Bioinformatics: Deep learning in protein structure predictions
Our ultimate goal is to model all proteins and their interaction partners in a cell at atomistic resolution. Given the recent revolution in structure prediction for both individual proteins and complexes this goal is not completely futuristic. The basis for this is the development of direct coupling analysis direct coupling analysis based contact predictions methods using direct coupling information and the recent advances in machine learning. In short predicted contacts can be used to predict the structure of individual protein as well as of protein-protein interactions. The major limitations of these methods has been that they can only be used to study very large protein families. To overcome this problem we have developed PconsC3 that performs significantly better for smaller protein families. We have also worked on our the development of our ProQ quality estimation programs together with the Wallner group.