An Approach to Developing Benchmark Datasets for Protein Secondary Structure Segmentation from Cryo-EM Density Maps.

Area of Science:

• Structural biology
• Computational biology
• Biophysics

Background:

• Deep learning methods are increasingly used for segmenting protein secondary structures from cryo-electron density (cryo-EM) maps.
• Current approaches are often tested on limited experimental datasets, hindering a full understanding of factors affecting performance.

Purpose of the Study:

• To develop a method for generating synthetic cryo-EM datasets with controlled variations in protein sequence identity, structural content, and data quality.
• To investigate the impact of secondary structure content and data quality on the performance of a deep learning segmentation tool, DeepSSETracer.

Main Methods:

• A novel approach was implemented to generate synthetic datasets with adjustable parameters for sequence identity, secondary structure content, and data quality.
• Generated datasets were used to train and test DeepSSETracer, a deep learning model for segmenting protein secondary structures from cryo-EM maps.

Main Results:

• The data generation approach successfully created test and training sets with specified characteristics.
• Results demonstrated that varying secondary structure content and data quality significantly influences the segmentation performance of DeepSSETracer.

Conclusions:

• Synthetic data generation is a viable strategy to systematically study factors affecting deep learning model performance in cryo-EM map analysis.
• Understanding the influence of data characteristics is crucial for optimizing deep learning-based secondary structure segmentation in cryo-EM.