Todays high-throughput molecular profiling technologies allow to routinely create large datasets providing detailed information about a given biological sample, e.g. about the concentrations of thousands contained proteins. A standard task in the context of precision medicine is to identify a set of biomarkers (e.g. proteins) from these datasets that can be used for disease diagnosis, prognosis or to monitor treatment response. However, finding good biomarker sets is still a challenging task due to the high dimensionality and complexity of the data and the often quite high noise level.In this work, we present an approach to this problem based on Deep Neural Networks (DNN) and a transfer learning strategy using simulation data. To allow interpretation of the results, we compare different approaches to analyze the learned DNN. Based on these interpretation approaches, we describe how to extract biomarker sets.Comparison of our method to a state-of-the-art L1-SVM approach shows that the new approach is able to find better biomarker sets for classification when small sets are desired. Compared to a state-of-the-art $$\ell _1$$-support vector machine ($$\ell _1$$-SVM) approach, our method achieves better results for the classification task when a small number of features are needed.