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

Jonas Pfab; Dong Si

Throughout the last decade, technological advancements in cryogenic electron microscopy (cryo-EM) have provided a new approach to solve the structure determination problem. Cryo-EM allows researchers to capture 3D images of large protein complexes by cooling them to cryogenic temperatures without the need of a costly crystallization of the protein [13] . In recent years the resolution of these images has increased to a nearatomic level which led to an exponential gain in popularity of the technology [14] . A 3D image captured through cryo-EM is an electron density map (see Fig 1) . It represents the volume of the protein complex through a 3D grid of voxels. Each voxel stores the electron density value, meaning the probability that an electron is present, for its location. Latest developments in the field of cryo-EM have allowed researchers to capture many high-resolution density maps from large protein complexes in the course of a single study [15] . This makes the ability to predict protein structures from cryo-EM maps fully automatically even more significant as it allows for a very high throughput of protein structures derived from cryo-EM maps. The structure of a protein can be estimated from its density map, intuitively by placing the amino acids in such way that the resulting structure fits well into the density map. While this sounds like a simple task it becomes very challenging due to lower local resolutions as well as noise created through errors in the cryo-EM. Currently, there are several existing methods that perform protein structure predictions based on cryo-EM data such as Phenix and Rosetta [16] , [17] . They both use conventional algorithms to place and connect the amino acids of the protein. However, they often struggle to identify amino acids as such and, therefore, tend to have a low coverage on large protein complex [18] . Furthermore, their long runtime is a problem. The prediction of small density maps can take up to several days rendering the method practically useless for the prediction of large proteins. The C-CNN method, another protein structure prediction application which utilizes deep learning techniques, achieves better results in terms of coverage [18] . However, it requires some manual input and the runtime could be further improved. Additionally, it predicts the location of Cα atoms without information about their amino acid type. Therefore, we present the DeepTracer, a new fully automated method for the prediction of protein structures from cryo-EM density maps, with the goal of improving the runtime by applying a 3D U-Net in combination with a modified travelling salesman algorithm and having the ability of recognizing the amino acid type from 3D cryo-EM density maps.