Protein secondary structure prediction with context convolutional neural network.

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Protein secondary structure (SS) prediction is crucial for understanding protein structure and function.
• Deep learning, including convolutional and recurrent neural networks, has significantly advanced SS prediction accuracy.
• Exploring novel architectures and training procedures is key to further improving SS prediction.

Purpose of the Study:

• To develop and evaluate a novel Contextual Convolutional Neural Network (Contextnet) for protein secondary structure prediction.
• To compare the performance of Contextnet against established models under similar training conditions.
• To investigate the impact of training procedures on Contextnet's accuracy.

Main Methods:

• Construction of a novel Contextual Convolutional Neural Network (Contextnet).
• Comparative performance analysis of Contextnet against popular models (CNN, RNN, CNF) on the Jpred dataset.
• Retraining and validation of Contextnet on the Cullpdb dataset, with comparisons to Jpred, ReportX, Spider3, and MUFold-SS on the CASP13 dataset.

Main Results:

• Contextnet demonstrated highly competitive performance compared to existing models on the Jpred dataset.
• Contextnet achieved superior Q3 accuracy on the CASP13 dataset when compared to Jpred, ReportX, Spider3, and MUFold-SS.
• Training procedures were identified as a significant factor influencing Contextnet's prediction accuracy.

Conclusions:

• Contextnet represents a promising novel architecture for protein secondary structure prediction.
• The study highlights the importance of both model architecture and training methodology for achieving high accuracy in SS prediction.
• Further research into optimizing training procedures can enhance the utility of Contextnet and similar models.