Protein Solvent-Accessibility Prediction by a Stacked Deep Bidirectional Recurrent Neural Network.

Area of Science:

• Computational Biology
• Biophysics
• Bioinformatics

Background:

• Residue solvent accessibility is crucial for understanding protein structure and function.
• Accurate prediction of solvent accessibility aids in protein analysis.

Purpose of the Study:

• To develop a deep learning method for predicting residue solvent accessibility.
• To improve the accuracy of solvent accessibility predictions using sequence-derived features.

Main Methods:

• A stacked deep bidirectional recurrent neural network was employed.
• Novel merging operators were introduced to capture long-range sequence information.
• A training database of 7361 proteins was utilized, with sequence-derived features representing each residue.

Main Results:

• The deep learning method achieved improved prediction quality compared to current approaches.
• Mean absolute error of 8.8% and Pearson's correlation coefficient of 74.8% on the CB502 dataset.
• Mean absolute error of 8.2% and Pearson's correlation coefficient of 78% on the Manesh215 dataset.

Conclusions:

• The proposed deep learning method effectively predicts residue solvent accessibility.
• The approach offers a valuable tool for protein structure and function studies.
• Enhanced prediction accuracy facilitates further bioinformatics research.