A Causal Model for Disease Pathway Discovery

Pathway provides a deep insight into the mechanism of the biological process. With the increasing of high-throughout gene expression monitoring technology, a lot of data driven methods have been proposed to reconstruct the pathways from the observation data. Low reliability of the discovered results, especially the direction of the regulatory relation, is the main challenge of the existing methods. In this work, a level-wise causal search (LWCS) based disease pathway discovery method is proposed. The following three steps are conducted in each searching level of LWCS to locate the causal variables: firstly, in the parents and children (PC) discovery step, structure learning approach is employed to discover the candidate causal genes; then, in the casual direction learning step, additive noise models are explored to determine the direction of the edges, finally, the trivial causal candidates are pruned and not contained in the further level search. The proposed method is tested and verified on real life gene expression data sets. The success of the proposed method reflects that the causality is a proper model to present the regulatory relations among the genes and phenotypes.