A variety of approaches aimed at machine learning probabilistic relational representations have been explored in recent years. These include learning algorithms for Probabilistic Relational Models, Bayesian Logic Programs and Stochastic Logic Programs. In this paper we investigate the question of whether there exist important applications which require the machine learning of such representations. In particular, applications of this kind must necessitate a) machine learning, b) probabilistic representation and c) relations. We argue in this paper that the automated extension of partial metabolic pathway descriptions fits all these criteria. As usual with such representations the learning task can be divided into sub-tasks involving model building and parameter estimation. In order to gain understanding of the sample size requirements for estimating the parameters of medium-sized metabolic networks we conducted experiments in which RMS error is estimated with increasing sample size. We separately investigated the effects of multiple pathway contributions to the concentrations of metabolites within the networks. For non-branching pathways the results indicate that parameter estimation is feasible from limited sources of data. However, the introduction of branches increases sample requirements considerably.