Simulations of chemical reactions in biological systems require an extensive sampling of the phase space, which demands a compromise between accuracy and computational cost. While semi-empirical methods resort to parametrization, neural networks (NNs) follow a data-driven approach to close the accuracy gap to higher order quantum mechanical methods. Transferable models, such as high-dimensional neural network potentials (HDNNPs) developed by Behler and Parinello or graph neural networks (GNNs), might lead to a faster and more flexible solution for the description of systems, where semi-empirical methods are insufficient and higher order methods too demanding. My research is therefore aimed at the exploration of ways to apply these models in order to move from QM/MM- to NN/MM-calculations. Both the disulfide shuffling in P4 and the phosphoryl transfer reaction in P5 represent ideal candidates for the training, evaluation and application of the developed models. The cooperation between enzyme research, which provides the required training data, and model development benefits both parties.