Bioinformatics in the Semantic Grid
by Kai Kumpf and Martin Hofmann
One of the major challenges in Grid computing is the semantics-driven retrieval of Grid services and distributed data. DB-Annotator is an annotation tool for data in Grid services developed in a collaborative project between the bioinformatics department and the department for Web applications at the Fraunhofer Institute for Scientific Computing (SCAI). Fully annotated data in the Grid is particularly important in bioinformatics.
DB-Annotator was conceived for the Resource Description Framework (RDF) annotation of structured information sources such as relational data or XML-based service descriptions. Only semantically annotated Grid services (GS) provide a means of finding data or compute-services that are suitable for the task at hand. Furthermore, they make the building of workflows (coupled services within the grid) a realistic goal. DB-Annotator will support several levels of semantic annotation to enrich Grid services, ranging from the services themselves to the data within the Grid.
The project DB-Annotator (see Figure 2) was designed for universal annotation of data that can be represented in tabular form, ie data that can be qualified by unique keys.
|Figure 1: Grid Service- Oriented Architecture. The multiple levels of semantic annotation within a Grid-Service Oriented Architecture (gSOA). Path 0 corresponds to the semantic enhancement of the UDDI service registry via RDF. If no additional semantic glue is needed, direct access to either data- or compute-services is provided (1a, 2). When retrieving data, the semantic description can either stem from UDDI or, more finely -grained, from semantic annotation of structured (most often relational) data. 1b/c correspond to retrieval of annotated data; both are equivalent when there exists a central RDF annotation repository that stores n:m ontology-class - (data) - relationships. Whereas 1a provides a complete view ofn the data via full-text search or keys, 1b/c allow querying by content.
The main thrust for the development came from the realization that most interesting data within biological databases resides within free-form text fields and is therefore not easily accesible for data mining. Relational databases (RDB) are usually the only choice for data retrieval within distributed organizations. Categorial description of the typed data residing in RDBs can be derived from the tables (entities) themselves, and from the columns (entity attributes) and the cell contents (instances of attributes). Still more fine-grained information comes from the notorious description fields, but putting this implicit information to use for machine-reasoning is unrealistic at this point.
The goal was thus to provide easy navigation through existing ontologies and data-source-independent RDF annotation of RDB data via drag and drop (see Figure 2). This data categorization can take place on three levels within a database (see Figure 3), not counting the database itself. A unique data key that can be linked to an ontological category will therefore appear as database:schema:table[:column:row], the latter two key parts being optional depending on the depth of the annotation.
|Figure 2: DB-Annotator was designed to allow RDF-annotation of structured data. Navigation through a chosen ontology is facilitated by a DAG view on the classes and their relationships (left). Any data that can be represented in tabular form and is accessible via unique keys can be annotated (upper right). Class attributes have to be filled in on annotation (lower right). In the case shown here the user has chosen a listing of available Grid Services that have to be semantically described. The major field of application for DB-Annotator however, is the annotation of tabular data from RDBs.
This concept of assigning ontology classes to data was inspired by existing annotation software for the Semantic Web. Once a data source has received ample annotation, data mining will be possible by employing standard methods on the categorized data. With annotated RDB, access to data services in the Grid will improve significantly, since querying by categories is enabled across several RDBs at once. Further, every result set will be enhanced with the respective category information, a data basis that could be used to support further automated annotation of newly inserted data.
|Figure 3: Selection modes in DB-Annotator. Data entities for annotation can either be the whole table, one column or one cell. The data key for annotation is constructed depending on scope of the selection. Only (multiple) annotation of a single cell corresponds to an extension of the data model. Selection of a row corresponds to selection of a record from the extended data model on the basis of table and ontology keys and as such need not be considered separately.
Since DB-Annotator produces standard RDF (subject-predicate-object triples like entity_xyz is-a category_abc), its output can also serve to enhance UDDI, a proposal made for the introduction of content-searchable Web services and a possible starting point for GS-workflow construction. Instead of the database unique keys, the RDF would contain qualifiers for the available Grid services. Eventually a semantically enhanced Grid Service-Oriented Architecture (gSOA SOA being a term borrowed from the Web services world) would look like Figure 1. Full categorization of both GS and the exchanged data leaves no more room for ambiguities a precondition for data exchange and manipulation within the Grid network of virtual organizations (VO).
Given the bulk of biochemical entities that go under various names and interact on different levels, the relevance of fully annotated data Grid services in bioinformatics, especially systems biology, is enormous.. Until now, ambiguities in synonyms and interactions had to be detected by human intelligence, meaning full-fledged systems biology services won't be successfully automated without proper categorization. Since DB-Annotator draws on a common set of agreed ontologies (eg gene ontology) for the annotation of both data services and data sets or subsets, it can provide the semantic glue on all levels required.
The project was launched in the department for bioinformatics at Fraunhofer SCAI. It is a collaborative project between the bioinformatics department and the department for Web applications. Preliminary work commenced late in 2003 with the aim of producing a prototype that would capture the basic functionality of the future system. The prototype, featuring 1:n mapping of a database to several ontologies and RDF flatfile storage, was completed along with a detailed requirement analysis this summer. A new structured development from scratch has now started and is expected to be finished early in 2005. Among the things that will be included in the new phase are n:m relationships between data and ontology classes and storage in a single, structured relational repository. The product is currently intended to provide support for semantic annotation of biological and chemical databases. The possibilities for Grid services annotation within UDDI are explored within the semantic grid research group under the auspices of the Global Grid Forum (GGF).
Globus Toolkit: http://www.globus.org
Global Grid Forum: http://www.gridforum.org
Gene Ontology: http://www.geneontology.org
Kai Kumpf, Institute for Algorithms and Scientific Computing (SCAI), Fraunhofer ICT Group, Germany
Tel: +49 2241 14 2257