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

Jonas Pfab; Dong Si

The U-Net model gets its name from the U-shape of its architecture which is optimized for segmentation tasks on medical images [19] . The detailed architecture of the model can be seen in Fig 2. The cryo-EM density maps are fed to the 64 # input layer. The output layer has the same 64 # shape with N different channels. The number of channels depends hereby on what the U-Net is predicting (e.g. the secondary structure U-Net has three channels for loops, sheets, and helices). The copyright holder for this preprint (which was not peer-reviewed) is the .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 As mentioned above, the U-Net should predict multiple things, including the amino acid positions, secondary structure elements, as well as the amino acid types. Therefore, the model consists of three parallel U-Nets as shown in Fig 3. The first one is responsible for predicting the amino acid positions. As the Ca atom of each amino acid is central to its position we can reduce this task to only predicting the Ca atoms of the protein structure. Additionally, this U-Net also predicts the backbone location of the protein, which is defined as the position of all Ca, C, and N atoms of the structure. This prediction will later on be useful in order to connect the amino acids into chains. Thus, the output of the Cα atoms U-Net has two channels. The second U-Net is responsible for predicting the secondary structure elements and, therefore, its output has three channels, one for each structural element (α-helix, loop, and sheet). Finally, the third U-Net is tasked to predict the amino acid types of the protein structure. Since there are 20 different types of amino acids occurring in nature its output has the same number of channels. In Now that we have looked at the architecture of the model, we can continue with the data collection. In order for the deep learning model to learn common noises and errors present in cryo-EM density maps we decided to train the U-Net using experimental data rather than density maps simulated from solved structures such as [18] . We downloaded the density maps from the EMDataResource website [20] in combination with their solved protein structures which function as the ground truth for the training process. We obtained the solved structures from the RCSB Protein Data Bank [21] . Since this paper is focused on high resolution maps, we only used density maps with a resolution of 4Å or better. In total we used 411 different density maps to build the training and validation sets.