We are developing a general methodology for model-based identification and analysis of complex nonlinear interaction networks from incomplete and noisy observations. We are employing rigorous theoretical and computational techniques for estimating the structural and dynamic properties of complex interaction networks by state-of-the-art identification and model selection methodologies, and for studying network robustness via probabilistic sensitivity analysis. We use fundamental laws of physics (such as thermodynamics) to appropriately constrain the systems under consideration and focus our identification and analysis methodologies only on physically realizable networks. Sensitivity analysis is used from start to finish: in model reduction to ease inference, in experimental design to enhance scientific discovery, and in final system analysis for controlled intervention.