Bioinformatics tools and techniques depend directly or indirectly upon experimental data. However, interpreting experimental data often requires access to significant amounts of additional information about the sample used in the experiment, the conditions in which measurements were taken, the equipment used to take the measurements, etc. Recent proposals for models that capture such experimental descriptions alongside experimental results include MIAME for transcriptome data [1] and PEDRo for proteome data [2]. However, if full use is to be made of such rich data models for genomic data, these models must also be associated with comprehensive software tools for data capture, dissemination and analysis.

In proteomics, which is rapidly evolving into a high-throughput experimental approach, there is (as yet) no standard representation for experimental data. As a result, limited tool support is available for disseminating, searching, comparing or analysing the results of experiments conducted using different techniques and equipment in different laboratories. Thus, while experimental results can be analysed, often in a labour-intensive manner in-house, the development of bioinformatics techniques for archiving, sharing and wider exploitation of proteomics results is still in its infancy. This paper seeks to contribute to the development of effective and systematic support for proteome data management by:

1. Describing a database for storing, searching and disseminating experimental proteomics data. This material should be relevant to the developers of future proteome data management systems in that it discusses and illustrates various design and implementation decisions that have an impact on the role and maintenance of the resulting database.

2. Making available data sets from several labs whose data have been included in the initial release of the database. These data sets themselves result from substantial experimental activities, and are representative of the sorts of information that in-house and public proteome data repositories must capture. As the database stores data in an XML format that conforms to the PEDRo (P roteomics E xperimental D ata R epo sitory) data model [2], this material provides concrete examples for other users of data that conform to this schema, and should be useful for validation of specific parts of the model as input to the Human Proteome Organisation Proteome Standards Initiative (HUPO-PSI) activity on models for proteome data [3].

The database described in this paper has similar objectives and functionality to various other databases for functional genomic data. In particular, like the Gene Expression Omnibus (GEO) [4], the Stanford Microarray Database [5] and ArrayExpress [6], it contains a single category of experimental data, while accommodating the production of that data using several different experimental techniques. Like ArrayExpress, and unlike GEO, for example, the data stored in PEDRo must conform to a rich, but nevertheless deliberately constraining, data model. This model is richer than that supported by the well established SWISS-2DPAGE database [7] in that it not only contains information on protein separation and identification, but also includes detailed descriptions of experimental samples, the mass spectrometric analyses conducted, and the software used to perform protein identifications.

Establishing the most appropriate kinds of data to include in a database such as PEDRo is not straightforward, as this depends on the use that is to be made of the data. In a large data repository, users may want to search for results based on widely varying criteria – for example, the proteins identified, the change in the level of a protein over time, the mechanism by which a sample was studied, etc. Furthermore, the users of a proteome data repository may themselves be diverse, and include: experimentalists with minimal direct experience of proteomics, but who are interested in proteins or organisms for which proteome studies have been conducted; proteome scientists who want to identify how successful specific techniques have been in different contexts; or mass-spectrometric analysts who want to compare their results with those of others.

This wide range of potential users encourages the creation of a rich repository for proteomics data that provides detailed descriptions of many different aspects of an experiment. However, populating a database such as PEDRo is not a trivial task, as several of the different kinds of data included in PEDRo currently have to be entered manually, which is time-consuming for data providers. Even though a data entry tool has been developed to ease data entry (available from [8]), experience populating the database suggests that the creation of a data set from scratch (e.g., for a sample analysed using a single gel, for which multiple identifications have taken place) can take around a week, but that creating subsequent data sets that share some aspects of the experimental set-up is significantly less time-consuming. In addition, widespread deployment of a standard model should lead to laboratory equipment, or associated software, producing data that conforms to the standard, so the longer-term position for high-throughput laboratories should involve much lower data capture costs. It is hoped that the early provision of a collection of data sets conforming to the widely discussed PEDRo model will be useful in informing ongoing activities on the HUPO-PSI proteome data standard [3].

Many bioinformatics databases, such as UniProt [9] and PDB [10], are associated with file formats that can be parsed by software that analyses or displays the data from the database. The Extensible Markup Language (XML [11]) has been developed in part to make the description, parsing and display of such files more systematic; thus there is a trend in bioinformatics towards the use of XML for storing or transmitting biological data [1, 2].

The PEDRo database makes extensive use of XML for capturing, transmitting, storing and searching proteomics data. In particular:

1. The data-capture process uses a software tool, illustrated in Figure 1, which prompts users for values for different fields, and includes facilities for importing substantial data files, such as those representing peak lists. The tool constructs data-entry forms from the XML Schema definition of the PEDRo model. An XML Schema describes the structure of an XML document, and thus makes explicit the hierarchical structure of the document, the elements that are contained within the document, the types of those elements, and the number of times different elements may occur. The result of the data capture process is thus an XML file that corresponds to the PEDRo schema. A fragment of the XML format for a PEDRo entry is provided in Figure 2.

2. The database stores the XML captured using the data entry tool directly, using Xindice [12], an open-source XML storage system. Several different storage options exist for XML data, including: (i) storing the XML directly in a native XML repository such as Xindice; (ii) storing the XML directly using the XML storage extensions provided by commercial relational database vendors; and (iii) mapping from the XML documents onto tables for storage in a relational database. We have chosen option (i) for PEDRo. Option (ii) was not adopted because there is not yet a standard for integrating XML storage with relational databases, although this is being developed [13]. Option (iii) was not adopted because we envisage that the data model used in PEDRo will evolve to reflect the HUPO-PSI standard [3], and we wanted to avoid the need to evolve both relational and XML versions of the database in parallel. Furthermore, the emphasis for the PEDRo database is on enabling users to identify relevant experimental data sets, rather than on conducting complex searches or analyses over such data sets. For the required tasks, the query facilities provided with XML databases such as Xindice, which tend for the meantime to be based upon XPath [14] are sufficient.

3. The data are presented to users by generating web pages from the stored XML using XSLT [15], which was designed to support exactly this sort of task. This means that it has been straightforward to develop reports from the stored form of the data. Furthermore, the download format for the data is as XML documents, in the hope that this will ease the development of tools for parsing and analysing data obtained from the database.

The software components used within PEDRo (and the role they play in data capture, storage and dissemination) are illustrated in Figure 3.

These PEDRo data are significant in biological terms. For example, they include the first direct comparison of proteomic responses in two fungal species, namely responses to amino-acid starvation in the baker's yeast S. cerevisiae and the pathogenic fungus, C. albicans [16]. In addition, they include the first proteomic analysis of the medically important pathogen C. glabrata (Stead et al., Proteomic changes associated with inactivation of the Candida glabrata ACE2 virulence-moderating gene, manuscript submitted), whose genome sequence has only just been completed. The Streptomyces coelicolor M600 data set is the largest proteomics time course analysis of this strain in terms of numbers of proteins identified. It adds significantly to our knowledge of expression of some of the 20 gene sets annotated as being determinants of the biosynthesis of secondary metabolites, including antibiotics. Somewhat similar experiments, but differing in many aspects of their metadata, are reported on the SWICZ database [17]. This provides an opportunity to evaluate PEDRo in the context of related data presented in different databases. Also included are data from an experiment investigating the proteomic analysis of the mouse jejunal epithelium and its response to infection with the intestinal nematode, Trichinella spiralis [18].

Web-based interfaces to biological databases tend to support one or more of the following tasks: browsing – interactively listing or navigating through database entries; searching – identifying database entries on the basis of simple restrictions on the values of one or more fields; visualising – presenting a visual representation of the data as a starting point for browsing; or querying – specifying a search that is to be conducted over the database using a query building interface or by providing inputs to pre-written (or "canned") queries. Functional genomics databases tend to emphasise browsing and searching. For example, the Stanford Microarray Database [5] supports browsing based around organisms and experiments, and more complex Boolean searches based on criteria such as experimenter, organism and category of experiment. ArrayExpress [6] supports browsing through experiments, arrays and protocols, and searching based on criteria such as species, experiment type and author. SWISS-2DPAGE [7] supports browsing by clicking on spots on gels, and searching based on criteria such as description, accession number or author.

PEDRo also emphasises browsing and searching. Figures 4 and 5 illustrate the web-based interface to PEDRo, which can be accessed at [8]. In essence, the records in the database can be accessed by browsing summaries of the entries in the database, or by searching using one or more criteria. These criteria were obtained through a systematic requirements analysis with potential users from several different research groups, who were asked to comment on early versions of the interface. Overall, PEDRo provides core data access facilities that are principally intended to allow users to identify data sets that are of interest to them. As such, the PEDRo database as described should not be seen as a comprehensive query or analysis environment for proteomics data, but rather as a repository through which experimental results can be made available to a wider community. Therefore, S. cerevisiae data from PEDRo will also be made available through GIMS [19], for example, to enable the integration of these data with other sequence and functional information.

A significant motivating factor behind the development of the PEDRo repository has been to allow informed discussion, assisted by concrete examples, into the level of detail and forms of model that are most appropriate for a proteome data repository. As the PEDRo model is being used as the starting point for the HUPO-PSI activity on models for proteome data, early validation of this model is important. The following observations have been made about the PEDRo model during the data capture process:

1. Sample description is neither very precise nor systematic. The effective description of samples is an open issue that spans different kinds of functional genomic data. For example, work is underway on the development of an ontology for characterising microarray experiments, focusing, in particular, on samples [20]. However, as the variety of organisms, genetic manipulations, extraction techniques, environmental conditions and experimental manipulations that may characterise a sample are extremely large, a mature solution to this problem may be some way off.

2. There is only limited support for relative protein abundance data (e.g. DIGE and stable isotope labelling strategies). Thus, for example, there is no place in the model to describe an expression ratio for a protein species derived from quantitative experimental strategies, only the ability to capture the 'raw' numbers. In fact, the PEDRo model was not designed to capture expression ratios, partly because such numbers are easily derived from the captured primary data, and partly because the particular method of their derivation may be contentious. It is hoped that the HUPO-PSI model will provide generic constructs for representing relationships between certain kinds of measurement (e.g. relative protein expression readings), to which can be attached the specific detail for individual techniques. However, it also seems important to avoid the pitfalls associated with overly permissive models, as these provide a less stable foundation for the developers of analytical tools than their more proscriptive counterparts.

3. The gel model is not particularly detailed. Thus, for example, there is no detailed description of the image analysis software used, the descriptions of individual spots are fairly minimal, and no details are captured on spot excision. An earlier critique of the PEDRo model for gels, and some possible extensions, is provided by [21]. It seems that, in order to provide insights for the developers of gel-based experiments, it would be appropriate for the model to be revised to provide additional details on gels.

Overall, the appropriate level of detail for a proteomics repository is somewhat subjective, but can usefully be based on guiding principles; agreement as to the principles should then avoid scope-based discussions at a very fine-grained level. The current PEDRo model essentially supports the principle that enough detail should be captured about an experiment to:

i. Allow results of different experiments to be analysed/compared.

ii. Allow suitability of experiment design and implementation decisions to be assessed.

iii. Allow protein identifications to be re-run in the future with new databases or software.

There is also an additional negative principle, to the effect that the model itself should not be designed to include dependencies on characteristics relating to the configuration or properties of an individual piece of equipment. Accordingly, we have attempted to allow experimental methods and results to be described in significant detail, but without including parameters and properties that are likely to be superseded rapidly when new models of equipment are introduced, and without including parameters that can only be understood with reference to the documentation of a particular product.

The data stored in PEDRo is more comprehensive for each experiment than is the case for most existing proteome databases. For example, in the longest established experimental proteomics resource, SWISS-2DPAGE [7], the emphasis is on annotated gels, and there is much less information collected on how the annotations were arrived at. Furthermore, there is an architectural distinction – SWISS-2DPAGE follows a more federated approach, with individual sites continuing to hold their own data. These other proteome data sources can be accessed through WORLD-2DPAGE, a web resource listing sites making available experimental proteomics data [22]. An example of a database that participates in WORLD-2DPAGE is the University of Alabama (UAB) Proteomics Database [23] which provides search and browsing facilities over data from its host university. As such, the emphasis is on annotated gels, and relatively few details are captured on sample processing, mass spectrometry or in silico analysis. Such design decisions are appropriate for certain categories of user of a proteomics database, but not for others. The UAB database has been designed to provide access to processed experimental results for biomedical researchers, but does not provide enough information to allow detailed comparisons of the ways in which the results were obtained.

The ProteomeWeb [24] provides a wider range of tools than PEDRo (for example, for computing theoretical maps), and supports browsing of annotated gels from several bacteria and archaea. Once again, though, the data provided for each experiment are less comprehensive than in PEDRo. ProDB [25] has a certain amount in common with UAB, in that it too provides search and browsing over a database of locally produced data. In addition, ProDB features an architecture that supports the plugging-in of data-loading and analysis tools. However, the level of detail supported by the model is not obvious from the paper, which gives only part of the model, and the database was not publicly accessible at the time of writing. In consisting of a collection of tools associated with a database, ProDB thus also has a certain amount in common with SBEAMS [26] which includes a relational database of proteomic data. The SBEAMS model emphasises the description and analysis of mass spectrometry data, but seems not to support open access to experimental data at the time of writing.

In terms of quantities of data, there are fewer data sets in PEDRo than in SWISS-2DPAGE, reflecting the fact that PEDRo is a newly created resource (Release 16 of SWISS-2DPAGE contains 34 reference maps), but somewhat more than in the UAB Proteomics Database. The Open Proteomics Database (OPD) supports the browsing and downloading of comparable amounts of data to those in PEDRo, and also includes mass spectrometry data, although quite a lot of the data are in flat-file format [27]. However, it is fair to say that none of the current databases is operating in the context of high-throughput experimentation, which will certainly be prevalent in the near future.

The need for wider and more systematic dissemination of experimental proteomics data is widely recognised, as argued in [27], and attested to by the ongoing work of the Proteome Standards Initiative [3]. As such, issues that need to be addressed include:

i. The nature and variety of information that should be recorded about proteomics experiments.

ii. The functionality that should be provided by repositories that make large-scale proteomic data available.

iii. The computational architecture that should be used to provide the functionality at (ii).

iv. The nature of the tools that should be developed for use with such a repository.

This paper has sought to address issues (i), (ii) and (iii), with a particular emphasis on (i). Following on from [2] we believe that the provision of a collection of representative proteomic data sets conforming to a consistent model is important to the ongoing process of developing a stable and effective de jure standard for proteome data representation and sharing. This paper describes a database that includes a rich collection of representative data sets. Furthermore, the paper describes the functionality (issue ii) and architecture (issue iii) of an exploratory system for disseminating such data. In the same way as we see models for representing proteomic data evolving in the light of practical experience, we anticipate that the PEDRo repository, and the overall understanding of the data access and dissemination requirements for proteomic data, will evolve as the opportunities presented by high-throughput experimental techniques and comprehensive data sets become more fully understood.

The database can be accessed using a web browser at http://pedro.man.ac.uk/, by following the Database link.