Protein secondary structure prediction using modular reciprocal bidirectional recurrent neural networks.

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in structural biology

Background:

• Accurate prediction of protein secondary structure is crucial for understanding protein function.
• Recurrent neural networks (RNNs) show promise for sequence-based structure prediction.
• Existing methods may not fully capture local and long-range dependencies in protein sequences.

Purpose of the Study:

• To develop an advanced supervised learning model for protein secondary structure prediction.
• To integrate local and long-range interactions for improved prediction accuracy.
• To enhance the prediction of protein structural elements from amino acid sequences.

Main Methods:

• Proposed Modular Reciprocal Recurrent Neural Networks (MRR-NN) to model adjacent secondary structure correlations.
• Introduced Multilayer Bidirectional Recurrent Neural Networks (MBR-NN) for long-range intramolecular interactions.
• Developed a combined network integrating MRR-NN and MBR-NN for comprehensive sequence analysis.

Main Results:

• The combined network achieved 79.36% accuracy (Q₃) on the PSIPRED dataset.
• Segment Overlap (SOV) improved to 70.09% using three-fold cross-validation.
• Demonstrated enhanced prediction by effectively utilizing neighboring effects and sequential dependencies.

Conclusions:

• The proposed integrated MRR-NN and MBR-NN model significantly improves protein secondary structure prediction.
• This approach effectively captures both local and global sequence information for accurate structural predictions.
• The developed method offers a powerful tool for bioinformatics research and protein structure analysis.