Section: New Results

KDDK in Life Sciences

Participants : Yasmine Assess, Emmanuel Bresso, Thomas Bourquard, Adrien Coulet, Marie-Dominique Devignes, Anisah Ghoorah, Renaud Grisoni, Jean-François Kneib, Florence Le Ber, Bernard Maigret, Jean-François Mari, Amedeo Napoli, Violeta Pérez-Nueno, Dave Ritchie, Malika Smaïl-Tabbone.

The Life Sciences constitute a challenging domain in which to implement knowledge-guided approaches for knowledge discovery. Biological data are complex from many points of views: voluminous, high-dimensional, deeply inter-connected, etc. Analyzing such data and extracting hidden knowledge has become a crucial issue in important domains such as health, environment and agronomy. More and more bio-ontologies are available and can be used to enhance the knowledge discovery process [88] , [117] . In the next few years, the experience of the Orpailleur team in KDDK applied to the Life Sciences will be further developed in two directions: the use of bio-ontologies to improve approaches for data integration and mining when applied to real-world data, and the study of the synergy between numeric and symbolic data-mining methods in life-science applications.

Relational data mining applied to complex biological object characterization and prediction

Inductive Logic Programming (ILP) is a learning method which allows expressive representation of the data and produces explicit first-order logic rules. However, any ILP system returns a single theory based on heuristic user-choices of various parameters and learning biases, thus ignoring potentially relevant rules. Accordingly, we propose an approach based on Formal Concept Analysis for effective interpretation of reached theories with the possibility of adding domain knowledge. Our approach was applied to the characterization of three-dimensional (3D) protein-binding sites, namely phosphorylation sites, which are the protein portions on which interactions with other proteins take place [33] . In this context, we defined a logical representation of 3D patches and formalized the problem as a concept learning problem using ILP. Another application of this KDDK methodology concerns the characterization and prediction of drug side-effect profiles (Journal manuscript in preparation). In this case, maximal frequent itemsets are extracted and allow us to propose relevant side-effect profiles of drugs which are further characterized by ILP.

Functional classification of genes using semantic similarity matrix and various clustering approaches

In the last report, we proposed a measure called IntelliGO which computes semantic similarity between genes for discovering biological functions shared by a set of genes (e.g., showing the same expression profile). This measure takes into account domain knowledge represented in Gene Ontology (GO) [83] .

Functional classification aims at grouping genes according to their molecular function or the biological process they participate in. Evaluating the validity of such unsupervised gene classification remains a challenge given the variety of distance measures and classification algorithms that can be used. We evaluated functional classification of genes with the help of reference sets. Overlaps between clusters and reference sets are estimated by the F-score metric. We test the IntelliGO measure with hierarchical and fuzzy C-means clustering algorithms and we compare results with the state-of-the-art DAVID functional classification method (Database for Annotation Visualization and Integrated Discovery). Finally, study of best matching clusters to reference sets leads us to propose a method based on set-differences for discovering missing information.

The IntelliGO-based functional clustering method was tested on four benchmarking datasets consisting of biological pathways (KEGG database) and functional domains (Pfam database) [13] . The IntelliGO measure is usable on line (see http://bioinfo.loria.fr/Members/benabdsi/intelligo_project/ ).

We are currently investigating the clustering problem when objects are not represented as feature vectors in a vector space but as a pairwise similarity matrix. In biology such similarity measures are often computationally expensive or incompatible with bona fide distance definition. Embedding techniques of pairwise data into Euclidean space aim at facilitating subsequent clustering of the objects  [115] . Spectral clustering methods are also relevant in this case  [127] . We are conducting comparative and large-scale gene clustering evaluation using the Intelligo measure and reference sets.

Analysis of biomedical data annotated with ontologies

Annotating data with concepts of an ontology is a common practice in the biomedical domain. Resulting annotations define links between data and ontologies that are key for data exchange, data integration and data analysis tasks. In 2011 we collaborated with the National Center for Biomedical Ontologies (NCBO) to develop of large repository of annotations named the NCBO Resource Index  [99] . The resulting repository contains annotations of 34 biomedical databases annotated with concepts of 280 ontologies of the BioPortal (http://bioportal.bioontology.org/ ). We proposed a comparison of the annotations of a database of biomedical publications (Medline) with two databases of scientific funding (Crisp and ResearchCrossroads) to profile disease research [18] . The annotation of these three databases with a unique ontology about diseases enable to consider their content conjointly and consequently to analyze and compare, for distinct disease (or family of diseases), trends in term of number of publications and funding amounts.

We started a new project that aims at exploring biomedical annotations with FCA techniques. One main challenge here is to develop a knowledge discovery approach that consider the knowledge represented in the ontologies employed for the annotations.

Connecting textual biomedical knowledge with the Semantic Web

A large amount of biomedical knowledge is in the form of text embedded in published articles, clinical files or biomedical public databases. It is consequently of high interest to extract and structure this knowledge to facilitate its consideration when processing biomedical data. We benefited from advances in Natural Language Processing (NLP) techniques to extract fine-grained relationships mentioned in biomedical text and subsequently published such relationships on line in the form of RDF triples  [91] , [90] . In a collaborative work with the Health Care and Life Science (HCLS) interest group of the W3C, we demonstrated how biomedical knowledge extracted from text, along with Semantic Web technologies has high potential for recommendation systems and knowledge discovery in biomedicine  [118] .