DeepTracer: Predicting Backbone Atomic Structure from High Resolution Cryo-EM Density Maps of Protein Complexes

Jonas Pfab; Dong Si

In this paper, we presented the DeepTracer, a fully automated tool which can accurately predict backbone protein structures based on 3D cryo-EM density maps. In terms of accuracy, we were able to outperform the established structure prediction method Phenix by over 20% in structure coverage with only a slight RMSD increase of 0.08. Compared to the deep learning-based C-CNN method we achieved a slightly better RMSD by a margin of 0.01 with a slightly decreased coverage of 3%. In contrast with the C-CNN method, this fully automated method requires no manually set parameters. And the vastly reduced runtime allows us to predict protein structures within minutes instead of days. This is significant, as it allows the researchers to apply the DeepTracer on very large density maps and expect results maximally within a couple of hours. Furthermore, we achieved preliminary results in the amino acid type prediction from 3D with an accuracy of 25.2%. This gives promise to the incorporation of amino acid sequence into the prediction pipeline. Since the predicted amino acids from 3D could potentially be utilized for the reference of future full-atom structure prediction. In the future, we expect to use the amino acid type information to extend the functionality of the DeepTracer such that it can predict the protein full atom structure, including the side chains of the amino acids.