Abstract by Wendy Billings
Physics and Astronomy
Dennis Della Corte
ProSPr: Protein Structure Prediction using Deep Learning
Deep convolutional neural networks have recently enabled spectacular progress in predicting protein structures, as demonstrated by DeepMind’s winning AlphaFold entry in the latest Critical Assessment of Structure Prediction experiment (CASP13). The best protein prediction pipeline leverages intermolecular distance predictions to assemble a final protein model, but the resources required to create these predictions for arbitrary sequences are not publicly available. Here we present ProSPr, a trained implementation of a comparable network, and make it freely available to the scientific community. We benchmark ProSPr predictions in the related task of contact prediction against the CASP13 contact prediction winner TripletRes. Further, we investigate the dependence of prediction quality on information derived from multiple sequence alignments and perform ablation tests. Access to ProSPr will enable other labs to build on best-in-class protein distance predictions and engineer superior protein reconstruction methods.