Finding Metabolic Pathways in Decision Forests

Abstract

Data of metabolite concentrations in tissue samples is a powerful source of information about metabolic activity in organisms. Several methods have already been reported that allow for inferences about genetic regulatory networks using expression profiling data. Here we adapted these techniques for metabolic concentration data. While somewhat accurate in predicting certain properties these methods encounter problems when dealing with networks containing a high number of vertices (τ 50). To circumvent these difficulties, larger data sets are usually reduced a priori by means of preprocessing techniques. Our approach allows to make network inferences using ensembles of decision trees that can handle almost any amount of vertices, and thus to avoid time consuming preprocessing steps. The technique works on a bootstrap principle and heuristically searches for partially correlated relations between all objects. We applied this approach to synthetically generated data as well as on data taken from real experiments.