We develop fast and powerful large-scale data integration methods which enable the reconstruction of context-specific molecular networks, e.g. for a given disease, patient group or individual patient. The fastcore family of algorithms allows the integration of large-scale gene expression data and other data types with generic metabolic reconstructions for producing specific molecular metabolic networks. This is powered by an efficient linear programming approach, which allows to obtain a close-to-optimal minimal network given a core set of metabolic reactions. Largely applied also by other teams, these algorithms have been included in a community-effort toolbox. This has lately been extended with a novel dynamic Flux Balance Analysis approach for multi-tissue metabolic modelling and allows now also for simulating disease-specific metabolic blood level alterations.

Based on these and other state-of-the-art computational biology, data science and machine learning approaches, we developed a variety of fruitful collaborations, notably in cancer research. E.g. we employ the reconstructed cancer specific molecular networks for identifying promising specific targets and to suggest novel treatment strategies. Drug repositioning thereby aims at reorienting approved drugs to novel disease indications. In proof-of-concept studies we used fastcore to predict a number of non-cancer drugs to be effective in colorectal cancer or melanoma, while not harming healthy control tissue. The experimental validation gave superior results compared to large-scale screening efforts.

Many techniques for analyzing our biological data generate latent spaces, new representations of data that preserve the important information.  This talk will start by examining a variety of methods for generating such latent spaces, their similarities, differences and individual strengths.  However, in order to be truly useful these latent spaces also need to be interpretable, to enable us to gain new insights into the biological systems being perturbed in our experiments.  We will investigate the challenges surrounding interpretability and some approaches which may hold promise for delivering the much needed powerful understanding of our data sets and ultimately our systems.