DCMA: faster protein backbone dihedral angle prediction using a dilated convolutional attention-based neural network.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein backbone dihedral angles are crucial for determining 3D protein structure.
• Existing computational prediction methods, especially deep learning models, are often resource-intensive and time-consuming to train.
• There is a need for efficient and accurate methods for predicting protein structural features.

Purpose of the Study:

• To introduce a novel, lightweight deep learning method named Dilated Convolution and Multi-Head Attention (DCMA) for predicting protein backbone torsion dihedral angles.
• To evaluate the performance of DCMA against existing state-of-the-art methods on benchmark datasets.
• To demonstrate DCMA's potential as an alternative for predicting other protein structural features.

Main Methods:

• DCMA utilizes a novel architecture comprising hybrid inception blocks and a multi-head attention block (I2A1 module).
• Hybrid inception blocks combine multi-scale and dilated convolutional neural networks to capture both local and long-range sequence-based features.
• The multi-head attention block further enhances the feature extraction capabilities.

Main Results:

• DCMA achieved better or comparable generalization performance on Critical Assessment of Protein Structure Prediction (CASP) benchmark datasets.
• The proposed DCMA model is an individual model, outperforming ensemble models in terms of efficiency.
• DCMA demonstrated a significantly shorter training time and lower computational resource requirements compared to existing heavyweight methods.

Conclusions:

• DCMA presents a highly efficient and effective lightweight alternative for predicting protein backbone torsion dihedral angles.
• The method's reduced computational demands and training time make it a practical tool for structural bioinformatics.
• DCMA's architecture holds promise for application in predicting diverse protein structural characteristics.