Mathematical models of complex biological networks are valuable to make predictions on system properties and to identify therapeutic targets. However, development, validation and analysis of predictive models is often hampered as absolute and quantitative measurement data are rarely available. Instead, the data are typically uncertain with respect to the measured variable or time, relational due to normalization, or data are given as conditional if-then observations. Many common approaches for model development and validation cannot deal with such semi-quantitative data and qualitative information. We present a framework for the guaranteed invalidation and parameter estimation of dynamical models using such data. For this purpose, the data are formally expressed by sets of equalities and inequalities containing binary variables. Then a mixed-integer nonlinear feasibility problem is constructed, which is subsequently relaxed into a mixedinteger linear program that can be solved efficiently. A model can be proved inconsistent, that is, invalid with the uncertain and semi-quantitative/qualitative data, if the solution set of the mixed-integer linear program is empty. To exemplify the approach, we analyze different models whether they can show adaptation to a step-input. First, we invalidate all but one model and, second, derive outer-bounds for those regions in the parameter space of the non-invalidated model that contain parameterizations for which it is consistent with the data.