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

Jonas Pfab; Dong Si

In addition to the false-positive percentage, the coverage, and the RMSD we also calculated the secondary structure and amino acid type accuracy for the DeepTracer using the confidence maps predicted by U-Net. We can see that the secondary structure prediction performs well with an average accuracy of 81.72%. We also show that deep learning could discern different types of amino acids based on the density maps. Particularly, as this is only the preliminary stage of predicting amino acid type directly from 3D cryo-EM density map. It gives promise for future improvements through a more elaborate training of the U-Net with more data. Furthermore, we plan to include the amino acid sequence as an input parameter, so that we can map the predicted sequence with the input sequence and refine our prediction. For this sequence mapping work, an amino acid type prediction accuracy of around 30% from 3D could be already helpful. Particularly, as we can see from Fig 12 that the U-Net tends to correctly predict the type of consecutive amino acids, likely due to high local resolutions. As it is very unlikely that this happens by chance, we can use this information to map the input sequence to our 3D backbone atom and amino acids prediction.