SegmA: Residue Segmentation of cryo-EM density maps

Abstract

1The cryo-EM resolution revolution enables the development of algorithms for direct de-novo modelling of protein structures from given cryo-EM density maps. Deep Learning tools have been applied to locate structure patterns, such as rotamers, secondary structures and C atoms. We present a deep neural network (nicknamed SegmA) for the residue type segmentation of a cryo-EM density map. The network labels voxels in a cryo-EM map by the residue type (amino acid type or nucleic acid) of the sampled macromolecular structure. It also provides a visual representation of the density map by coloring the different types of voxels by their assigned colors. SegmAs algorithm is a cascade of CNNs and group rotational equivariant CNNs. A data gathering algorithm was designed for creating datasets that will give best results when used for SegmAs training. At resolution of 3.2{degrees}A SegmAs accuracy is 80% for nucleotides. Amino acids which can be seen by eye, such as LEU, ARG and PHE, are detected by SegmA with about 70% accuracy. In addition SegmA detects regions where the exact labeling is of low confidence due to resolution, noise, etc. Removing those "unconfident" regions increases the amino acid detection accuracy to 80% The SegmA open code is available at https://github.com/Mark-Rozanov/SegmA_3A/tree/master.