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

Jonas Pfab; Dong Si

In order to get a better idea of the implications of the results we take a closer look at the protein structures predicted by the DeepTracer and Phenix for the EMD-6272 density map (rotavirus VP6) and compare them to the solved structure (see Fig 9) . We can see in the highlighted areas that the DeepTracer incorrectly connected amino acids in two cases. However, all predicted amino acids are still connected as a single chain. In comparison to that the prediction of Phenix is fragmented into multiple chains. Furthermore, amino acids appear to be missing from the prediction explaining the coverage difference of 8.1% (see Table 1 ). The predicted secondary structure elements in the Phenix prediction, particularly for sheets, appear also much less accurate than the ones made by the DeepTracer. Another noteworthy observation, although not a fair comparison as Phenix predicts more than just the backbone structure, is that the prediction runtime of Phenix for the EMD-6272 map was 1 hour and 45 minutes while the runtime of the DeepTracer for the same map was 40 seconds. As mentioned in the introduction, a central goal of the DeepTracer was to improve the runtime of protein structure predictions. This makes it possible to predict ever larger density maps which allows for a more versatile application of the method. In order to compare the runtime of the Deep-Tracer with the C-CNN method we measure the execution time of each method for the prediction of multiple density maps. A scatter plot of the measured times is depicted in Fig 10. Fig 10. Scatter plot of prediction runtimes (y-axis) of the DeepTracer (blue) and the C-CNN method (red) for various density maps against the number of predicted atoms (x-axis). In order to see the runtimes more clearly, we use a logarithmic scale with a base of 4.