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

Jonas Pfab; Dong Si

The preprocessing steps can be split into preparing the cryo-EM density maps and creating the different labels based on the solved structures. For the experimental density maps we have to make sure that the voxel size equals one, meaning that every voxel measures 1Å. Furthermore, we need to adjust the dimensions of the map to the 64 # input dimensions of the model as seen in Fig 2. This is done by splitting the maps into several 64 # big cubes. In order to avoid errors at the borders of the cubes we overlay them partly such that we only use the 50 # center voxels of each cube to reconstruct the output maps to the original dimensions of the density map. By splitting the 411 density maps that we collected in this manner, we retrieve 16,070 training examples. To create the labels that the model can be trained on, we use the solved structure to create masks by filtering out the corresponding atoms and setting all voxels within a certain distance to those atoms to one. For the Cα mask we filter only the Cα atoms of the protein structure and set all directly neighboring voxels around them to one. The backbone mask is created similarly except that we use all Cα, C, and N atoms. To generate the secondary structure masks, we filter out all atoms for each structural element and set all voxels within a distance of two around each filtered atom to one. Finally, to create the amino acid type masks we filter out all atoms for each amino acid type and set all neighboring voxels to one for each atom. In total this gives us 25 different masks for each training example (Cα mask, backbone mask, three secondary structure masks, and 20 amino acid type masks).