DeepACLSTM: deep asymmetric convolutional long short-term memory neural models for protein secondary structure...

Area of Science:

• Computational biology
• Bioinformatics
• Structural biology

Background:

• Protein secondary structure (PSS) prediction is crucial for tertiary structure prediction, protein function understanding, and drug design.
• Experimental PSS determination is costly and time-consuming, necessitating efficient computational methods.
• Current deep learning methods for PSS prediction often overlook the feature vector dimension of protein matrices.

Purpose of the Study:

• To develop an efficient computational method for predicting 8-category PSS using only sequence information.
• To explore the utility of the feature vector dimension in protein feature matrices for enhancing PSS prediction.
• To improve the accuracy of PSS prediction by capturing complex sequence-structure relationships.

Main Methods:

• Proposed DeepACLSTM, a novel deep neural network integrating asymmetric convolutional neural networks (ACNNs) and bidirectional long short-term memory (BLSTM) networks.
• ACNNs were used to extract local amino-acid contexts, while BLSTM captured long-distance interdependencies.
• The method leverages the feature vector dimension of protein feature matrices for prediction.

Main Results:

• DeepACLSTM demonstrated superior performance compared to state-of-the-art baselines on three benchmark datasets (CB513, CASP10, CASP12).
• The method successfully predicted 8-category PSS from protein sequence and profile features.
• Results confirmed the effectiveness of combining ACNNs and BLSTM for PSS prediction.

Conclusions:

• DeepACLSTM is an efficient and accurate method for 8-category PSS prediction.
• The study highlights the importance of the feature vector dimension for extracting complex sequence-structure relationships.
• The developed method offers a promising computational approach for advancing PSS prediction.