Systems biology iteratively combines experimentation with mathematical modeling, and the complexity of cellular networks constitutes an important challenge for this approach. In addition, limited mechanistic knowledge, conflicting hypotheses, and relatively scarce experimental data hamper the development of mathematical models as systems analysis tools. From a structural network analysis point of view, defining and identifying suitable sub-structures, or motifs, in complex interaction networks significantly helps understanding biological functions. However, current approaches only allow for function prediction at the motif level: by recognizing a known motif in a given network, one can assign its previously established function. We introduce the concept of reaction motifs and develop probabilistic, predictive models for metabolic network functions. Importantly, the framework leverages information ‘hidden’ in the correlation structure of motifs in metabolic networks to automate, for instance, gap filling and genome annotation. At a more detailed level, methods to systematically develop and discriminate between predictive dynamic models are still lacking. To address this problem, we developed a computational method that incorporates all hypothetical mechanisms about the architecture of a biological system into a single model, and automatically generates a set of simpler models compatible with observational data. For the short-term dynamic control of the transcription factor Msn2 in yeast, iterations between model predictions and rationally designed phosphoproteomics and imaging experiments identified a single, highly plausible circuit topology. Overall, novel mathematical and computational methods may allow for the systematic construction of systems biology models despite the prevailing uncertainty in network components and functions.