Identifying functions or pathways shared by genes responsible for cancer is still a challenging task. This paper describes the preparation work for applying Formal Concept Analysis (FCA) to biological data. After gene transcription experiments, we integrate various annotations of selected genes in a database along with relevant domain knowledge. The database subsequently allows to build formal contexts in a exible way. We present here a preliminary experiment using these data on a core context with the addition of domain knowledge by context apposition. The resulting concept lattices are pruned and we discuss some interesting concepts. Our study shows how data integration and FCA can help the domain expert in the exploration of complex data.
Over past few years, large volumes of transcriptomic data were produced but their analysis remains a challenging task because of the complexity of the biological background. In the eld of transcriptomics, biologists analyze routinely the transcription or expression of genes in various situations (e.g., in tumor samples versus non-tumor samples).
Some earlier studies aimed at retrieving sets of genes sharing the same
transcriptionl behaviour with the help of Formal Concept Analysis (see, e.g., [
In this paper, we are interested in applying knowledge discovery techniques for analyzing a di erentially expressed gene set and identifying functions or pathways shared by these genes assumed to be responsible for cancer. Knowledge discovery aims at extracting relevant and useful knowledge patterns from a large amount of data. It is an interactive and iterative process involving a human (analyst or domain expert) and data sources. We show how various gene annotations and domain knowledge are integrated in a database which is then queried for building in a exible way formal contexts. We present here a preliminary experiments using these data. It was performed on a core context with the addition of domain knowledge (by context apposition). The considered domain knowledge are the hierarchical relationships between molecular pathways. Pruning the obtained lattices allows us to retrieve interesting concepts which we discuss. The results obtained from both experiments are also compared.
The plan of the paper is as follows: Section 2 introduces Formal Concept Analysis, Section 3 explains the data resources which are integrated, Section 4 focuses on the application of FCA, Section 5 discusses the results and Section 6 concludes the paper and presents future Work. 2
We introduce here the basics of Formal Concept Analysis that are needed to understand what follows. Let G and M be the set of objects and set of attributes respectively and I be the relation between the objects and the attributes I G M , where g 2 G, m 2 M , gIm is true i the object g has the attribute m. The triple K = (G; M; I) is called a formal context. If A G, B M are arbitrary subsets, then a Galois connection denoted by 0 is given by: A0 := fm 2 M j gIm 8 g 2 Ag B0 := fg 2 G j g I m 8 m 2 Bg (1) (2)
FCA framework is fully described in [
In this section, we introduce and describe the biological data on which we are working. 3.1
Molecular Signature Database (MSigDB) is an up-to-date database which
contains data from several resources such as KEGG, BIOCARTA, REACTOME,
and Amigo [
For our study, we used MSigDB Version 3.0. One entry, shown below in XML format, describes the gene set corresponding to the GO term 'RNA Polymerase II Transcription Factor Activity Enhancer Binding' (all the attribute names are underlined). The Members attribute contains the list of gene symbols belonging to the gene set. MSigDB was chosen as the main source for describing genes because it gathers up-to-date informations about many aspects of human genes.
<GENESET Standard Name =\RNA Polymerase II Transcription Factor Activity Enhancer Binding" Systematic Name = \M900" Historical Names =\" Organism =\Homo sapiens" Geneset Listing URL =\" Chip = \Human Gene Symbol" Category Code =\c5" Sub Category Code =\MF" Contributor =\Gene Ontology" Contributor Org =\GO" Description Brief =\Genes annotated by the GO term GO:0003705. Functions to initiate or regulate RNA polymerase II transcription by binding an enhancer region of DNA." Description Full ="" Members =\ MYOD1, TFAP4, EPAS1, RELA, MYF5, MYEF2, NFIX, PURA, HIF1A" Members Symbolized = \MYOD1, TFAP4, EPAS1, RELA, MYF5, MYEF2, NFIX, PURA, HIF1A" Members EZID =\ 7023, 2034, 5970, 3091" Members Mapping = \ MYOD1, 4654-TFAP4, TFAP4, 7023-EPAS1, EPAS1, 2034-RELA, RELA, 5970-MYF5, MYF5, 4617-MYEF2, MYEF2, 50804-NFIX, NFIX, 4784-PURA, PURA, 5813-HIF1A" Status =\public" > </GENESET> 3.2
Besides the gene annotations included in MSigDB, many types of domain knowledge are interesting to use when analyzing genes. The rst type of such do4 http://www.genome.jp/kegg/pathway.html main knowledge are the hierarchical relationships between GO terms or between KEGG pathways. Indeed, the KEGG hierarchy for human groups the KEGG pathways into 40 categories and 6 upper level categories. Figure 1 illustrates the KEGG hierarchy detailing on one upper-level category and one category.
In our study we have genes described by pathways involving them which may in turn be present in some category of pathways. For example, if a gene is involved in a pathway apoptosis it will also be in the category 'Cell Growth and Death'. In order to facilitate the knowledge discovery, it is important to identify the relevant data sources, organize, and integrate the data at one single database. In our case, the relevant primary data sources are MSigDB, KEGG PATHWAYS database, and AmiGO database. 4
Once the data are integrated in our database the next step is to build formal contexts for applying FCA. Our experiment focuses on applying FCA to a core context describing genes by MSigDB-based attributes and shows its extension based on the addition of domain knowledge. 4.1
The experiments described here are based on three published sets of genes
corresponding to Cancer Modules de ned in [
We apply FCA for analyzing a context describing genes of each Cancer Module with MSigDB-based attributes. Table 2 shows ve genes (involved in Cancer Module 1) as a set of objects described by attributes which are the memberships to gene sets from MSigDB. For example, CCT6A is in the set of genes (gene set) whose standard name is Reactome Serotonin Receptors. Interestingly, by querying our integrated database the analyst is able to select the prede ned gene sets to include in the formal context.
In order to extend the analysis of a list of genes, we need to take into account the domain knowledge. Hence, the same experiment was conducted with the addition of the KEGG hierarchy knowledge to the core contexts resulting in three extended contexts. All KEGG categories and upper-level categories were added as a set of attributes. If a gene is member of a KEGG pathway which in turn belongs to a category and an upper level category then a cross ' ' is added in the corresponding cells in the extended context.
Table 2 shows ve genes (from Cancer Module 1) with the addition of one KEGG category (kc) and one KEGG upper level category (kuc). In the given example CCT6A is involved in pathway KEGG PPAR Signaling Pathway which belongs to the category kc:Endocrine System and upper level category kuc:Organismal Systems. The lattices were generated and the statistics for each Cancer Module are given in Table 3. The concepts were ltered and ranked based on same criteria as in the rst experiment. Data Sets No. of Genes No. of Attributes No. of Concepts Levels Module 1 361 3496 9,588 12 Module 2 378 3496 6,508 11
Module 5 419 3496 5,004 12 5
In this study, biologists are interested in links between the input genes in terms of pathways in which they participate, relationship between genes and microRNAs etc. We obtained concepts with shared transcription factors, pathways, positions of genes and some GO terms. After the selection of concepts with higher support, we observed that there were some concepts with pathways from KEGG and REACTOME as their intent. These pathways are either related to cell proliferation or apoptosis (cell death). The addition of domain knowledge e ectively gives an opportunity to obtain the pathway categories shared by larger sets of of genes. Table 4 shows the top-ranked concepts found in each module. For example, in module 5, we have con rmation that Cytokine Cytokine Receptor Interaction pathway comes under the category Signaling Molecules and Interaction and upper level category Environmental Information Processing (Concept ID 4938). The absolute support and stability of the concept containing only the category Signaling Molecules and Interaction and upper level category Environmental Information Processing as its intent are higher (Concept ID 4995, Table 4) .
To sum up, we were able to discover interesting biological properties of subsets of genes in the three test data sets. As for example, the Focal Adhesion pathway was found to be associated to 17 genes in both modules 1 and 2; the Dataset Concept Intents
ID Module 1 9585 9571 9566 KEGG category Immune System was found to be shared by 11 to 25 genes in the three cancer modules (Table 4). Given the test data sets, these results are hopeful and constitute interesting positive control. This con rms that FCA-based analysis o ers a powerful procedure to deeply explore sets of genes. Our study shows how Formal Concept Analysis can be applied to complex biological data. Data integration and FCA give the exibility of using various types of attributes (pathways, GO terms, positions, microRNAs and Transcription Factor Targets) for analyzing a list of genes. Our approach gives an insight into how domain knowledge can be introduced in the analysis with the help of FCA. As for future work, we plan to apply our approach to experimental gene lists and take into account gene-gene relationships (physical Protein Protein Interactions), term-term relationships (Gene Ontology relationships, namely is-a, part-of, and regulates) and relationships between gene positions. Moreover, in order to e ciently deal with the relationships present within the data we can use Relational Concept Analysis.