Automated Threshold Selection for Cryo-EM Density Maps

Recent advances in cryo-EM have made it possible to create protein density maps with a near-atomic resolution. This has contributed to its wide popularity, resulting in a rapidly growing number of available cryo-EM density maps. In order to computationally process them, an electron density threshold level is required which defines a lower bound for density values. In the context of this paper the threshold level is required in a pre-processing step of the backbone structure prediction project which predicts the location of Cα atoms of the backbone of a protein based on its cryo-EM density map using deep learning techniques. A custom threshold level has to be selected for each prediction in order to reduce noise that could irritate the deep learning model. Automatizing this threshold selection process makes it easier to run predictions as well as it removes the dependency of the prediction accuracy to the ability of someone to choose the right threshold value. This paper presents a method to automatize the threshold selection for the previously mentioned project as well as for other problems which require a density threshold level. The method uses the surface area to volume ratio and the ratio of voxels that lie above the threshold level to non-zero voxels as metrics to derive characteristics about suitable threshold levels based on a training dataset. The threshold level selection was tested by integrating it in the backbone prediction project and evaluating the accuracy of predictions using automatically as well as manually selected thresholds. We found that there was no loss in accuracy using the automatically selected threshold levels indicating that they are equally good as manually selected ones. The source code related to this paper can be found at https://github.com/DrDongSi/Auto-Thresholding.