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

Jonas Pfab; Dong Si

Before we place any Cα atoms, we try to identify separate chains in the protein structure. This will help us later when connecting the atoms and improves the runtime of the connection process. In order to discern the chains, we identify connected areas of voxels that have a confidence value of 0.5 or higher. An example of the chains identified from a backbone prediction can be seen in Fig 5. We can now mask all other confidence maps with each chain area that we identified in order to continue the prediction independently for all chains. The copyright holder for this preprint (which was not peer-reviewed) is the . https://doi.org/10.1101/2020.02.12.946772 doi: bioRxiv preprint Next, we place the actual Cα atoms in two simple steps, using the masked Cα confidence maps. First, we find the initial atom locations by calculating local minima in the confidence map within a neighborhood of one voxel and a minimum value of 0.6. The value of 0.6 was chosen after some testing as it produced the most accurate results. This gives us a list of voxel indices. However, as atoms can have floating point coordinates, we try to improve this initial location by calculating the center of mass for 3 # cubes around each local minimum. The resulting values are used as the coordinates for the Cα atoms of the protein structure.