Simplified ontologies allowing comparison of developmental mammalian gene expression

The Developmental eVOC ontologies presented are simplified orthogonal ontologies describing the temporal and spatial distribution of developmental human and mouse anatomy.


Background Ontologies and gene expression
Biological investigation into mammalian biology employs standardized methods of data annotation by consortia such as MGED (Microarray Gene Expression Data Society) and CGAP (Cancer Genome Anatomy Project) or collaborative groups such as the Genome Network Project group at the genome Sciences Centre at RIKEN, Japan [1]. Data generated by these consortia include microarray, CAGE (capped analysis of gene expression), SAGE (serial analysis of gene expression) and MPSS (massively parallel signature sequencing) as well as cDNA and expressed sequence tag (EST) libraries. The diversity of data types offers opportunity to capture several views on concurrent biological events, but without standardization between these platforms and data types, information is lost, reducing the value of comparison between systems. The ter-minology used to describe data provides a means for the integration of different data types such as EST or CAGE.
An ontology is a commonly used method of standardization in biology. It is often defined as a formal description of entities and the relationships between them, providing a standard vocabulary for the description and representation of terms in a particular domain [2,3]. Given a need and obvious value in the comparison of gene expression between species, anatomical systems and developmental states, we have set out to discover the potential and applicability of such an approach to compare mouse and human systems.
Many anatomical and developmental ontologies have been created, each focusing on their intended organisms. As many as 62 ontologies describing biological and medical aspects of a range of organisms can be obtained from the Open Biomedical Ontologies (OBO) website [4], a system set up to provide well-structured controlled vocabularies of different domains in a single website. The Edinburgh Mouse Atlas Project (EMAP) [5] and Adult Mouse Anatomy (MA) [6] ontologies are the most commonly used ontologies to describe mouse gene expression, representing mouse development and adult mouse with 13,730 and 7,702 terms, respectively. Mouse Genome Informatics (MGI), the most comprehensive mouse resource available, uses both ontologies. Human gene expression, however, can be represented as developmental and adult ontologies by the Edinburgh Human Developmental Anatomy (HUMAT) ontology [7], consisting of 8,316 terms, and the mammalian Foundational Model of Anatomy (FMA) [8], consisting of more than 110,000 terms. Selected terms from the above ontologies have been used to create a cross-species list of terms known as SOFG Anatomy Entry List (SAEL) [9]. Although these ontologies more than adequately describe the anatomical structures of the developing organism, with the exception of SAEL, they are structured as directed acyclic graphs (DAGs), defined as a hierarchy where each term may have more than one parent term [6]. The DAG structure adds to the inherent complexity of the ontologies, hampering efforts to align them between two species, making the process of a comparative study of gene expression events a challenge.
Efforts are being implemented in order to simplify ontologies for gene expression annotation. The Gene Ontology (GO) Consortium's GO slim [10] contains less than 1% of terms in the GO ontologies. GO slim is intended to provide a broad categorization of cDNA libraries or microarray data when the fine-grained resolution of the original GO ontologies are not required. Another set of simplified ontologies are those from eVOC [11]. The core eVOC ontologies consist of four orthogonal ontologies with a strict hierarchical structure to describe human anatomy, histology, development and pathology, currently consisting of 512, 180, 156 and 191 terms, respectively. The aim of the eVOC project is to provide a standardized, simplified representation of gene expression, unifying different types of gene expression data and increasing the power of gene expression queries. The simplified representation achieved by the eVOC ontologies is due to the implementation of multiple orthogonal ontologies with a lower level of granularity than its counterparts.

Mammalian development
The laboratory mouse is being used as a model organism to study the biology of mammals [12]. The expectation is that these studies will provide insight into the developmental and disease biology of humans, colored by the finding that 99% of mouse genes may have a human ortholog [13], and cDNA libraries can be prepared from very early mouse developmental stages for gene expression analysis.
The study of developmental biology incorporates the identification of both the temporal and spatial expression patterns of genes expressed in the embryo and fetus [14]. It is important to understand developmental gene expression because many genetic disorders originate during this period [13]. Similarities in behavior and expression profiles between cancer cells and embryonic stem cells [15] also fuel the need to investigate developmental biology.
Using mice as model organisms in research requires the need for comparison of resulting data and provides a means to compare mouse data to human data [13]. The cross-species comparison of human and mouse gene expression data can highlight fundamental differences between the two species, impacting on areas as diverse as the effectiveness of therapeutic strategies to the elucidation of the components that determine species.

Cross-species gene expression comparison
Function of most human genes has been inferred from model organism studies, based on the transitive assumption that genes sharing sequence similarity also share function when conserved across species [16]. The same principle can be applied to gene regulation. The first step is to find not only the orthologs, but the commonly expressed orthologs. We predict that although two genes are orthologous between human and mouse, their expression patterns differ on the temporal and spatial levels, indicating that their regulation may differ between the two species.
The terminology currently used to annotate human and mouse gene expression can be ambiguous [17] among species, which is a result of different ontologies being used to annotate different species. Although the EMAP, MA, HUMAT and FMA ontologies describe the anatomical structures throughout the development of the mouse and human, their complexities complicate the alignment of the anatomy between the two species. With the alignment of terms between a mouse and human ontology, the data mapped to each term become comparable, allowing efficient and accurate comparison of mammalian gene expression. A SAEL-related project, XSPAN [18], is aimed at providing a web tool to enable users to find equivalent terms between ontologies of different species. Although useful, the ontologies used describe only spatial anatomy and are not temporal.
We have attempted to address the issue by developing simplified ontologies that allow the comparison of gene expression between human and mouse on a temporal and spatial level. The distribution of human and mouse anatomy terms across development match the structure of the human adult ontologies that form the core of the eVOC system.
Due to the ambiguous annotation of current gene expression data between human and mouse, and the lack of data mappings accompanying the available ontologies, the ontologies presented here have been developed in concert with semiautomatic mapping and curation of 8,852 human and 1,210 mouse cDNA libraries. We have therefore created a resource of standardized gene expression enabling cross-species comparison of gene expression between mammalian species that is publicly available.

Ontology development
The ontologies were originally created to accommodate requests by the FANTOM3 consortium [19] for a simple mouse ontology that could be used in alignment to the human eVOC ontologies. The FANTOM3 project was a collaborative effort by many international laboratories to analyze the mouse and human transcriptome. The aim was to generate a transcriptional landscape of the mouse genome that led to the evolutionary and comparative developmental analysis in mammals. The ontologies presented here provided the FANTOM3 consortium with a platform to compare the human and mouse transcriptome in the context of mammalian development.
Shared structure between the ontologies ensures effective interoperability on the developmental and species levels. The importance of shared structure between two ontologies becomes apparent when attempting to align them for comparison. If two terms in an ontology are mapped to each other, ontology rules infer that the children terms in each of the ontologies share the same characteristics. For example, if gene X is mapped to 'heart' in a human ontology and gene Y is mapped to 'cardiovascular system' in mouse, we can infer that because 'cardiovascular system' is the parent of 'heart' in both ontologies, gene X and gene Y have an association with respect to their expression in the cardiovascular system although their annotations are not identical. This is especially important when the granularity of annotation in one species is different to that of another.
Terms from the EMAP, MA and HUMAT ontologies have been used to create 28 mouse and 23 human ontologies, representing the 28 Theiler stages and 23 Carnegie stages of mouse and human development, respectively. The 28 Theiler stages represent mouse embryonic, fetal and adult anatomical development, whereas the 23 Carnegie stages represent only human embryonic development. Human adult is represented by the Anatomical System ontology of the eVOC system, upon which the other ontologies are based. The terms from the source ontologies (EMAP, MA and HUMAT) have been mapped to the equivalent term in the developmental eVOC ontologies to ensure interoperability between external ontologies and eVOC. Terms from the mouse have also been mapped to those from human to enable cross-species comparison of the data mapped.
The integration of the ontologies is described in Figure 1, where 'Mouse eVOC' refers to the individual mouse ontologies and 'Human eVOC' refers to the individual human ontologies (including the adult human ontology). The EMAP and MA ontologies represent mouse pre-and post-natal developmental anatomical structures, respectively, and, therefore, exhibit no commonality. The mouse developmental eVOC ontologies integrate the two ontologies by containing terms from, and mappings to, both the EMAP and MA ontologies. Of the 2,840 terms in the individual mouse ontologies, 1,893 and 237 map to EMAP and MA, respectively. The human developmental eVOC ontology is an untangled version of the HUMAT ontology and has one-to-one mappings to the mouse developmental ontology, providing a link between the terms and data mappings between the mouse and human ontologies.
The presence of species-specific anatomical structures posed a challenge when aligning the mouse and human terms. An obvious example is the presence of a tail in mouse but not in human. We decided that there would simply be no mapping between the two terms. Further challenges involved structures such as paw and hand. The two terms cannot be made identical because it is incorrect to refer to the anterior appendage of a mouse as a hand. However, due to the fact that the mouse paw and human hand share functional similarities, the two terms are not identical, but are mapped to each other based on functional equivalence.
In order to provide simplified ontologies, the 28 mouse and 23 human ontologies were merged to create two ontologiesone for each species. In addition, a Theiler Stage ontology was created that represents the Theiler stages of mouse development. The human stage ontology is represented by the current eVOC Development Stage. A cross-product of two terms (one from the merged and one from the stage ontology) for a species can, therefore, represent any anatomical structure at any stage of development.
The relationship between the developmental mouse and individual ontologies is illustrated in Figure 2, where the term 'brain' is mapped to 12 terms in the individual ontologies and, therefore, occurs in 12 of the 28 Theiler stages. All terms in the individual ontologies that are derived from EMAP or MA for mouse, and HUMAT for human are mapped to the corresponding term by adding the term's accession from the external ontology as a database cross-reference in the eVOC ontologies. Figure 3 shows that the database cross-reference is the accession of the EMAP term, indicating that 'intestine' of the 'Theiler Stage 13' ontology is equivalent to the term represented by 'EMAP:600'. This feature allows cross-communication, and thereby integration, of the EMAP, MA, HUMAT and eVOC ontologies.
The ontologies presented here are simplified versions of existing human and mouse developmental and adult ontolo-gies, containing 1,670 and 2,840 terms, respectively. Table 1 shows the number of terms and database cross-references for the individual mouse and human ontologies. The Theiler Stage 4 ontology contains 12 terms and has 9 mappings to the EMAP ontology. The mouse and human stages have been aligned in the table, showing that mouse Theiler stage 4 is equivalent to human Carnegie stage 3, based on morphological similarities during development [20]. The Carnegie Stage 3 ontology contains 13 terms and has 11 mappings to the HUMAT ontology. The difference in the number of ontology terms and external references is attributed to the addition of terms to maintain the standard structure of the eVOC system. In this example, the term 'germ layers' is in the eVOC ontologies, but not in the EMAP or HUMAT ontologies. Many eVOC terms are mapped to more than one term in the external referencing ontology as an artifact of the simplification of the ontologies, resulting in a one-to-many relationship between eVOC and its reference ontology. For example, 'myocardium' at Theiler stage 12 in the eVOC ontologies is mapped to five EMAP identifiers. Each EMAP identifier references a cardiac muscle, but at a different location. eVOC does not distinguish between cardiac muscle of the common atrial chamber (EMAP:337) and cardiac muscle of the rostral half of the bulbus cordis (EMAP:330). Compared to their counterparts, the Developmental eVOC ontologies represent 22% of both the human HUMAT and mouse EMAP ontologies, with the only relationship between the terms being 'IS_A'. Note that relationships within the eVOC ontologies indicate only an association between parent and child term and do not systematically distinguish between is_a or part_of relationships. As eVOC moves to adopt relationship types from the OBO Relation Ontology [21], relations will be reviewed and curated. Using a principle of data-driven development, eVOC terms are added at an annotator's request, resulting in a dynamic vocabulary describing gene expression.

Data mapping
The resources providing ontologies to annotate gene expression do not always provide the data themselves. In order to obtain mouse and human data, one would have to search separate databases for each species. An example of this would be searching MGI for mouse gene expression data, and ArrayExpress for human. Apart form having to access different databases to obtain data, the terminology used to describe the data is ambiguous and differs in the level of granularity, impacting on the accuracy of inter-species data comparison. The ontology terms have, therefore, been used to annotate 8,852 human and 1,210 mouse cDNA libraries from CGAP [22].
The mapping process revealed inconsistencies in the annotation of the human and mouse CGAP cDNA libraries, requiring manual intervention and emphasizing the need for a standardized annotation. All genes associated with the libraries have been extracted by association through UniGene. A gene was considered to be associated with a cDNA library if at least one EST was evident for the gene in a particular library. The result is a set of 21,152 human and 24,047 mouse genes from UniGene that are represented by CGAP cDNA libraries and annotated with eVOC terms, and represent the set of human and mouse genes for which there is expression evidence. CGAP represents an ascertainment bias where there is a strong over-representation for cancer genes, and, therefore, future efforts for this research will include obtaining a wellrepresented, evenly distributed dataset of human and mouse gene expression. The list of human and mouse orthologs were extracted from HomoloGene to represent the 16,324 human-Venn diagram illustrating the integration of mouse and human ontologies represented by the eVOC system Figure 1 Venn diagram illustrating the integration of mouse and human ontologies represented by the eVOC system. The total number of terms in each ontology is in parentheses. The numbers in each set are the number of terms in the intersection represented by that set. 'Mouse eVOC' represents the 28 individual mouse ontologies and 'Human eVOC' represents the 23 individual human and adult ontologies; therefore, the numbers in parentheses refer to the total number of terms in all the eVOC ontologies for each species. The intersection of the Mouse eVOC with the EMAP and MA ontologies represents the number of terms in Mouse eVOC that have database cross-references to EMAP and MA. Similarly, the intersection of the Human eVOC and HUMAT sets represents the number of Human eVOC terms that map to HUMAT terms. The number within the arrows represents the number of mapped human and mouse eVOC terms. mouse orthologs. Two genes were considered to be orthologs if they shared the same HomoloGene group identifier.

Data mining
Genes may be categorized according to their eVOC annotation on a spatial or temporal level, or a combination of both. An example of this would be genes expressed in the heart at Theiler stage 26 for mouse. For the purposes of this study, we searched for human-mouse orthologs that are expressed in the normal postnatal and developmental brain of both species, where a gene is classified as normal if its originating library was annotated as 'normal'. Research involving gene expression of the brain aims at identifying causes of psychological and neurological diseases, many of which originate during development. With the use of mice as model organisms in this kind of research, it is important to identify genes that are co-expressed in human and mouse on the temporal and spatial levels. The results of our analysis show that of the available 16,324 human-mouse orthologs, 14,434 can be found in CGAP libraries for both human and mouse. When looking at brain gene expression, we could segregate genes according to their spatial and temporal expression patterns. We found that of all the orthologs expressed in the brain, 10,980 genes were expressed in the post-natal brain of both species whereas 1,692 genes were expressed in the developing brain of both species. Of these two sets of genes, 90 genes were found to have biased expression for developmental brain ( Table 2) where developmentally biased genes are those that are expressed during development and not the post-natal organism in either human, mouse or both species (see Additional data file 1 for illustration). The 9,378 genes found to have a bias for post-natal brain gene expression can be found in Additional data file 2. It is important to note that only genes whose orthologs also have expression evidence were considered for analysis. This small number of genes found to be biased for expression during brain development in both species may be a result of data-bias due to the difficulty involved in accessing developmental libraries. Our future efforts will include expanding the data platforms to provide data that are representative of the biology. This analysis does, however, demonstrate the usefulness of the ontologies in performing cross-species gene expression analyses. The GO categories that are highly associated with the 90 genes biased for developmental brain expression were extracted with the use of the DAVID bioinformatics resource [18]. The human representatives of the human-mouse orthologs cluster with GO terms such as 'nervous system development' and 'cell differentiation', suggesting a shared role for development of the mammalian brain, and, therefore, may be potential targets for the analysis in neurological diseases. Given the existence of ascertainment bias on these kinds of data, it was still surprising to see how many genes passed the stringent selection criteria. Searching the Online Mendelian Inheritance of Man (OMIM) database implicated some of the 90 genes, such as GOPC, ARX and DEK, in diseases such as astrocytoma, lissencephaly and leukemia.
To assess the similarity in expression across major human and mouse tissues other than brain, the expression profiles of the 90 genes with bias for developmental expression were determined for developmental and adult expression in the following tissues: female reproductive system, heart, kidney, liver, lung, male reproductive system and stem cell. These tissues were chosen based on the availability of data for each tissue in the developmental and adult categories. For each ortholog-pair, we determined the correlation between their expression profiles (Additional data file 3). We found that, according to the cDNA libraries, one mouse gene was found to be expressed in all the tissues in both post-natal and development (Twsg1), and three mouse genes were expressed only in the mouse brain (Resp18,Gm872,Barhl1) as opposed to all other tissues (see Additional data file 4 for expression profile). The highest correlation score between an ortholog-pair is 0.646 (HomoloGene identifier: 27813), having identical expression profiles during development (expressed in liver and stem cell), but differing during post-natal expression (expression in mouse heart, kidney and stem cell but not in their human counterparts). The correlations observed suggest that the expression profiles of orthologs across these major tissues are only partially conserved between human and mouse. This finding strengthens our understanding of orthologous gene expression in that although two genes are orthologs, they do not share temporal and spatial expression patterns and, therefore, probably do not share a majority of their regulatory modules [23].
Developmental gene expression may be subdivided into embryonic and fetal expression, which in turn may be categorized further according to the Theiler and Carnegie stages for mouse and human, allowing a high-resolution investigation of gene expression profiles between the two species. This stage by-stage expression profile for human and mouse will allow investigation into common regulatory elements of codevelopmentally expressed genes and give new insight into the characterization of the normal mammalian developmental program.

Conclusion
The developmental mouse ontologies were developed in collaboration with the FANTOM3 consortium to have the same structure and format as the existing human eVOC ontologies The first three columns display the individual mouse ontologies, the number of terms in each ontology, and the number of external references of each. The last three columns display the individual human ontologies, the number of terms, and the number of external references of each. The external references refer to the EMAP and MA ontologies for mouse, and to HUMAT for human. The alignment of the rows between the mouse and human ontologies represents the alignment of the Theiler and Carnegie stages of development based on morphological similarities. For example, the Theiler Stage 4 ontology contains 12 terms and has 9 mappings to the EMAP ontology. Mouse Theiler Stage 4 is equivalent to human Carnegie Stage 3. The Carnegie Stage 3 ontology contains 13 terms and has 11 mappings to terms from the HUMAT ontology.
to enable the comparison of developmental expression data between human and mouse. The developmental ontologies have been constructed by integrating EMAP, MA, the developmental Human Anatomy and the human adult eVOC ontol-ogies. The re-organization of existing ontological systems under a uniform format allows the consistent integration and querying of expression data from both human and mouse databases, creating a cross-species query platform with one- The table lists the HomoloGene group identifier, Entrez Gene identifier and gene symbol of the 90 human-mouse orthologs found to have an expression bias towards the embryonic and fetal stages of brain development, without expression during postnatal development. Genes were considered for analysis only if they have an ortholog, and if the ortholog also has expression evidence based on eVOC annotation.

Genes showing developmental expression bias in human and mouse brain
to-one mappings between terms within the human and mouse ontologies.
The ontologies have been used to map human and mouse gene expression events, and can be used to identify differential gene expression profiles between the two species. In future, the ontologies presented here will be used to investigate the transcriptional regulation of genes according to their characteristics based on developmental stage, tissue and pathological expression profiles, providing insight into the mechanisms involved in the differential regulation of genes across mammalian development.

Ontology development
The ontologies were constructed using the COBrA [24] and DAG-edit [25] ontology editors. Each term has a unique accession identifier with 'EVM' as the namespace for mouse and 'EV' for human, followed by seven numbers. This is consistent with the rules defined by the GO consortium [26].

Data mapping
Mouse and human cDNA libraries were obtained from the publicly available CGAP resource and mapped (semi-automated) to the entire set of eVOC ontologies. The eVOC ontologies consist of Anatomical system, Cell type, Developmental stage, Pathology, Associated with, Treatment, Tissue preparation, Experimental technique, Pooling and Microarray platform. The 'age' annotation of the mouse CGAP libraries was manually checked against the Gene Expression Database (version 3.41) [27] to determine the Theiler stage of each library. Due to the lack of a resource providing the Carnegie stage annotation for cDNA libraries, the human cDNA libraries were annotated according to the age annotation originally provided by CGAP. Genes associated with each mouse and human cDNA library were obtained from NCBI's UniGene [28]. A list of human-mouse orthologs were obtained from HomoloGene (build 53) [29].

Data mining
The genes were filtered according to the presence or absence of expression evidence and homology. A gene passed the selection criteria if it has an ortholog and if both genes in the ortholog pair have eVOC-annotated expression. According to eVOC annotation, genes were categorized into those that showed expression in normal adult brain and those expressed in normal developmental brain, many genes appearing in more than one category. Genes expressed in normal adult brain were subtracted from those with expression in normal developmental brain to establish genes whose expression in the brain occurs only during development. The expression profiles of the developmentally biased genes annotated to female reproductive system, heart, kidney, liver, lung, male reproductive system and stem cell for post-natal and developmental expression were determined according to the eVOC annotation of the cDNA libraries, and the correlation coefficient of the ortholog-pairs were calculated.