The integration of bio-molecular data from diverse sources such as public databases or clinical parameters (annotation) is a key challenge in the process of the analysis of high-throughput experiments. While the interpretation of the outcome of a standard experiment used to depend on the knowledge of one human expert from the field, today's screening tools produce data quantities not manageable by human inspection. After receiving a list of differentially regulated genes from a microarray gene expression experiment comprising several hundreds or thousands of entries, it is not efficient to start the interpretation of these results by manually searching through publications. As a first step, broad biological themes should be identified and followed into a more detailed inspection.

Using network technologies, the availability of data sources is no more the limiting factor, but if accomplished manually, the obstacles for their integration are numerous. Often data are made available in different formats (HTML pages, flat files, direct database access) and very heterogeneous layouts. In addition the nomenclature (e.g. gene names) as well as the relationship of different systems to each other are often inconsistent and require many manual selection and modification steps. At the same time, as much knowledge as possible should be integrated about the data features analyzed, since interesting unknown pathways and interconnections might be hidden behind biological complexity.

As the first aspect, FACT accomplishes the task of integrating heterogeneous information sources by abstraction from the specific data types to one basic concept. (figure 1, lower part). The smallest entities are data features, which are single items of information, either a name/value pair or additional textual description thereof. This could be a list of ids of clones with their relative expression as measured on a cDNA microarray. They are grouped into data sets, combining data features relating to the same experiment or a group of annotation terms for a certain set. In our example the clone/ratio pairs measured in one hybridization would be stored as one data set. The data sets originate from specific data sources, defining distinct types of experiments or annotation sources. One data source would be "cDNA microarray measurements with textual clone ids and numerical results". This architecture follows the idea, that the primary data must be represented at a sufficient level of abstraction to make the data independent of the source technology 1. The concept applies to experimental as well as to annotation data. Meta-data about the different data sources is stored in dedicated tables of the underlying database, describing the source with date and type of data.

Differences in nomenclature and the problem of relating one type of experiment to an other, as the second obstacle for data integration, has been addressed for gene and protein centered research by the development of the GeneOntology system (GO) 2. This hierarchical framework of a directed acyclic graph of annotations for gene attributes has become a de facto standard, which can be employed in the functional analysis of experiments. Similar projects have been initiated for example to organize the classification of molecular interactions in pathways and molecular complexes (Genome Knowledgebase / Reactome 3). Using these resources FACT is able to compare experimental results from technically distant applications.

However, most experiments differ in focus and design and usually no standard solution for their interpretation can be applied. The third aspect for an integrating approach therefore is high flexibility concerning the application of diverse analysis methods that have already been developed for the interpretation of results or might be used in future.

FACT consists of a MySQL database, a core library (as an Application Programming Interface, API) and various modules written in the language Perl. We also created an interface for the web-based usage of all functions of FACT 4. The database reflects the transformation of heterogeneous data into a generalized format, storing information in DateSet and DataFeature tables (figure 2). The core software framework supplies the basic functionality for data access in an object-oriented fashion and for adding and operating modules. These modules are adaptors and can be classified into three categories (figure 3):

A variety of modules for the handling of different data sources (table 1) as well as for the application of basic data analysis and display functions (table 2) were developed. Most of the current functionality is focused on handling human and murine gene annotation information. Additional functionality can be added in a "plug-and-play" fashion, since new modules can be loaded dynamically into FACT.

To read in experimental results, a simple list (with terms or term-value pairs) or table can be used, or more specialized parser functions can be employed to read and decipher the notations of different types of results. One parser (2_Colums) reads tab-, or semicolon-separated lists and stores the data as terms of the data type that is passed as a parameter (e.g. gene symbol or clone id) and the respective value. Another function (LongList) expects terms only (list of genes that are to be annotated) or Affymetrix probe ids (AffyCelFile). The parser for CGH results translates ISCN notations of cytogenetic alterations 9 into the distinct chromosomal bands that are affected while reading the data file. The bands are stored with the alteration -1 (loss of genomic material), +1 (gain), or +2 (high level amplification).

Data sources that can then be used for annotating these experimental results contain among others EnsEMBL databases 8, with functions providing numerous gene annotations on the human and murine genome (gene name, accession numbers, genomic localization, GO terms and other ids). The annotation data is fetched by using the EnsEMBL API or direct sql queries. Chromosomal locations expressed as cytogenetic bands can be translated in megabasepair positions. This can permit the direct comparison of results from genomic and expression experiments. Most common identifiers (IMAGE IDs, DDBJ/EMBL/GenBank accession numbers, international clone names, MGD (Mouse-Genome Database) IDs or Probe-IDs as used on expression microarrays in the Affymetrix system) are recognized and used by the different annotation modules. Additionally, homologous genes, sequence features, InterPro protein domains, CpG content and PubMed references can be acquired. The euGenes database 10 modules delivers an additional set of broad annotations (Gene Symbol and Name, GDB ID, OMIM ID, Genomic Localization, GO terms, etc.). The BBID module searches for representations of affected pathways in the Biological Biochemical Image Database 11. We store links to the images which sometimes allow the clarification of interactions better than textual description alone. Additionally data is used from STRING 12 and Reactome 3 to point out protein interactions and involvement in molecular complexes.

Annotation steps can be concatenated, allowing deriving from specific (e.g. Affymetrix Probe-IDs) to more general terms (e.g. gene symbols). This broad annotation that is facilitated by FACT is crucial for the researcher to acquire a complete picture of his data. The user can export the combined lists of annotated data in HTML, XML or text format. If desired the system can send them by email.

Modules to correlate these annotated datasets with each other have been developed for FACT, incorporating existing algorithms or presenting new approaches (table 2). For example, a counting procedure (SimpleCount) reports the number of occurrences of each annotation term. The module is independent of the type of data, one or more data sets can be added up and a threshold for reporting can be defined. The results are displayed graphically in a chart and in a table format, directing the researcher's attention to distinct characteristics (figure 4). A more specific approach to interpret list of genes lies in the explicit usage of GO terms. As originally introduced by Khatri et al. in the program OntoExpress 13 there are different methods and implementation to search for those parts of the GO tree that appear more often in the gene list analyzed than by chance alone. For FACT we used an implementation from the GeneMerge program 14 and the GO::TermFinder perl module 15. Occurrences of terms in a gene list are normalized against a background, which might contain the annotations of all terms spotted on a chip or the entire genome, by using the hypergeometric tail probability (with Bonferroni correction if desired, example: figure 5). In our implementation, the function can be applied on GO data as well as on any other kind of annotation to detect overrepresented terms. Alternatively the Fisher's exact test can be used for this purpose. We included part of this method taken from the EASE program 16 in a FACT function. We also developed a combined algorithm, which calculates a K-Means similarity matrix based on the experimental values and reports on significant terms with the Fisher's exact test within the identified groups afterwards (goCluster, G. Wrobel, available at http://www.bioconductor.org). Another analysis module dedicated to the comparison of genomic and expression data runs pairs of two-sided T-Tests on the groups of over- and under-expressed genes against the groups of enhanced/amplified and deleted genomic regions. This allows detecting a correlation between an altered genomic locus and the corresponding expression pattern. FACT can also represent experimental values or frequency counts in the genomic context which can be helpful for the identification and representation of localization effects in the genome. This is achieved by reading in the genomic locations of data sets and generating bar chart images on top of prepared chromosome ideograms (figure 6). The MedLiner module presents a simple literature mining tool which uses the Bio::Biblio perl functions to find citations that are shared by two or more gene (Figure 5). Example output files are also available at 4.

The Flexible Annotation and Correlation Tool has proven especially helpful with the interpretation of results from genomic and expression microarray experiments, but most functions can be applied to a broad variety of experimental data.

We demonstrate the benefits of FACT for the functional interpretation of the results from a comprehensive analysis of gene expression patterns in the development of non-melanoma skin-cancer conducted within our group [L. Hummerich et al.: Identification of novel tumor-associated genes in the process of squamous cell cancer development; submitted]. Using two different sets of microarrays with 15.000 and 20.000 cDNA fragments respectively, the chemically induced multi-step development of squamous cell carcinoma was monitored. We used the dorsal skin of mice as a well-studied system for the development of epithelial cancer 18 with the carcinogen 7,12-dimethylbenz-[a]-anthracene and the tumor promoter 12-O-tetradecanoylphorbol-13-acetate as inducing agents. Expression values were measured at different time-points of tumor formation. Genes with differentiated expression patterns are expected to play a role in the development of human epidermal tumor development as well.

Preprocessed results where loaded into the FACT system as text files containing the murine clone identifier and expression values. Using mainly information from the Ensembl database, FACT annotation functions acquired corresponding gene names, genomic localizations, functional information from Gene Ontology, homologues human gene names and the genomic localization of those genes (supplement S1, complete set of annotated data). We applied different FACT analysis functions to explore the expression data (for example the TermFinder function to search for enriched functional groups of genes, figure 5). The function SimpleCount was used to search for enriched functional categories of annotations. Using the information of homologues genes, we were able to identify the human chromosomal band 1q21, as a region of accumulated differentially expresses genes in murine skin cancer formation (figure 4). The visual representation of the loci using the ChromPlot function highlights these findings (figure 6). With the MedLiner module we were to confirm that several regulated genes were collectively mentioned in previous publications (figure 7). A number of genes (S100A3, S100A6, S100A8, S100A9) which are part of the S100 family of calcium-binding proteins are involved in the regulation of AP-1 and NFκB-dependent transcription. Using FACT we were able to focus our analysis and to gain understanding of the relevance of the results from the microarray study. With the application, it was possible to find human homologues of the murine tumor-associated genes and to confirm the involvement of the S100 protein family in human epidermal malignancies.

There are several large-scale database projects that incorporate an immense spectrum of information about genes and gene product (Ensembl, euGene, LocusLink/EntrezGene, etc.). The Ensembl project for example also allows the user to display his own selected data sources in the context of the full genome annotation through the Distributed Annotation System 19, and it offers the possibility of mining the data of several genomes using the EnsMart software 20. FACT uses Ensembl, LocusLink/EntrezGene and euGene data and complements them with other annotation resources; it allows the user to apply different analysis functions on the combined data.

Recently, a variety of computational tools have been introduced to aid in the interpretation of results, some of which are of interest concerning FACT. The majority of the programs use GO annotations to gain an interpretation of gene expression data. OntoExpress 13 was introduced in 2002 and offers the options to use hypergeometric, chi-square, binomial and Fischer's exact test to score annotation term derived from gene lists. It also allows the appliance of different methods (False Discovery Rate (FDR), Bonferroni, Holm, Sidak) for the multiple experiment correction 17. The program can also include KEGG pathway information and chromosomal localization and is now part of the Onto-Tools collection to offer further functionality 21. EASE 16/ DAVID 22 offer a broad variety of annotation options in their latest version including all major database identifiers, protein domain and pathway information. The Fisher's exact test is used for the detection of enriched terms. GoMiner 23 and numerous other tools listed at the GO website 24 can be used for the annotation of gene lists with GO terms. GeneMerge 14 uses the hypergeometric tail probability with Bonferroni correction to test GO terms, genomic localizations and KEGG pathway information and is used in parts within the FACT system. FACT uses the available Perl code of GO::TermFinder 15 for the GO annotation and the detection of significantly overrepresented terms using the FDR. It also includes a function combining K-means clustering and Fisher's exact test. GEPAS 25 and GECKO 26 are two recently introduced large software packages that include functional analysis and visualization steps. In contrast to FACT they are focused on the initial statistical evaluation and on the analysis of gene expression microarray data. GFINDer 27 is a system that offers annotations on GO, pathway information, protein domains and genetic disorders. It analyses with count functions and appropriate tests (Hypergeometric, Binomial, Fisher's, Z or Poisson Test).

This list of available tools is far from complete and not all aspects are covered. The FACT system was developed with the focus to include and extend the functionality of tools like these. To our knowledge, the individual programs do not offer the same degree of flexibility and openness to different data sources and analysis methods. New functions can be added to FACT by simply uploading the respective module. The system is designed as an open framework for the explorative analysis using a variety of methods on annotational data. It is not restricted to or focused on Gene Ontology-based interpretation or the analysis of gene expression data alone and should facilitate the development and application of new analysis approaches. The system is constantly being extended to include additional aspects. With the submission as an open-source project we want to encourage other researchers to participate in this development.