Using Curriculum Learning in Pattern Recognition of 3-dimensional Cryo-electron Microscopy Density Maps.

Area of Science:

• Structural Biology
• Biophysics
• Computational Biology

Background:

• Cryo-electron microscopy (cryo-EM) enables atomic structure determination of proteins.
• Medium-resolution cryo-EM density maps (5-10 Å) present challenges for atomic structure resolution.
• Deep neural networks (DNNs) show promise for predicting protein secondary structures from cryo-EM data.

Purpose of the Study:

• To investigate the effectiveness of curriculum learning for training DNNs on cryo-EM density maps.
• To address challenges in DNN training accuracy due to variations in cryo-EM data quality.
• To improve the derivation of atomic structures from medium-resolution cryo-EM maps.

Main Methods:

• Utilized 1,382 3D cryo-EM images from the Electron Microscopy Data Bank.
• Implemented and investigated three distinct curriculum learning strategies for DNN training.
• Focused on strategies that manage the reuse of training data and mitigate the 'forgetting problem' during continuous training.

Main Results:

• Curriculum learning significantly improved the prediction accuracy of the trained DNNs.
• The performance enhancement was observed when the 'forgetting problem' was effectively addressed.
• The study demonstrated the utility of curriculum learning for handling diverse cryo-EM data quality.

Conclusions:

• Curriculum learning offers a more effective approach to training DNNs with variable-quality cryo-EM density maps.
• This method enhances the potential for accurate protein secondary structure prediction and atomic structure determination.
• Addressing data quality variations through curriculum learning is crucial for advancing cryo-EM structural biology.