For any model based applications, the quality of the metabolic network model determines the outcome of the application. Therefore, it is critically important to determine how accurate a genome scale model is in describing the cellular metabolism of a given strain. Due to the complexity involved in a genome-scale model, a good agreement between measured and computed substrate pickup rates and product secretion rates does not necessarily mean the model is accurate. To address this challenge, in this work we present a novel system identification based framework to extract the qualitative, biological knowledge embedded in the quantitative simulation results from the metabolic network models. This framework bridges the gap between the complicated numerical results generated from different genome-scale models and the knowledge that can be easily understood by biologists. The effectiveness of the proposed framework is demonstrated by analyzing two published genome-scale models of Scheffersomyces stipitis, pin-pointing their limitations, as well as guiding the development of a modified genome scale model.