(6a) Transformer Networks for Fast and Improved Protein Structure Prediction

We put forward a novel neural-network based protein structure prediction pipeline which maps the protein sequence to structure problem as a language translation problem (i.e. used by Google translator). It uses a transformer network at the heart of its core learning module. Transformer networks have shown to be capable of 'paying attention' to the entire amino acid sequence (rather than fragments) while discerning biophysical and biochemical rules that decide its structure. We have trained our network over every single protein structure deposited in public databases. We have used it for successfully recapitulating structures of 'in-house' de novo proteins, which were not a part of the training set, with sub 4-angstrom RMDSs. Our future plan is to use a top-down implementation of this tool to identify all possible sequences that will not hamper the structural integrity of a given protein. This method can then be used to re-design enzymes with enhanced cofactor/ substrate specificity, designing high-affinity high-avidity nanobodies to target disease causing antigen proteins, and also for redesigning porins for use in drug-delivery devices.