Protein secondary structure prediction with partially recurrent neural networks

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Accurate protein secondary structure prediction is crucial for understanding protein function.
• Recurrent neural networks (RNNs) offer potential for sequence-based prediction tasks.
• Elman networks, a type of RNN, can process sequential data effectively.

Purpose of the Study:

• To evaluate the efficacy of partially recurrent neural networks for protein secondary structure prediction.
• To compare the performance of a modified hierarchical Elman network against feedforward networks.

Main Methods:

• Utilized a reference dataset of 91 proteins for training and 15 non-homologous proteins for testing.
• Implemented a modified, hierarchical Elman network architecture with feedback connections.
• Compared performance against a three-layer feedforward network and a cascaded network.

Main Results:

• The Elman network achieved 67.83% accuracy on the training set and 63.98% on the test set.
• The best feedforward network achieved 62.7% accuracy on the test set.
• A cascaded network incorporating the recurrent network reached 64.49% predictive performance.

Conclusions:

• Partially recurrent neural networks, particularly the modified Elman architecture, show promise for protein secondary structure prediction.
• The implemented recurrent network outperformed standard feedforward networks in this task.
• Cascading recurrent and feedforward networks can further enhance predictive accuracy.