A Tool for Segmentation of Secondary Structures in 3D Cryo-EM Density Map Components Using Deep Convolutional Neural...

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Cryo-electron microscopy (cryo-EM) enables atomic structure determination, but medium-resolution maps (5-10 Å) present challenges.
• Identifying protein secondary structures (helices, β-sheets) in these maps provides crucial constraints for atomic modeling.
• Advancements in deep learning offer new avenues for analyzing complex biological data.

Purpose of the Study:

• To develop an effective software bundle, DeepSSETracer, for detecting protein secondary structures from medium-resolution cryo-EM maps.
• To integrate deep learning models with visualization software for user accessibility.
• To provide a robust tool for enhancing atomic structure resolution from cryo-EM data.

Main Methods:

• Developed DeepSSETracer, a software bundle featuring a U-Net deep neural network architecture.
• Trained the model using curriculum learning and gradient of episodic memory (GEM).
• Integrated the deep learning model with ChimeraX for visualization and analysis.

Main Results:

• DeepSSETracer achieved residue-level F1 scores of 0.72 for helices and 0.65 for β-sheets on 28 cryo-EM map components.
• The software bundle processes a 446-amino acid map in approximately 6 seconds on a single CPU and GPU.
• DeepSSETracer demonstrated superior β-sheet detection compared to Emap2sec+ (F1 score 0.65 vs. 0.42).

Conclusions:

• DeepSSETracer is an effective tool for secondary structure detection in medium-resolution cryo-EM maps.
• Integrating deep learning with ChimeraX provides a promising approach for structural biology applications.
• The developed software aids in improving atomic model accuracy from challenging cryo-EM density maps.