Hedelius, Bryce; Millecam, Todd; Wingate, David; Della Corte, Dennis (Brigham Young University)
Faculty Advisor: Della Corte, Dennis (BYU College of Physical and Mathematical Sciences, Physics); Wingate, David (BYU College of Physical and Mathematical Sciences, Computer Science)
Substantial progress has been made in the past several years towards the accurate prediction of protein tertiary structures from primary sequence, aided greatly by the integration of machine learning. Current success is based on two-stage protocols: first, the training of a deep convolutional neural network (CNN) to predict macromolecular structure restraints, and second, the use of these restraints to construct a folded three-dimensional structure of the target protein. Such a two-stage folding protocol was used by DeepMind in the recent Critical Assessment of Structure Prediction (CASP13), which outperformed all established groups. However, DeepMind has not expressed a plan to publish the code of their AlphaFold protocol. Here we present ProSPr, a network representing the first part of the AlphaFold pipeline for predicting interatomic distances, and demonstrate its abilities in the contact prediction task relative to other state-of-the-art methods. We also investigate and report on the roles of certain input features in prediction quality. ProSPr is made freely available to the scientific community both as source code and a Docker container, which we anticipate will encourage the development of better techniques for assembling protein structures from restraints.