Despite the growing accessibility of sequencing data, three-dimensional protein structure determination remains a major scientific challenge. Experimental methods are prohibitively expensive or impossible, and predictive methods often lack accuracy, resolution, and/or computational efficiency. Leveraging recent advancements in language translation methods and the Protein Data Bank, now with more than 140,000 protein structures, we will investigate the ability of machine learning methods to accurately and efficiently transform a sequence of amino acids into a complete (all-atom) protein structure. Neural machine translation methods can be utilized to translate the “language of amino acids” into the “language of angles” that define how each atom in a protein is placed in relation to the last.<br/><br/>Recent work has demonstrated that Recurrent Neural Networks (RNNs) are effective at protein structure prediction but take weeks or months to train. Meanwhile, in the field of language translation, attention-based models that forgo recurrence altogether are able to achieve similar performance to RNN based models while taking significantly less training time. Motivated by the performance of attention-based architectures, we present a novel network for complete protein structure prediction based on the Transformer network for machine translation. We then evaluate the performance of this network and compare it with other deep learning methods built for the same task.<br/><br/>This work utilizes Tensorflow and Pytorch, two GPU-accelerated machine learning libraries.

Exploring sequence-to-sequence learning methods for end-to-end, complete protein structure prediction

2019 Fall National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.