More recently, ontologies have been shown to have significant benefits for the analysis of data resulting from high-throughput technologies [8] and for automated reasoning applications [9-12], which has led to organized attempts to improve the structure and formal rigor of ontologies in ways that will better support computational analysis and reasoning [13].

To help meet the data annotation needs of model organism researchers, Bard and colleagues developed an ontology of cell types [14] that has been widely used for the annotation of data in genome and other biological databases, including biorepository data. To enhance the CL's utility for computational analyses, we developed a systematic approach for the ontological representation of cells as described below, adhering to the principles outlined in [13].

The Cell Ontology (CL) currently contains representations of some 863 cell types, covering cell types from the major model organisms, including prokaryotic, fungal, animal, and plant cell types. This broad scope is accomplished primarily through the use of terms and definitions that can be applied to cell types from multiple types of organisms.

Cell types in CL are classified on the basis of a plurality of structural, histological, functional, and lineage properties. While some terms are provided with natural language definitions, most of the information in CL is conveyed through hierarchical classifications of the cell types along the following multiple axes of classification:

   cell by organism

   cell by histology

   cell by function

   cell by lineage

   cell by nuclear number

   cell by ploidy.

Each cell type within the CL is related to at least one supertype via the is_a (subtype) relation, but most CL cell types have multiple supertypes, and are thus classified on multiple axes, a phenonmenon referred to as 'is_a overloading' [15]. CL is currently undergoing review, addressing specifically the problems that arise from the use of multiple modes of classification.

An example of the CL's use of multiple hierarchies is classification of Langerhans cell within the cell by organism and cell by nuclear number hierarchies as a subtype of animal cell and single nucleate cell, respectively (Figure ​(Figure1).1). Langerhans cell is further classified within the functional hierarchy (Figure ​(Figure1)1) as a subtype of:

   transporting cell

   professional antigen presenting cell

   circulating cell

   motile cell

   defensive cell

   phagocyte (sensu Vertebrata).

Following a path along both is_a and develops_from relations – the latter is the relation "used to code developmental lineage relationships" – reveals that Langerhans cells are of mesodermal and hematopoietic lineage (Figure ​(Figure1)1) [14]. is_a and develops_from are the only relations used in the current version of the CL.

To enhance the CL's utility for computational analyses, we developed a systematic approach for the ontological representation of cells that:

i) separates classification via the is_a relation from the assertion of structural, functional, and lineage properties by using formally defined, property-specific relations, such as has_function

ii) systematically includes both species-neutral and species-specific representations of cell types

iii) defines cell types on the basis of specific combinations of surface proteins used for identification of the cells via flow cytometry.

We have applied our method to develop an extension of the CL, DC-CL, for dendritic cells (DC), focusing on the DC types observed in mice and humans. The approach we propose increases the amount and accuracy of information contained in DC-CL, enhances its support for cross-species data integration, and optimizes it for the analysis of flow cytometry data.

The CL currently contains representations of six DC types:

   dendritic cell

   plasmacytoid dendritic cell

   myeloid dendritic cell

   immature myeloid dendritic cell

   mature myeloid dendritic cell

   Langerhans cell

which are defined on the basis of structural, functional, and lineage characteristics. The definitions are formulated in a highly general way to ensure broad applicability, but unfortunately at the cost of a precise specification of cell types. For example, plasmacytoid dendritic cell is defined as:

A dendritic cell type of distinct morphology, localization, and surface marker expression from other dendritic cell types and associated with early stage immune responses, particularly the release of physiologically abundant amounts of type 1 interferon in response to infection.

In the immunological literature and research community, the term 'dendritic cell' does not refer to a single, clearly distinguishable cell type; rather it refers to cells from a variety of sub-populations that have different morphologies, are distributed across different microenvironments within the body, express different microbial receptors and surface molecules, and different cytokines [16-18]. Cells in the various subpopulations are referred to using a common term because they are optimized to play a particular role in an immune response, that of priming an immune response by stimulating naïve T cells in the T cell zones of secondary lymphoid tissue to proliferate and execute their effector activities, but the cells in the various sub-populations are equipped to detect different pathogens and modulate distinct classes of immune responses. The structural, functional, and lineage similarities and differences between these sub-populations are not yet well understood, however, and their study is challenging because they are sensitive to changes in the cells' microenvironment. The subpopulations are typically characterized using a combination of variables, including flow cytometry or immunohistochemistry markers, function, and anatomical location, but the issue of how to define distinct DC subtypes is still an area of active debate within the immunology community.

i) structural components, such as granules in the cytoplasm or molecules on the cell surface,

ii) functions cells of the given type perform,

iii) dispositions, such as the tendency to participate in certain types of processes,

iv) anatomical locations in which the cells are found, and

v) lineage relationships.

Among these properties, we emphasize structural components, as these are most often used to identify cell types for study and because their use for classification facilitates maintenance of a single-hierarchy classification of DC types (Figure ​(Figure2).2). More precisely, we classify DCs by surface protein expression. For other types of cells, other structural components may be used, as for example in the classification of eosinophils, basophils, and neutrophils, which can be classified on the basis of types of cytoplasmic granules.

We next identified a set of ontology development principles designed to maximize both an ontology's utility for computational analysis and reasoning and its interoperability with existing resources. The basis of this approach is use of relations from the OBO Foundry Relation Ontology (RO) [19] to link terms both within DC-CL and also from DC-CL to other OBO Foundry ontologies (Figure ​(Figure3),3), as described in [13,20]. Relations in the RO are formally defined to support automated reasoning [19]. In addition, in constructing DC-CL we employed accepted principles of ontology development outlined by the OBO Foundry [13], including the use of genus-differentia definitions [21].

We use RO relations wherever possible, and define new relations where needed, following the approach used in RO. Relations in RO are defined in accordance with the distinction between types and instances [19], corresponding to the standard Description Logic distinction between A-boxes and T-boxes used within the OWL/Semantic Web community [22]. Types are general; they are the kinds of things that exist and are documented in scientific textbooks (e.g. human, red blood cell); instances are particular; they are the specific examples upon which experiments are performed (e.g. J. Craig Venter, the many red blood cells in my body). The relations between types represented in an ontology are defined in terms of relations between the corresponding instances of those types of the sort that can be observed in experiments. Thus, a type-level relation R will be defined in terms of the corresponding instance-level relation R as follows [where italics indicates type-level relations, bold indicates instance-level relations, type-level variables are denoted by upper case (X, Y, Z, ...), and instance-level variables are denoted by lower case italics (x, y, z, ...)]:

X stands_in_R_to Y = _def for every instance x of type X, there exists at least one instance y of type Y, such that x stands_in_R_to y.

We use R(X, Y) and XRY to abbreviate: all Xs stand in R to some Y. Thus, has_part(human, brain) abbreviates: every instance of human has some instance-level part which is an instance of brain.

Defining type-level relations in terms of all instances of X ensures that assertions of relations between types hold universally, i.e. XRY will hold only if all instances of X stand in R to some instance of Y. This universality in turn ensures the possibility of transitive reasoning by ensuring that relations transitive at the instance level are transitive at the type level. Thus, if R is transitive and R is defined in terms of R and all instances of X, then, if XRY holds at time t and YRZ holds at t, XRZ also holds at t.

We have developed DC-CL using a systematic approach for the ontological representation of cells that:

i) separates classification via the is_a relation from the assertion of structural, functional, and lineage properties by using formally defined, property-specific relations, such as has_function

ii) systematically includes both species-neutral and species-specific types

iii) defines cell types on the basis of specific combinations of surface proteins used for identification of the cells via flow cytometry.

The use of property-specific relations, such as has_function, to incorporate structural, functional, and lineage properties has many benefits. First, this approach eliminates many of the errors that frequently result from multiple uses of the is_a relation [34-36] in what has been called 'is_a overloading' [15]. Second, the is_a relation can only be used between entities of the same ontological category (higher level types, such as those found in the Basic Formal Ontology described below), while specific relations can be used to relate cells to entities in other categories, such as functions (has_function), molecules (has_part), and processes (participates_in), that are represented within their own ontologies. DC-CL is formally connected to the hierarchical structure and relations of these ontologies, as well as the data annotated in their terms, thereby providing significant additional information and opportunities for data integration. The use of property-specific relations also allows us, without sacrificing expressive power, to maintain a policy of single inheritance (each representational unit in the ontology has maximally one single asserted is_a parent), which brings benefits such as clearer statement of definitions, easier and more reliable curation, ability to use more powerful reasoning tools, and the ability to have a unique measure of distance between any two terms on the same branch of an ontology. Finally, the use of property-specific relations enhances ontologies for computational analyses because each relation can be defined with its own inference properties.

The inclusion of species-specific cell types allows for the more specific annotation of data and for the incorporation within DC-CL of species-specific properties, many of which have important functional consequences. For example, the plasmacytoid DCs observed in humans (CD11c^-) express Toll-like receptors (TLR) TLR7 and TLR9, while the plasmacytoid DCs observed in mice (CD11c^low) express all mouse TLRs except for TLR3 and TLR4, with consequent differences in the types of pathogens human and mouse plasmacytoid DCs can detect [37]. We avoid use of species of origin as a basis of defining types, however, and only define types based on the presence or absence of specific surface proteins. Thus, the plasmacytoid DCs observed in humans are instances of the type CD11c^- plasmacytoid dendritic cell, while the plasmacytoid DCs observed in mice are instances of the type CD11c^low plasmacytoid dendritic cell, where the two types are defined by the patterns of surface protein expression given in the above definitions. Plasmacytoid DCs observed in a third species to have either pattern of surface protein expression would be instances of the corresponding type. In addition, we only include assertions about the cell types that hold across all species in which the type is observed. In this way, the inclusion of species-specific DC types in DC-CL facilitates understanding of the similarities and differences between mouse and human immunology and improved capacity for generating hypotheses about the human immune response from the interpretation of the results of mouse experiments. In this way DC-CL also fosters the advance of translational medicine.

To define cell types on the basis of species of origin, or to include assertions that hold for the type in one species but not another, we recommend the creation of species-specific extensions rather than the inclusion of such types in CL or DC-CL. This approach allows for the representation of detailed, species-specific information without using multiple modes of classification (structure and species of origin) or including conflicting assertions in the core ontology. The approach of more specific extensions of a core template ontology has been used successfully in the creation of species-specific anatomy ontologies as extensions of the Common Anatomy Reference Ontology (CARO) [38] and in the creation of ontologies of specific infectious diseases as extensions of the core Infectious Disease Ontology (IDO) [39].

The use of specific combinations of surface proteins to define DC subtypes has advantages both for the creation of DC-CL and for its application to the analysis of cellular data. A primary means by which experimentalists distinguish cell types is by distinguishing patterns of protein expression using flow cytometry. Defining DC subtypes in terms of flow markers allows easy incorporation into DC-CL of new discoveries about DCs deriving from experiments involving flow cytometry to isolate or analyze cell populations. Similarly, defining DC subtypes in terms of flow markers optimizes DC-CL for the annotation, analysis, and integration of flow cytometry data and of data deriving from experiments in which fluorescence-activated cell sorting is used as a source of cells. Just as the Gene Ontology has been shown to offer significant benefits for the computational analysis of high-throughput data in the study of gene expression using hybridization microarrays [8], we anticipate similar benefits from the use of an ontology of cell types to support analysis of high-throughput, multidimensional flow data.

The relations has_high_amount and has_low_amount defined in terms of the geometric mean are used in the definition of cell types and are not meant to replace more complicated statistical methods for the analysis of flow cytometry data, such as is described in [40], or other cellular data. Such statistical methods can be applied to the analysis of individual flow data sets, while ontology definitions need to hold universally, across different experimental designs, protocols, and equipment and across differences in the resulting distributions of fluorescence intensities for reference cells. Indeed, the ontology definitions should hold across different assays for surface protein expression, and should not be tied directly to flow cytometry. We have therefore taken a relatively simple approach to the formulation of cell definitions that hold universally and that are supported by our current understanding of DC biology. It is our hope, however, that our work, taken together with [40], will encourage the use of more objective criteria in the analysis of flow cytometry data and in the description and analysis of cell types in general.

The classification of DCs is still an area of active research, thus DC-CL will continue to undergo revisions to keep current with new research results and new technologies for the characterization of cell types [28]. Because ontologies are based on an open world assumption, in contrast to relational databases, they are easily extended to include new subcategories. In addition, the formulation of DC-CL definitions as logically conjoined statements of the from XRY makes it easy to add or remove surface proteins from the definition of any cell type and to use a reasoner to assess consequences of the revision on the DC-CL hierarchy. Thus, newly discovered surface markers can be easily incorporated into the ontology. Furthermore, the system we have outlined is readily applicable to subcellular localizations other than the cell surface and to other cellular components such as mRNA molecules or cytoplasmic granules. In addition to defining more localization-specific relations like has_plasma_membrane_part, the general RO relations has_part and lacks can be used. For all of these relations, too, cellular components other than proteins and protein complexes can be used as arguments. Morphological characteristics such as size and shape can also be used to define cell types using the has_quality relation to link to the relevant qualities in PATO, the ontology of phenotypic qualities [41]. In this way, the DC-CL framework lends itself quite readily to the incorporation of new information as knowledge of DC biology increases.

We have built our representations of cell types in DC-CL by relating terms in the domains covered by the OBO Foundry ontologies using relations from the Foundry's relation ontology (RO) and creating new relations as needed. The OBO Foundry [6] was created in 2006 by a group of developers of OBO ontologies on the basis of an evolving set of principles designed to foster the pursuit of best practice in ontology development [13]. Its ontologies are designed to represent in an interoperable fashion the biomedical reality from which data are sampled. Their development within the framework of a common top-level ontology (Basic Formal Ontology, BFO, [42]) and consistent employment of a common set of relations allows Foundry ontologies to be used together as modules of a larger system.

There are many benefits to building DC-CL from OBO Foundry ontologies. In addition to the formalism underlying Foundry ontologies ensuring their support for sophisticated computation both within and between ontologies, building from Foundry ontologies means extensive use of existing ontology resources, both eliminating redundant effort and providing a significant head-start to ontology development. By building on OBO Foundry ontologies, DC-CL is automatically interoperable with other ontologies that also build from Foundry ontologies and with the large information resources, such as UniProt, that use Foundry ontologies for their annotations, representing a wide base of existing annotations. Finally, as OBO Foundry ontologies, and in particular GO, are widely used, use of Foundry ontologies in constructing DC-CL improves the chances that DC-CL will be accepted by the biological ontology and database communities.