Accurate representation of phenotypes using ontologies is important in biology and biomedicine. This paper describes the OWL translation of our methodology for representing phenotypes using ontologies in the OBO Foundry.
In simple terms, a phenotype is a collection of characteristics that arise through the expression of the genes of an organism, in an environment. Some examples of phenotypes are the red eyes of a typical fruitfly, Drosophila melanogaster; the length of a body part such as the tail of a mouse; high blood pressure in the arteries of a human, sensitivity to a chemical or to light; reduced mass of an organ or part of an organism such as bone.
In biology, we are often interested in how variations in genotype and environment can lead to variations in phenotype. There is a wealth of research to draw from, but much of the knowledge is currently encoded in natural language and free text in the literature and biological databases[?] - if this knowledge can be captured using ontologies and ontological formalisms we increase the range of computational methods that can be used to analyse this data[?].
This paper describes the OWL translation of a formalism for representing phenotypes making use of the ontologies that comprise Open Bio-Ontologies Foundry (http://www.obofoundry.org). Many of these ontologies such as the Gene Ontology[?], the Foundational Model of Anatomy[?] and the Cell Ontology[?] represent canonical biology, which is to say the biology of “normal”, “typical” or healthy organisms. We use these ontologies as building blocks that can be combined with an ontology of qualities to construct descriptions of phenotypes, many of which deviate from canonical biology. This methodology can be used in a variety of different kinds of investigations.
The reader of this paper should be aware that the end-user of the software systems and databases built around the formalism presented do not work directly at level of OWL constructs, or at their OBO-Formalism equivalents. They typically interact via intermediate representations and user interfaces such as Phenote1 - more details can be found on the phenotype ontology website2. We present the OWL version of the formalism in order to give it a grounding in logic, and as a means of leveraging OWL-based tools and reasoning engines, although space dictates we do not discuss these applications. 1 http://www.phenote.org 2 http://www.phenotypeontology.org 1.1
We use the Manchester OWL syntax [?] for writing our OWL representations3. Most of the OWL descriptions in this paper can also be translated to OBOFormat equivalents; we do not include these here for brevity, details can be found at http://tinyurl.com/326feg and the OboInOwl page4.
Our examples are all drawn from OBO Foundry Ontologies[?] for the most part. OBO ontologies uniquely identify classes via a numeric identifier, but in the examples given we replace these with class labels in order to enhance readability. We do not always specify the ontology used in the examples given, the full list can be found on the OBO Foundry website5. 2
The biological and biomedical literature contains descriptions of phenotypes such as the following: 1. red eye (a typical or wild-type specimen of the fruitfly Drosophila melanogaster has red eyes. 2. curvature of a wing 3. high blood pressure 4. high glucose concentration in the blood 5. short tail 6. invaginated cell membrane
How should we go about representing phenotypes such as these in OWL? We will first review two existing W3C recommendations (instance and class partitions), and then describe our approach. 2.1
The W3C Working Group note Representing Specified Values in OWL: value
partitions and value sets[?] provides guidelines for two different ways of representing
what it calls specified values: (
The first method involves subdividing a quality such as curvature into owl Individuals, such as flat and curved, represented in Manchester Syntax as:
3 http://www.co-ode.org/resources/reference/manchester syntax/ 4 http://tinyurl.com/34ddl2 5 http://obofoundry.org
We may also want to use differentFrom statements to state that flat and curved are distinct individuals:
We can relate individuals to color hues using an OWL functional property6: So an instance of a curved fruitfly wing7 could be represented as:
or the logically equivalent:
The W3C document notes some problems with this approach, such as the inability to further partition the values. For example, we could not create fiat partitions dividing curved into highly curved, mildly curved, etc8. This is a major disadvantage for representing phenotypes, as multiple levels of partitioning may be required to accurately represent the biology. On top of the listed problems, we would also add that this solution is ontologically unsatisfying, as we hold that these “values” correspond to types (also known as universals) and are instantiated multiple times in nature, and are best represented as owl Classes, not owl Individuals. 2.2
The second method described in the W3C technical note involves partitioning qualities into owl Classes. We can write the same example in Manchester Syntax as follows:
This states that the class CurvatureValue is equivalent to the union of the 2 classes Flat and Curved - this also states the list is closed - there are no other classes that are also CurvatureValues. We could, however, create subclasses of Curved.
In addition we can specify disjointness conditions: 6 for simplicity, we omit treatment of 3D objects that may have surfaces of different
curvature, and assume any entity has a single curvature 7 Represented in the Fly Anatomy Ontology http://obo.sourceforge.net/cgi-bin/detail.cgi?fly anatomy 8 in this example it may be possible to combine value partitions but this is inherently
problematic
The functional property linking curvature individuals to individuals with curvature is the same as for the enumeration pattern:
Manchester Syntax even has a macro feature for conveniently representing value partitions:
This combines the unionOf axioms with the disjointWith axioms into a handy syntactic idiom.
So an instance of a curved fly wing could be represented as:
It is important to note that an important consequence of this pattern is that the existence of each individual of type CurvedWing also entails the existence of a distinct quality individual of type Curved. This is in contrast to the enumeration pattern, in which there is a single quality individual curved, which is “shared” by all curved things.
This gives the class partition pattern various advantages over the enumeration of individuals pattern. It allows the quality classes to be further partitioned into more refined classes. We also hold that this pattern is ontologically preferable, and that distinct quality instances, such as my hair color or the shape of a particular organ, exist in reality. 2.3
Our approach, which has been dubbed the “EQ” model, is highly similar to the class partition pattern described above, but differs in a few key respects. Our treatment is based on a previously published formal analysis of qualities[?], and uses an ontology of qualities called PATO[?][?].
In our formalisation, a phenotype is one or more qualities which are borne 9 by bearer entities in some organism . The relation between a quality and its bearer entity is one of inherence. An example of an instance of a phenotype is the particular instance of high curvature inhering in a particular instance of wing in a particular fruitfly. 9 the phenotype of an organism is influenced by the genetics of that organism, mediated by the environment; description of these crucial factors are omitted for reasons of space
Note that sometimes qualities are referred to as properties, both colloquially and in the philosophical literature, but we avoid this usage, due to the contradictory use of the term property within the OWL community to denote what we would call a relation. Here we also distinguish between instances, the particular instances in reality, and the individuals that represent them.
All qualities are linked to their bearers by means of the inherence relation, which can be represented in OWL as:
Note that this approach has no need of relations such as hasColor (although these could optionally be added as sub-relations of inheresIn). Rather than creating distinct class partitions for hasColor, hasShape and so on, we have a single subclass hierarchy, with all quality classes ultimately inheriting from the root Quality class.
We can describe the fly wild-type phenotype high curvature wings using the class expression: A-1 Curved that inheresIn some Wing
We can also describe a curved wing: A-2 Wing that hasQuality some Curved
These two expressions are intuitively similar, but crucially distinct, in that the first represents a quality, and the second represents the three dimensional object that is the bearer of the quality. The existence of an instance of one entails the existence of some instance of the other at any given time. Thus either class is eligible as the description of the phenotype. We choose A-1 as the “normal-form” for phenotypes.
Our approach is most consistent with the second approach in the W3C note above, as we use owl Classes to denote our Quality universals. However, we differ from this approach in that we have no specific need for multiple relations such as hasColor, hasSize, hasShape and so on. Our ontology of qualities is a single hierarchy, with all classes a subclass of the root class Quality.
To support the description of phenotypes, we have developed an ontology of qualities called PATO10. The ontology is constructed as far as possible according to OBO Foundry11 principles, with each class in the ontology having a single is a (subclass) parent. The ontology is maintained in OBO format and is available from OBO, but is also available in OWL12.
The ontology is divided along granular levels [?] [?]. On the one hand we have physical qualities, such as mass, velocity and color; next we have cellular qualities such as ploidy (which inheres in a cell or cell nucleus by virtue of the number of chromosomes it has) and cellular potential (capability of differentiating into a range of other cell types; finally we have organismal qualities. Note that of course cells and organisms have physical qualities such as mass, by virtue of the physical qualities of their subparts. Physics is sufficient to give an account of 10 This originally stood for Phenotype and Trait Ontology. The full name is somewhat misleading as the classes in the ontology represent qualities, rather than full-blown phenotypes 11 http://www.obofoundry.org 12 http://purl.org/obo/owl/PATO physical qualities - however, other qualities are emergent and a description must be given at a higher level of granularity.
Although the ontology is developed with the intent of use for phenotypes in biology and biomedicine, it is our hope that the physical quality sub-hierarchy of PATO will prove useful in other scientific realms. 3.3
We follow [?] in distinguishing between monadic and relational qualities. The former are qualities that inhere in and depend on a single entity. The latter depend on some additional entity or entity type. For example, shape is a monadic quality because shapes exist in totality in a single shape-bearing. A quality such as sensitivity, by contrast, has a single bearer but is always with respect to some other entity or kind of entity (such as in, for example, sensitivity to sunlight).
Neuhaus et al[?] represent relational qualities in a first-order logic framework using an inherence relation with a higher arity - the additional entity universals are referenced as additional arguments.
As all relations in OWL are binary, we must introduce an additional relation to represent relational qualities - we call this the “towards” relation.
Using this relation we can represent the phenotype skin photosensitivity to UV light using the class expression: A-1 Sensitive that towards some UltravioletLight and inheresIn some Skin
Note this is ontologically unsatisfactory, as any instance of photosensitivity is with respect to the universal UltravioletLight, rather than any one particular instance of a light wave or particle bundle. However, for the purposes of DL reasoning the representation is practical, and a reasoner such as Pellet will correctly reason that A-1 is subsumed by an expression such as A-2: A-2 Sensitive that towards some Radiation and inheresIn some Organ
We could instead use the UltravioletLight class as an individual, as in A-3: A-3 Sensitive that towards value UltravioletLight and inheresIn some Skin
However, this has the undesirable consequence of taking us to OWL Full, and will result in omissions in the reasoning process. 3.4
Descriptions of phenotypes may be temporally qualified. The qualities in question may only inhere in the bearer entity over some interval - for example, organismal development involves changes in morphology of parts of organisms. Sometimes the quality may only be observed over a certain interval - for example, a particular assay of high blood pressure.
This raises interesting representation issues - all relations in OWL are binary, and there is no simple way to add a temporal index to binary instance-level relations, such as the relation between a quality and its bearer. There are a number of representation patterns for getting around this limitation [?] [?], such as reifying relations, but unfortunately they are all unsastifactory from an ontological and practical point of view.
Consider an portion of tissue, such as the optic placode13, that changes its shape from flat to curved or invaginated during a specific stage of development. How do we represent this using binary relations?
We are unsatisfied with all the solutions afforded us within the constraints of binary relations - Our current compromise approach would be to represent this kind of anatomical entity as follows: A-4 OpticPlacode that hasQuality some (Flat that during some Stage11) and hasQuality some (Invaginated that during some Stage12) (note that here we have an object-centered representation rather than qualitycentered, using the hasQuality relation, the inverse of inheresIn)
This approach is unsatisfactory because it involves stating artefactual classes such as “flat during stage 11”. This approach also causes problems when reasoning, as illustrated below:
Assume also that Invaginated and Flat are disjoint (a reasonable assumption, given some fiat[?] quality boundary), and subclasses of Shape
Now we add the restriction that all organism parts have exactly one shape, as follows14:
However, if flat and invaginated are disjoint classes, and we declare that each organism part can only assume a single shape, this will result A-4 being equivalent to the empty set, an undesirable result.
An alternate representational paradigm involves representing “time-slices” of three dimensional objects (such as an organism part) at different times. These slices can then be related via binary relations to qualities that obtain over intervals. Each slice can then be linked back to the representation of the three dimensional object. This approach is sometimes called the perdurantist or 4dimensionalist approach, and the time-slices are sometimes called histories or snapshots. 13 the part of an organism that will develop into the eye 14 the intent is to specify a Qualified Cardinality Constraint, as is allowed in OWL-1.1, but the authors are not sure if this is the correct for in the Manchester syntax
For example, we can represent an optic placode whose shape changes between flat and invaginated during a stage of development using the following class expression: A-5 OpticPlacode that hasTimeSlice some (Slice that hasQuality some Flat and during Stage11) and hasTimeSlice some (Slice that hasQuality some Invaginated and during Stage12)
However, this solution is unsatisfactory from both an ontological and a practical perspective. On the one hand we must introduce additional artefacts, that have no place in upper ontologies such as BFO[?] and DOLCE. Our commitment to OBO Foundry principles leads us to reject the existence of unusual entities such as time-slices. In addition, the time slice paradigm must be anticipated in all ontologies used, such that for example the domain and range of relations can be 3D objects or time-slices of 3D objects.
At best, the above construction is unwieldy and will require specific tool support if these constructs are to be generally useful.
Another approach is to reify the relation in question, in this case inheresIn , modeling it with an OWL class rather than an OWL property. We omit a discussion of this approach for reasons of space, but we find this solution even more unsatisfactory.
An extension to OWL allowing would simplify this problem immensely. For example, if relations could take a third temporal argument as illustrated in the following expression: A-6 Organism that (hasQuality some Flat at Stage11) and (hasQuality some Invaginated at Stage12)
The proposed AT operator could be used to index any relation at the class or instance level in OWL. We recognise that temporal extensions to DL expressivity would introduce many challenging problems, but we believe that this issue must be addressed sooner rather than later in order to accurately represent biology.
One possibility is for 3-ary temporal expression such as the above to be “compiled down” to standard binary relation oriented description logics bu automatically introducing artefactual entities such as time-slices “behind the scenes”. 4
Representation of and reasoning over phenotypes is of huge importance to the life sciences and biomedicine. Our methodology for describing phenotypes through the composition of quality classes with classes of quality-bearers via the inherence relation works well for a variety of phenotypes, and promotes the reuse of ontologies in a modular way. The ontology of qualities, PATO, can be used across a variety of organisms for phenotypes across multiple levels of granularity.
We find the ability to compose classes using description logic type expressions useful, and is a representational technique the general life sciences and biomedical informatics and database communities would find useful in general. In particular, the elegance of Manchester OWL Syntax renders complex ontological representations accessible to ordinary mortals15. This is a significant result for taking OWL out of the ivory tower into the real world. To illustrate this, here is a class expression involving multiple levels of nesting utilising classes from multiple ontologies to represent the phenotype high permeability of mitochondrial cristae in the axons of pyramidal cells in the CA1 hippocampal field: A-1 HighPermeability that inheresIn some (MitochondrialCristae that partOf some (Axon that partOf some PyramidalCell that partOf some CA1Field))
The ability to attach these class expressions to named phenotypes, such as those that may be found in the Mammalian Phenotype Ontology[?] and the Plant Trait Ontology[?] will be important for comparing data across organisms, and we have commenced an effort to create these expressions16.
We have found that limiting expressions to binary relations has negative implications for ontologically sound and intuitive representations of biology, particularly with respect to time.
For lack of space we did not address various other important topics such as the representation of experimental assays and quantitative measurements. We also avoided discussion of representation on the complex relationship between genotypes, environments and phenotypes. However, we note that the methodology employed in this paper for representing phenotypes is also applicable to representing environment types. 5
The ideas in this paper arose through discussions with many individuals, including Fabian Neuhaus, Barry Smith, Alan Rector, Nigam Shah, William Bug, Mark Gibson, John Day-Richter, Michael Ashburner and the members of the FlyBase consortium, Monte Westerfield and the members of ZFIN Consortium, Pankaj Jaiswal and the participants on the obo-phenotype mailing list.