PASCAL - Pattern Analysis, Statistical Modelling and Computational Learning

Statistical relational learning for supervised gene regulatory network inference
Céline Brouard, Julie Dubois, Christel Vrain, Marie-Anne Debily and Florence d'Alché-Buc
In: Machine Learning in Systems Biology 2009, 5-6 Sept 2009, Ljubljana.


Starting from a known gene regulatory network involved in the switch proliferation/differenciation in keratinocytes cells, we have developed a new approach to learn rules that can explain the presence or absence of regulation between two genes. For this purpose, we have used experimental data (gene expression) as well as knowledge such as GO annotations and positions of genes on chromosomes. In the context of statistical relational learning, we have learned the concept of transcriptional regulation between two genes, represented by a predicate "regulate". A network of genes extracted from Ingenuity has been used for labelling couples of genes, and experimental data as well as prior knowledge have been encoded into a first order representation (ground atoms and rules) \cite{Dub07}. We have first applied a pure inductive logic programming approach, Aleph\cite{Sri04} and we have compared it to a statistical relational learning approach \cite{Get07}, called Markov Logic Network, introduced by Domingos et al.\cite{Low07,Dom08} In this framework, a set of weighted logical rules is represented by a random Markov network: nodes correspond to ground atoms, rules allow to form cliques, and the weights of the rules are associated to the corresponding cliques. Making a decision corresponds to computing the posterior probability of the labels given the input description. We have used Aleph to produce a large set of rules, thus fixing the structure of the random Markov network, and we have applied a discriminative learning method to get the weights associated to the rules implemented by Alchemy, a source code implemented and described in (Kok et al. 2005). Among the rules ranked by Alchemy, we have found interesting regulatory patterns which show that first, Ingenuity can be cross-validated by experimental data and provide consistent information and second, new rules can be used to suggest new candidates for regulators and regulees. However, in terms of performance, Alchemy only marginally improves performance upon Aleph. Results are discussed and compared to those obtained by (Huynh and Mooney 2008) on other relational tasks.

EPrint Type:Conference or Workshop Item (Poster)
Project Keyword:Project Keyword UNSPECIFIED
Subjects:Learning/Statistics & Optimisation
ID Code:6719
Deposited By:Florence d'Alché-Buc
Deposited On:08 March 2010