Single-sequence-based prediction of protein secondary structures and solvent accessibility by deep whole-sequence...

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in protein science

Background:

• Protein structure prediction is crucial for understanding protein function.
• Traditional methods heavily rely on evolutionary information from multiple sequence alignments.
• Long Short-Term Bidirectional Recurrent Neural Networks (LSTM-BRNNs) show promise in capturing sequence dependencies.

Purpose of the Study:

• To develop and evaluate a single-sequence-based protein structure prediction method.
• To assess the performance of LSTM-BRNNs for predicting various protein structural properties.
• To provide an efficient computational tool for large-scale protein analysis.

Main Methods:

• Utilized Long Short-Term Bidirectional Recurrent Neural Networks (LSTM-BRNNs) for prediction.
• Developed a method named SPIDER3-Single, operating solely on single protein sequences.
• Evaluated prediction accuracy for Q3, solvent accessible surface area, secondary structure, and backbone angles.

Main Results:

• SPIDER3-Single achieved a Q3 accuracy of 72.5% and a 0.67 correlation coefficient for solvent accessible surface area.
• The method accurately predicted eight-state secondary structure, main-chain angles, half-sphere exposure, and contact number.
• Outperformed evolutionary-based methods for proteins with limited sequence homologs.

Conclusions:

• Single-sequence-based prediction using LSTM-BRNNs is a viable and accurate approach.
• SPIDER3-Single offers a computationally efficient alternative for protein structure property prediction.
• The method is accessible via the SPIDER3 server and as a standalone download.