PhenomeNET-Map).
Neither version of these ontologies contains the DO or ORDO ontologies, despite there being a signi cant overlap between the four ontologies. Since neither DO nor ORDO contain axioms that follow a similar pattern to the axioms in HP and MP, we rely exclusively on lexical mappings to integrate DO and ORDO. We use the AgreementMaker Light (AML) [9] in its default settings to generate mappings between HP and DO, HP and ORDO, MP and DO, MP and ORDO, and DO and ORDO. We then add an equivalent class axiom for each mapping AML identi es and that has a score by AML over greater than 0:7. The resulting ontology contains HP, MP, ORDO, and DO, and can be used to generate mappings between these ontologies.
All versions of the PhenomeNET ontology contain the classes from the HP and MP ontologies as well as the subclass axioms between named classes asserted in these ontologies. Furthermore, the PhenomeNET ontology imports the ChEBI [17] and Mouse Pathology [18] ontologies using an OWL import statement. Additionally, PhenomeNET includes all classes from the UBERON anatomy ontology [13], the Gene Ontology [19], the BioSpatial Ontology [20], the Zebra sh Anatomy ontology [21], the PATO ontology [2], the Cell Ontology [22], and the Neuro-Behavior Ontology [23]. However, these ontologies are not directly imported but rather pre-processed so that all disjointness axioms from these ontologies are excluded while all other axioms contained within them are included in the PhenomeNET ontology. The aim of this pre-processing step is to avoid unsatis able classes due to di erent conceptualizations between anatomy and phenotype ontologies, or within anatomy ontologies (Zebra sh Anatomy and UBERON).
Mappings between ontologies included in PhenomeNET are generated using the ELK reasoner [24]. We use ELK to classify the PhenomeNET ontology and identify pairs of equivalent classes C1 and C2 that belong to the ontologies to be aligned. These constitute equivalent class mappings. Furthermore, subclass and superclass mappings are generated through queries for sub- and superclasses using ELK.
Ontology HP-MP HP-MP+mappings HP-MP+DO-ORDO Within PhenomeNET, we use an ontology consisting only of the (rewritten) axioms in MP and HP as well as equivalent class axioms derived from explicit mappings between HP and MP (expressed as xref annotation properties). For the evaluation, we further used the AML [9] to generate additional mappings. The AML mappings were generated using the default settings of AML with a con dence cuto of 0:7. In the case of DOID and ORDO mappings we additionally included 18 mappings derived from BioPortal. Our systems relying on these mappings were submitted as separate submissions.
Initially, we developed our matching system to take into account not only the direct sub- and super-classes, but also all inferred classes. We modi ed our system to output only the most speci c mappings instead for the evaluation; Table 2 shows both the number of direct and inferred mappings. Our submission consists of two modules: PhenomeNetBridge and PhenomeNetMatcher. The PhenomeNetBridge module wraps the SEALS infrastructure for the evaluation, and the PhenomeNetMatcher module performs the mappings, using one of three ontologies. Source code for the matching system, including parameter les, and the generated alignments, are available at http://github.com/bioontology-research-group/OAEI2016. Code to generate the PhenomeNET ontology is available at https://github.com/bio-ontology-research-group/phenomeblast/tree/master/ xphenotypes.
The PhenomeNET ontology is primarily intended to integrate the HP and MP ontologies. Using the axioms in the ontology alone (PhenomeNET-Plain submission), we identify 745 equivalent classes between the HP and MP ontologies (see Table 2). These correspond to a recall of 40.8% with respect to the reference mappings provided (see Table 3). Additionally, a large number of sub- and super-class mappings can be identi ed based on querying the ontology using the ELK reasoner [24] for sub- or super-classes in the two ontologies.
The number of pairs of equivalent classes identi ed increases to 1,536 when adding explicit mappings derived from AML. Of these, 370 are generated both by automated reasoning and are included in AML, 791 are generated from the AMLderived equivalent classes axioms, and 375 could only be derived through the automated reasoning. Total recall with respect to the reference mappings is 100% in this version of PhenomeNET. Additionally, we observe an improvement in the number of equivalent class mappings when adding the ORDO and DO ontologies to the PhenomeNET ontology. The increase in mappings (from 1,536 to 1,582 classes) is a result of additional inferences obtained from adding the mappings from HP and MP to ORDO and DO, and combining them with the axioms in the PhenomeNET ontology. For example, we infer a new mapping between decreased IgG level (MP:0001805) and agammaglobulinemia (HP:0004432) based on the equivalence axioms between both classes and agammaglobulinemia (DOID:2583) generated by AML (based on the shared synonym \hypogammaglobulinemia" between the class in DO and MP). Table 3 summarizes our results with respect to the reference mappings provided in the challenge. 2.2
PhenomeNET is primarily designed for ontologies that follow the Entity-Quality de nition pattern based on the PATO ontology. Neither ORDO nor DO follow this pattern, and ORDO and DO are primarily included in the PhenomeNET ontology through equivalent class axioms based on lexical mappings generated by AML. We achieve a recall of 99.9% with the PhenomeNET-Full ontology. Notably, the mappings we generate are increased by including HP and MP. For example, we identify a mapping between mandibulofacial dysostosis (ORPHANET:155899) and treacher collins syndrome (DOID:2908), based on common AML-generated mappings to mandibulofacial dysostosis (HP:0005321). 2.3
In order to carry out the nal evaluation, the OAEI utilized the SEALS infrastructure executed in a Ubuntu Laptop with an Intel Core i7-4600U CPU @ 2.10GHz x 4 and allocating 15Gb of RAM. The system carried out the evaluation according to following criteria: { Precision and Recall with respect to a voted reference alignment automatically generated by merging/voting the outputs of the participating systems. { Recall with respect to alignment manually generated. { Manual assesment of a subset of generating mappings. { Performance in other tracks.
Di erent mappings were used to evaluate the participating systems: i) Silver standard with vote 2, ii) Silver standard with vote 3, iii) manually dataset and manual assessment. In the rst dataset, PhenomeNET including all mappings reached an F-measure of 0:82 in the HP-MP task, and 0:89 in the DO-ORDO task. In the second evaluation, although the system PhenoMP was able to nd the largest number of mappings in HP-MP task, it reached an F-measure of 0:76 in the HP-MP task and 0:94 in the DO-ORDO task. When evaluating against manually created mappings, PhenomeNET achieved a recall of 0:897 in the HPMP task but could not generate any new mappings between DO and ORDO. For this task, PhenomeNET achieved a precision of 1:0 in the manual assessment of a subset of the generated mappings. 3 3.1
PhenomeNET is a system to match phenotypes; as such, it is not a system that can be applied to match ontologies in general. The axiom-based approach in PhenomeNET can be applied to any ontologies that utilize PATO and the Entity-Quality de nition patterns [2]. In particular, PhenomeNET can not only be used to integrate MP and HPO, but also has been used to further integrate yeast, y, worm, slime mold, and sh phenotypes [1, 25]. Furthermore, the combination of semantic matching (using automated reasoning) and lexical matching in PhenomeNET mitigates some of the limitations of using lexical approaches alone, and we demonstrate this by inferring several hundred mappings between HP and MP that cannot be inferred using AML.
However, relying on manually created axioms also has several limitations. In particular, the axioms are created by domain experts, and only about half the classes in MP and HP are constrained by an Entity-Quality based axiom. Furthermore, the quality of the axioms is di cult to assess, and there are distinct di erences between HP and MP in how the classes are constrained. 3.2
One of the main limitations in PhenomeNET is the need for manually created axioms that constrain classes in phenotype ontologies. A possible solution to this approach would be to generate phenotype ontologies fully automatically using anatomy and physiology ontologies as templates and applying the axiom patterns we use in the PhenomeNET [26].
Another limitation of PhenomeNET is the reliance on OWL 2 EL which limits the expressivity of axiom patterns. The choice is mainly due to the size of the PhenomeNET ontology and the complexity of reasoning. However, more complex axiom patterns would enable more comprehensive classi cation of phenotypes involving absences and abnormalities [14]; experiments with an updated ontology will likely require improvement in OWL reasoning technologies. 4
We have developed an ontology matching system for disease and phenotype
ontologies. We generated three di erent version of the PhenomeNet ontology, each
with di erent information and ontologies included. PhenomeNET is primarily
based on deductive inference and automated reasoning, and while it can utilize
lexically derived mappings in the ontology generation process, it does not on
its own include any lexical matching algorithms. Our results demonstrate that
a combination of lexical and semantic approaches may improve upon mappings
between ontologies generated using only one of these methods.
7. Kohler, S., Doelken, S.C., Mungall, C.J., Bauer, S., Firth, H.V., Bailleul-Forestier,
I., Black, G.C.M., Brown, D.L., Brudno, M., Campbell, J., FitzPatrick, D.R.,
Eppig, J.T., Jackson, A.P., Freson, K., Girdea, M., Helbig, I., Hurst, J.A., Jahn, J.,
Jackson, L.G., Kelly, A.M., Ledbetter, D.H., Mansour, S., Martin, C.L., Moss, C.,
Mumford, A., Ouwehand, W.H., Park, S.M., Riggs, E.R., Scott, R.H., Sisodiya,
S., Vooren, S.V., Wapner, R.J., Wilkie, A.O.M., Wright, C.F., Vulto-van Silfhout,
A.T., Leeuw, N.d., de Vries, B.B.A., Washingthon, N.L., Smith, C.L., Wester eld,
M., Scho eld, P., Ruef, B.J., Gkoutos, G.V., Haendel, M., Smedley, D., Lewis, S.E.,
Robinson, P.N.: The human phenotype ontology project: linking molecular biology
and disease through phenotype data. Nucleic Acids Res 42(D1) (2014) D966{D974
8. Smith, C.L., Goldsmith, C.A.W., Eppig, J.T.: The mammalian phenotype ontology
as a tool for annotating, analyzing and comparing phenotypic information. Genome
Biol 6(
Nature Genetics 25(
Reasoning 53(