Fig. 1: Overview of the EM2NA framework.
From: Automated detection and de novo structure modeling of nucleic acids from cryo-EM maps
a The workflow of map segmentation for protein and nucleic acid (NA). The input for EM2NA is a cryo-EM density map and optional DNA/RNA sequences. The input map is first passed to stage-1 network to segment the nucleic acid region. b The workflow for DNA/RNA model building. The segmented nucleic acid map region from the stage-1 network is passed to the stage-2 network for predicting the backbone atom probabilities and nucleotide types at each voxel. The output atom probability map is converted to the backbone points by shifting integer grids to local maximum coordinates. The points are then traced to multiple paths by solving a Vehicle Routing Problem. With input sequences, a path-sequence alignment is utilized to assign nucleotide types to the backbone traces, where the base pairing feature is also considered for double helical parts. Finally, the full-atom DNA/RNA structure is constructed from the backbone traces. A post-refinement is recommended to refine the model, using e.g. phenix.real_space_refine. c The Swin-Conv (SC) UNet architecture for the stage-1 network in a and the stage-2 network in b. The input map is cut into overlapped boxes of 48 × 48 × 48 Å 3, which are fed to the network. The predicted boxes are assembled to the final map as output.
