Ensemble deep learning models for protein secondary structure prediction using bidirectional temporal convolution and...

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Protein secondary structure prediction (PSSP) is crucial in computational biology.
• Existing deep learning models struggle with comprehensive long-range feature extraction in long protein sequences.

Purpose of the Study:

• To develop a novel deep learning model for improved PSSP.
• To enhance the extraction of local, global, and multi-scale long-range features in protein sequences.

Main Methods:

• Proposed a novel deep learning model incorporating Bidirectional Temporal Convolutional Network (BTCN), Bidirectional Long Short-Term Memory (BLSTM), and Multi-scale Bidirectional Temporal Convolutional Network (MSBTCN).
• Implemented feature fusion from 3-state and 8-state PSSP.
• Compared various model combinations, including BLSTM with TCN variants, and evaluated forward vs. reverse prediction.

Main Results:

• The proposed MSBTCN effectively captures multi-scale long-range features.
• Fusing 3-state and 8-state PSSP features improved prediction accuracy.
• Reverse prediction demonstrated superior performance, indicating the influence of later amino acid positions.

Conclusions:

• The novel deep learning architecture significantly advances PSSP accuracy.
• The findings suggest that incorporating multi-scale features and considering reverse prediction enhance model performance.
• The developed methods outperform existing state-of-the-art approaches on benchmark datasets.