Merge process

The goal was to maintain meaningful anatomical groupings for use by biologist end-users, as well as to enhance the description logic axiomatization to support more advanced reasoning that had been limited or absent in the original four source ontologies. This merger consisted of two phases. First, classes from one or more source ontologies that had similar text definitions, labels, synonyms to Uberon classes, or that already had equivalence or taxon-specific subclass relations to Uberon (see Methods), were manually compared. Where needed, the inclusion criteria in the Uberon class were broadened, or in some cases, narrowed based on more expert definitions coming from one of the source ontologies. Second, classes from any source ontology that were not represented in Uberon were placed in an OWL-formatted extension ontology, which imports the core Uberon and extends classes in the core. As an example of the first situation, the class ‘pectoral girdle skeleton’ was previously present in Uberon, as well as in VSAO [16] and a few other source ontologies. Because the definitions were fully expert-vetted, the central representation in Uberon was adjusted to conform to VSAO when these classes were imported. In contrast and as an example of the second situation, the class ‘extracleithrum’ was not previously represented in Uberon, and thus was given a new identifier in the Uberon namespace (UBERON:4200022)¹ and placed in the extension ontology (see Figure 1).This arrangement enabled distribution of editing rights for the more taxonomically specific classes to the domain experts. The contribution and evolution of each source ontology is highlighted in Figure 2, where the pre-merge cross-references (Xrefs) are highlighted, as well as the temporal relationships between these source ontologies and Uberon. An example of how the various classes relate to one another is shown in Figure 1.

Challenges

Reconciling terminological differences from different domains

The process of integrating AOs built for different purposes also highlighted inconsistencies in terminology between and within zoological and/or medical nomenclatures. A good illustration of this problem comes from the terms used for different regions of the limb. In the comparative morphology literature, “-podium” terms are commonly used in reference to the vertebrate appendicular skeleton [25, 26]. These terms were introduced by Haeckel [27] in his 1895 treatise, “Systematische Phylogenie,” to refer to skeletal elements and their developmental anlagen; they were not originally intended to refer to composite bone/flesh limb segments. In the developmental literature, however, authors sometimes use these terms to refer to “limb segments” (e.g. [26]), usually in the context of the limb buds [26–30]. To reconcile these different uses, we label terms such as ‘skeleton of manual acropodium’ to refer to the skeleton and ‘manual acropodium region’ to refer to the limb segment. Another challenge is the use of some terms (such as acropodium) to refer to different structures in different contexts. In some papers “acropodium” only refers to the phalanges (e.g. [31]), but in others it refers to the entire manus skeleton excluding the mesopodium (e.g. [32, 33]). In keeping with the definitions created by Haeckel we decided to use ‘acropodium’ (or ‘acropodial skeleton’) for the phalanges and introduce a new term, the ‘digitopodium’, for the metapodium/acropodium complex (Figure 3).

Documenting and implementing design patterns

Another challenge in arriving at a cohesive unified ontology was the difference in modeling strategies, design patterns, and terminological conventions across the source ontologies. An example of this was the representation of joints in different ontologies prior to the merge. In TAO, classes representing different skeletal elements had relationships to specific joints; for example, both the ‘quadrate and the ‘anguloarticular’ had overlaps relationships to ‘quadrate-anguloarticular joint’. In Uberon, the dependency is reversed – for any given skeletal element there is no assumption of a relationship to a particular joint class, but joints are assumed to imply the presence of certain elements. For example, in Uberon the ‘quadrate-articular joint’ has two connected_to relationships to the ‘quadrate’ and to the ‘articular’ (anguloarticular). In comparing the two styles, we decided to opt for the latter pattern. This was because although the representation in TAO was valid at the level of teleosts, the broader taxonomic diversity represented in Uberon encompasses variation in the pattern of connectivity of skeletal elements across vertebrates. In Uberon, joints are defined by the skeletal elements they connect to rather than by the overlap of various skeletal elements, and hence the modeling pattern is more stable because individual bones may vary in their connectivity across species.

Another example of a difference in modeling pattern is how the various ontologies represented the connection of teeth to skeletal elements. When the ontologies were initially combined, representation of teeth associated with different bones was variably represented using, for example, connected_to, part_of, overlaps, etc. For example, the relationship of teeth to jaws is represented via an attachment relationship (attaches_to) in Uberon, in comparison to a part_of relationship in TAO. Further, many fishes have teeth in places that would be unusual for a mammal, such as in the pharynx (see Figure 4). We therefore documented a set of core design and modeling patterns [34] that would be applied throughout the ontology, and we proceeded to unify the combined ontology along these lines. As a result, all the teeth in Uberon, whether associated with the mandible and maxilla as in mammals or with the vomer as in fishes, have the same relationship (attaches_to) to supporting structures. This is a major advantage for users wishing to query over the combined ontology and provides a modeling pattern that is more stable across taxa.

A final example that illustrates the need for a common design pattern is the representation of muscles and their attachment. For example, in the course of vertebrate evolution, the jaw closing (jaw adductor) muscles have changed from single, simple muscles to complex, highly differentiated groups of muscles several times. The TAO and the ZFA have a single class, the ‘adductor mandibulae complex’ (TAO:0000311, ZFA:0000311), even though the adductor mandibulae in most teleosts has at least four distinct parts (called A1, A2, A3, Aw). Similarly, in the AAO, there was a single class called ‘jaw muscles’ (AAO:0000247). Amphibians are usually described as having a separate internal and external adductor mandibulae (or levator mandibulae internus and externus, respectively). This broad representation, based on subclass and partonomy alone, was insufficient to describe the complexity. A common design pattern that enabled granular representation and better classification of the individual muscles in the adductor complex that was based additionally on innervation, connectivity, and function was implemented, for example, for the ‘masseter muscle’ (UBERON:0001597):

'capable of part of' some ‘GO:mastication’

'has muscle insertion' some 'mandible'

'has muscle origin' some 'zygomatic arch'

'has muscle origin' some ‘maxilla’

'part of' some ‘cheek’

innervated_by some ‘masseteric nerve' (a branching part of the trigeminal nerve)

As is clear in the aforementioned examples, maintaining a consistent style of modeling across a large ontology is challenging. In our ontological work, we follow the software methodology of documenting common ‘design patterns’ [35], which already has been applied to biological ontologies [36]. For every design pattern a document is created and represented in OWL as part of the ontology itself [34, 37]. In this way, documents are coupled directly to the structures with which they relate, facilitating application. A script is used to generate web pages for every document. The document pertaining to the modeling of joints, for example, is available and includes status, contributors, discussion and a summary of issues as implemented in the ontology [38].

Attribution

Allocation of responsibilities, coordination of editing rights, and attribution were critical parts of the merging process among our different communities. Major contributions consist of additions to Uberon via the tracker, design documents, meetings, and workshops from multiple domain experts. Proposed changes to the core are submitted through the Uberon tracker [41] and vetted by the larger community. Because it is very important to this community and “scientific good practice” to keep track of contributions and to facilitate further discussion regarding design or definitions, a more sophisticated approach to attribution is needed. Contributions to the ontology are marked as metadata in the ontology, at the level of classes, axioms, and the ontology itself, using references to design documents and to individuals using database-cross-reference or ORCID IDs. Design documents also have authors and can be included directly in the ontology as instances (see above) via an import file, as well as online and linked via the Uberon.org website.

Ontology integration challenges

Each of the contributing groups brought a new perspective that enriched the whole. The teleost and amphibian ontologies brought into Uberon a broad set of terms and axioms, many unique to these taxonomic groups, that not only enhanced the anatomical knowledge contained in the whole, but challenged and improved the structure itself. For example, vertebrates have a variety of scale-like elements: fishes have at least three types, amphibians have one, whilst other types of bony deposits (osteoderms) are found throughout other extant and extinct amniotes [43]. These osteoderms are not suspected to be homologous to scales in fishes [44, 45]. Similarly, the vHOG and the VSAO provided higher-level knowledge and grouping terms that supplemented, refined, and revised the classification of scales and other osteoderms within the merged Uberon.

The integration also involved reaching agreement to resolve the alternate perspectives. Consequently, the resulting ontology is more complex than any ontology developed specifically for a single project, such as Phenoscape or Bgee, because it necessarily must accommodate different modeling requirements and therefore a multiplicity of axes of classification. For some users simplified subsets of Uberon might be more useful, for example, for text-miners where automated entity matching on the labels and synonyms would be sufficient. For other purposes, such as gene expression comparisons (Bgee) or identification of candidate genes responsible for major evolutionary change (Phenoscape), full axiomization for a specific anatomical system are useful. We have addressed this by providing many different subsets of Uberon, described on the uberon.org website. In particular, one can utilize taxon specific slims (portions of the ontology) of Uberon as well as a variety of ontology slims that leverage the OBO subset annotation. For instance, the Experimental Factor Ontology (EFO [46]) was recently updated to use a slim of Uberon and these are recorded in Uberon using the EFO_slim annotation property.

Another agreement reached was in the choice of primary labels for classes. We strive to use the most species-neutral terminology possible. As a result of this decision, we renamed existing classes in Uberon such as ‘hand’, ‘foot’, ‘finger’, ‘muscle of hand’, etc. to ‘manus’, ‘pes’, ‘manual digit’, ‘muscle of manus’ etc. We carried this all the way down to named repeating subunits – e.g. ‘distal phalanx of manual digit 1’. The original labels are retained as synonyms. In some cases, the rationale for changing the label was deeper than terminological preference. For example the class UBERON:0006807 was previously labeled ‘lateral epicondyle of humerus’. However, it was pointed out by comparative morphologists on the Phenoscape team that the ‘lateral’ position did not apply to all taxa, due to the changes in posture of limbs that has occurred evolutionarily. We decided to apply a more species-independent terminology and rename this class ‘ectepicondyle of humerus’. The previous label was retained as a synonym, and tagged as the ‘human preferred’ synonym (i.e. the synonym that is preferred by communities looking for a more “Homo sapiens-centric” viewpoint). Each class effectively has a unique primary label, though applications can easily use the different synonym types as they prefer and more can be added as needed.

Ontology integration and alignment is a pervasive challenge for a wide range of use cases in biomedicine. As we have highlighted here, there are numerous communities that have a specific need to record anatomical data. While the anatomy of creatures as diverse as elephants and chickens may not at first glance appear to have a lot in common, any comparative morphologist’s primary objective is to examine those similarities. Anatomy ontologies therefore pose a special problem to ontological alignment - there exist many groups working on similar content without coordination and in different domains. Similar anatomical features end up being represented differently due to different modeling strategies, ontology development principles, and biological perspectives. Ideally, we would have been able to make use of standard tools to reconcile and merge these anatomy ontologies. In practice, we ended up using our own ad-hoc scripting and tooling, which allowed us to perform operations on large batches of the ontology at any one time. Once the initial merge had taken place, curators were able to make edits and corrections at a more fine-grained level using standard editing environments like Protégé. Our automation tools are open and freely available (see Methods), but it is not clear how applicable they would be to other domains, as they were heavily customized towards the specific anatomy ontologies we were working with. We would advise that groups developing anatomy ontologies do so in an integrated modular fashion from the outset, rather than integrating post-hoc, which is always a costly and time-intensive task.

Reconciling common terminological differences

We have encountered many difficulties in the process of defining classes of anatomical entities, often because commonly used categories are “fuzzy.” For example, most of the large muscles of the body are “skeletal” muscles. “Skeletal muscles” are attached to skeletal elements, are “voluntary” muscles innervated by somatic motor neurons, and are derived from somites. In contrast, branchiomeric or pharyngeal arch muscles have long been termed “visceral” muscles, because they are part of the gill arch apparatus, which is derived from cranial mesenchyme (cranial paraxial mesoderm), rather than the somites. The term “visceral” refers to viscera or internal organs. The term “visceral muscle” refers, in most cases, to smooth, involuntary muscles, innervated by autonomic motor neurons. However, branchiomeric muscles are voluntary muscles and they are innervated by large motor neurons that are, like skeletal muscles, considered alpha motor neurons. Furthermore, some voluntary muscles, such as the intrinsic muscles of the tongue, have no attachments to the skeleton and yet are referred to as skeletal muscles. Voluntary muscles are also “striated” muscles, based on their banded appearance under a light microscope. However, cardiac muscle is a form of striated muscle even though it is generally considered an “involuntary” visceral muscle, and is innervated by autonomic motor neurons. To complicate matters further, some invertebrates, such as arthropods, have at least two types of striated muscles. Thus, the traditional terminology presents many complex, contradictory issues. Reconciling complex fuzzy nomenclature is an ongoing and future goal, and we intend to work with the respective communities (the FEED group for this example) to provide competency questions and get feedback on our modeling approaches. In this way, we hope to improve our communication, specificity, and interoperability amongst different content specialists.

Extension beyond vertebrates

Our long term goal is to create a collection of federated anatomical ontologies covering multi-cellular organisms. Many classes in Uberon are applicable outside Vertebrata (e.g., ‘notochord’) and some are widely applicable across metazoans (e.g., ‘muscle tissue’). However, the majority of classes in Uberon are vertebrate specific. Some groups of taxa will be included within Uberon, while others will be developed as separate ontologies that will work together with Uberon as part of a multi-species OWL import chain. In particular, the new poriferan (sponge) ontology (Thacker, this issue), cephalopod ontology [47] and ctenophore ontology are being developed as separate ontologies that reuse a set of CARO and Uberon classes. As the Arthropod Anatomy Ontology (see https://code.google.com/p/arthropod-anatomy-ontology/) is developed further, a number of classes from Uberon (for example, ‘mushroom body’) will be obsoleted and ceded to this ontology. We welcome input from domain experts to extend Uberon or “bud off” new ontologies covering organisms such as cnidarians, annelids, or echinoderms, all of which currently have superficial coverage. While this sounds straightforward, it is actually quite challenging. As we have seen in the development of Uberon, it is often the case that vocabularies are developed for a specific taxon or purpose. Through the Phenotype RCN we are providing training for anatomists and evolutionary biologists to assist in the task of cataloguing the “parts of life”.

In summary, formal integration of anatomical systems from major vertebrate groups into Uberon has expanded and enabled our efforts to investigate disease [48–51] and evolution [14, 18] using comparative anatomical approaches.

Ontology merge

Prior to the merge, we had four source ontologies (VSAO, AAO, TAO, vHOG) and a target ontology (Uberon). In earlier work, we had enhanced the set of equivalence axioms connecting Uberon classes with source classes. The goal then was to eliminate redundant classes (i.e. equivalent classes) by merging source classes into the Uberon ontology, such that the resulting ontology was cohesive, did not lose information, and that duplication was minimized. As a first step, we enhanced the set of equivalence axioms connecting target classes with source classes. These were a mixture of direct equivalence axioms (e.g. Uberon:femur equivalentTo VSAO:femur) and taxonomic equivalence axioms (e.g. Uberon:femur and part_of some Amphibian EquivalentTo AAO:femur). The overall approach is shown in Figure 5 and is detailed here.Having the equivalence axioms in place, we then divided all source classes into those for which an equivalence axiom (either direct or taxonomic equivalence) existed in Uberon, and those source classes without equivalent classes in Uberon. For example, the amphibian and teleost classes for ‘quadrate bone’ (TAO:0000621 and AAO:0000525) were in the former group because both had equivalence axioms linking them to UBERON:0006597 (See Figure 6).

For those classes for which equivalence axioms existed, we verified that all axioms were also present in the Uberon class; if they were not isomorphic, the Uberon class was annotated with a structured comment. For example, the TAO had an axiom stating that the ‘quadrate’ overlaps the ‘quadrate-anguloarticular joint’. The equivalent relationship was not added to the Uberon class; instead, a comment was added stating that the TAO contained this more specific relationship. The original textual definition from TAO was transferred across as an “external definition” annotation. This allowed us to conservatively automate transferal of information from the sources without making logical statements that did not hold in the wider taxonomic context. Conversely, if there were no additional axiomatic constraints in the source class, the source class was made obsolete and directly replaced by its equivalent Uberon class. Synonyms were added to these Uberon classes as needed to retain information coming from the source class.

For every source class that did not have an equivalence axiom in Uberon, we created an obsolescence record, created a new Uberon class, and transferred the axioms across from the source class whilst mapping all entities using the equivalence axioms. For example, the TAO class ‘quadrate posterodorsolateral process’ (TAO:0001611) was not previously represented in Uberon, so this generated a new class (UBERON:2001611). The same label, definition, and other descriptive information were retained. If the source class also contained a part_of relationship (e.g. the TAO class for ‘quadrate posterodorsolateral process’ part_of TAO class ‘quadrate’ (TAO:0000621)), then the relation was rewritten to use the equivalent Uberon class (in this example, UBERON:0006597). All new classes created in this way are considered part of Uberon, but for administrative reasons were placed in a separate ontology file, ‘phenoscape-ext.owl’ or “Ext” for short. This allowed ontology editors to refine the results of the translation process and add new classes if they were not present in the core. The separation of axioms into a core ontology file and the extension Ext file allows a division of labor and editorial rights. All original Uberon terms are included in the core file, that is those with equivalences to the original source classes. Terms unique to the other ontologies were placed in the Ext file.

Merging of TAO

The TAO contained 3048 classes, with a number of TAO classes that had been obsoleted and replaced by VSAO classes. Prior to the merge, Uberon had equivalence axioms to 970 TAO classes, with the remaining 2078 classes in TAO representing largely teleost-specific structures (TAO version from 9 August 2012). The latter classes were obsoleted and replaced by 2078 equivalent new classes in the Uberon Ext ontology (these have IDs of the form UBERON:2nnnnnn, the final six digits were preserved for convenience), and corresponding axioms preserved or rewritten (see above). Additional manual reconciliation took place after this automated process. Because TAO was originally created by cloning ZFA, the ontology contained a large number of classes (in particular cell types and brain regions) that were not truly generalized beyond Danio rerio; we obsoleted these as they added little value beyond what was already present in the species-specific Zebrafish anatomy ontology. 1630 classes were retained in Uberon Ext from TAO.

Merging of AAO

The process for AAO followed that of TAO. The AAO contained 1595 classes (version from 13 April 2012), with 789 having taxonomic equivalents in Uberon. The remaining classes were obsoleted with equivalent classes added to Ext (in the ID range UBERON:3nnnnnn). Due to problems arising from the automated process by which much of the AAO was created, emphasis was put on manually enhancing these classes. Many classes had ambiguous names such as ‘processus dorsalis’ (AAO: 0000655) and lacked definitions. Working with experts in this field we generated less ambiguous names and definitions.

Merging of VSAO

VSAO contained 201 classes. The editors of Uberon and VSAO coordinated during the development of the VSAO, and therefore much of the VSAO was already present and well aligned with Uberon at the outset of the merge process. There were 54 classes not represented (version from 16 July 2012) and these became classes in Ext in the UBERON:40nnnnn ID space. VSAO definitions and axioms were retained because they had been expert-vetted and carefully curated [16].

Merging of vHOG

Because both Uberon and vHOG had their origins in aligning existing species-specific anatomy ontologies, the majority of classes in vHOG were already represented in Uberon. The manually verified groupings in vHOG were transferred into ‘homology notes’ annotations within the Uberon ontology, and were used to refine and verify taxon equivalence axioms.

Editing and releases

The Ext component is edited using Protégé 4. The core Uberon ontology is imported, but not directly edited by Ext editors (Protégé visually distinguishes between imported classes and declared classes, providing editors with the appropriate visual cues). Maintenance and editing of the core ontology file requires specialized tooling to keep the ontology in synchrony with ssAOs, GO, and other ontologies. However, the Ext ontology contains subclasses of core classes and can be maintained by curators using Protégé without this specialized tooling. The distinction between core and Ext files is an editorial artifact and disappears in the released file. We used Protégé 4.2, plus the OBO-Edit graph component (Dietze, in prep) from Obo-Edit [52], plus the depictions plugin (see below) and additional plugins that support an OBO style of development. Additional file 1: Table S1 holds a list of object properties and example uses. The release files associated with the ontology merge described in this manuscript are available at: http://purl.obolibrary.org/obo/uberon/releases/2013-02-21/

Multiple forms of Uberon are described on Uberon.org, such as:

http://purl.obolibrary.org/obo/uberon/ext.owl as well as taxon-specific files and simple pre-reasoned files with only three primary relations (subclass, part_of, and develops_from).

All tools used to generate and maintain the ontology are open and freely available and documented here:

https://github.com/obophenotype/uberon/wiki/Editor-tools-guide

Centralized reasoning using Jenkins

As described [11], Uberon includes a large number of axioms to ensure both internal consistency and consistency with other ontologies. We require that the ontology always satisfy these conditions. In addition, we require checks that the ontology did not violate certain syntactic conditions (e.g. every class should have exactly one label and no more than one text definition). Some of these checks are difficult to run locally on the ontology developer’s machine, so we made use of the OBO library Jenkins Continuous Integration server [53]. Every commit performed by an editor triggers a validation and build pipeline, which also generated an obo-format file which is used in Phenoscape character annotation in Phenex and entity matching through CharaParser [54].

Consistency checks with ssAOs

We have implemented a process by which the Xrefs and equivalence axioms are checked between the single-species anatomy ontologies and the ontologies now integrated into Uberon. For example, if a zebrafish class has an Xref to a TAO class that has not been obsoleted and replaced by an Uberon class (using the replaced_by annotation property), these will be unsatisfiable and will be fixed in the source zebrafish file.

Integration of depictions

This update also includes images depicting instances of ontology classes in a variety of species. The initial set of depictions were generated by querying DBpedia for images using the existing Uberon to Wikipedia mappings (see original Uberon paper [11] for details). Due to the use of DBpedia/Wikipedia, this initial set is mostly human specific, but we have been manually augmenting this using the image depiction plugin [55].