Identifying functions shared by genes responsible for cancer is a challenging task. This paper describes the preparation work for applying Formal Concept Analysis (FCA) to complex biological data. We present here a preliminary experiment using these data on a core context with the addition of domain knowledge. The resulting concept lattices are explored and some interesting concepts are discussed. Our study shows how FCA can help the domain experts 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. Some earlier studies aimed at retrieving sets of genes sharing
the same transcriptional behavior with the help of Formal Concept Analysis [
This paper focuses on the preparation of biological data to data mining guided by domain knowledge. The objective is to apply knowledge discovery techniques for analyzing a list of differentially expressed genes and identifying functions or pathways shared by these genes assumed to be responsible for cancer. Section 2 explains the proposed approach for FCA-based analysis of biological data. Section 3 focuses on the conducted experiment. Section 4 discusses the results. Section 5 concludes the paper. c 2012 by the paper authors. CLA 2012, pp. 339–344. Copying permitted only for private and academic purposes. Volume published and copyrighted by its editors. Local Proceedings in ISBN 978–84–695–5252–0, Universidad de M´alaga (Dept. Matem´atica Aplicada), Spain.
We rely on the standard definition of FCA fully described in [
The first set of attributes M1 refers to four types of attributes, “Location”, “Pathway”, “Transcription Factors” and “GO Terms” (see Table 1). For our convenience we have named MSigDB categories as types of attributes and used only C1, C2, C3 and C5. The category C4 was not used as it keeps information on sets of genes related to a certain kind of cancer, which is not useful for the current problem. Thus we have a first context K1 = (G, M1, I1) where I1 denotes the relation stating that gene gi has an attribute mj in M1.
Types of Attributes C1: Positional Gene Sets
Description Data Provenance Location of the gene on the chro- Broad Institute mosome.
Pathway C3: Motif Gene Sets Transcription Factors C4: Computational Gene Cancer Modules Sets C5: Gene Ontology (GO) Biological Process, Cellular Com- AmiGO Gene Sets ponents, Molecular Functions
The second set of attributes M2 is related to the so-called “categories” where a category makes reference to a set of attributes with the “Pathway” type. For example, “Cell Growth and Death” is an example of category (see Figure 1). The categories in M2 determine a second context, K2 = (G, M2, I2) where M2 is the set of categories and I2 denotes the relation between a gene and a category. It can be noticed that the categories are only related to the “Pathway” type and that they can be considered as domain knowledge.
Moreover, the third set of attributes, namely M3, refers to the so-called
“upper categories”, which are defined as groupings of categories. Actually, we
have for the type “Pathway” a hierarchy of categories with two levels, categories
and upper categories (see Figure 1). The upper categories in M3 define a third
context K3 = (G, M3, I3) where M3 is the set of upper categories and I3 denotes
the relation between gene gi and an upper level category mj . Upper categories
are also related to the “Pathway” type and as categories, they can be considered
as domain knowledge too.
The framework described above was applied on three published sets of genes
corresponding to Cancer Modules defined in [
The exploration of a given concept lattice is carried out following the “Iceberg metaphor”, i.e., the lattice is explored level by level according to the support of
g din ) n m i r
B e P T T O A (G s r o t p e c e
R in y) Sero(tPonathwa
g alin n ig )
S y R wa
A th PP Pa ( ) r o t c 2 Fa 0 n
F2 tio U3 crip O s
P an V$ Tr (
ponenTterm)
Com(GO lar bly llu em e s C s
A 2 on) 1 i chr5(qLocat PBGSTenPBeH0s3L × × × × ×× × × CQMCNYTGC6PAT1 ×× ×× × × × Table 2. A toy example of formal context including domain knowledge. e rin c o d n m E te c: s k y
S l a ism n a rg s uc:Ostem k y
S
each concept, where the the support of a concept is the cardinality of the extent.
In addition, we also used stability for extracting interesting frequent and stable
concepts [
Module 5 419 3496 5,004 12 In this study, biologists are interested in links between the input genes in terms of pathways in which they participate, relationships between genes and their positions etc. We obtained concepts with shared transcription factors, pathways, locations of genes and GO terms. After the selection of concepts with a high support (≥ 10), we observed that there were some concepts with pathways either related to cell proliferation or apoptosis (expert interpretation). The addition of domain knowledge gives an opportunity to obtain the pathway categories shared by larger sets of genes (as categories and upper categories are there for maximizing the grouping of objects, see below).
Table 4 shows the top-ranked concepts found in each module. For example, in Table 4, we have the concept C4938:(KEGG Cytokine Cytokine Receptor Interaction, kc:Signaling Molecules and Interaction, kuc:Environmental Information Processing) and the concept C4995:(kc:Signaling Molecules and Interaction, kuc:Environmental Information Processing). These two concepts are such as C4938 ≤ C4995, meaning that C4995 has greater support than C4938. Moreover, we observed that the introduction of categories and upper categories in the global context allows us to consider concepts that otherwise would not be frequent. Actually, the role of categories and upper level categories is to facilitate the observation of sets of related genes.
This is a general way of obtaining larger sets of objects to interpret. When available, one can introduce a hierarchy of attributes –this is domain knowledge– and then insert the levels of each attribute in this hierarchy as a new attribute in the context. As a result, some classes of objects, that could not emerge before, will appear based on these hierarchical indications. Given the test data sets, the preliminary results obtained here constitute an interesting and positive control, and confirm that FCA-based analysis offers an efficient and practical procedure to explore complex and large sets of genes. 5
The preliminary study presented here shows how FCA can be applied to complex biological data and can give flexibility in using various types of attributes for analyzing a list of genes. In addition, domain knowledge can be introduced and guide the analysis. Dataset Concept Intents
ID Module 1 9585 9571 9566 9402 9078
As for future work, we plan to take into account relationships between genes and between terms (Gene Ontology relationships) and use the framework of relational concept analysis.