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

Jonas Pfab; Dong Si

When looking at the results reported in Table 1 , we can note a couple of things. First, we see that the percentage of false positives predicted by the DeepTracer has increased by 1.9% compared to the C-CNN method. This can be explained through the manual thresholding which reduces noises in the experimental density maps and is utilized in the C-CNN method. However, considering that this step introduces manual effort to the prediction process, this is an acceptable concession as our new method is fully automated. Next, we see that the coverage percentage of the Deep-Tracer beats Phenix by over 20%. However, it still lacks behind the C-CNN method by 3%. On the contrary, Phenix performs best for RMSD, beating the DeepTracer by 0.08 and the C-CNN method by 0.09. If we compare the DeepTracer to Phenix based on these two metrics, we can see that the coverage improvements significantly outweigh the small deterioration in RMSD. However, compared to the C-CNN method there is no significant difference regarding the RMSD and coverage. As mentioned in the introduction, we wanted to focus not only on the prediction accuracy, but also on the runtime of the predictions. And here we can discern immense differences. For the prediction of smaller structures. the C-CNN method took at least 20 minutes while the DeepTracer ran small predictions in a matter of seconds. However, the bigger implications lay in the prediction of larger proteins. For around 6000 atoms the C-CNN method took over 24 hours to finish, whereas the DeepTracer only needed around 20 minutes, for some even less than 5 minutes. Particularly, the trendline is interesting as it gives us an idea of how the methods will behave for very large proteins. The C-CNN trendline seems almost linear above 2000 predicted proteins and the runtime seems to quadruple when 6000 predicted atoms are reached. If this trend continues then we could expect a 75-fold runtime increase in order to predict 100,000 atoms. The runtimes of DeepTracer, however, seem to remain flat once a certain number of atoms is reached which would mean that very large proteins could be predicted within a reasonable runtime. This can be explained by the chain identification step that is applied before the prediction of any atom. While larger proteins consist of more chains than smaller ones, the number of atoms in the chains does not necessarily grow. As each chain is processed separately this means that if the chain lengths remain constant and only the number of chains gets larger the runtime grows linearly with the number of chains.