In this thesis I examine the parameterization of two networks with discrete-time dynamics.
The parameterization problem is rephrased as supervised learning problem.
The similarities between the examined networks and recurrent neural networks are explotied by using optimization algorithms used for learning deep neural networks.
The first network has similar dynamics to real protein-protein interaction (PPI) networks, but solves
an easy toy problem.
By examining this network I identified potential problems hindering optimization by causing disappearing gradients or by creating highly non-linear error surfaces.
The larger examined network is a PPI network describing cancer cells.
This network was parameterized to predict the results of in vitro cell viability assays using cancer cell lines.
The parameterization achieved an acceptable result for predicting individual experiments.
However the prediction of EC50 values for drug--cell line pairs is lacking.
This result can be likely attributed to using a weak model, and too little input data.