Multi-S3p: protein secondary structure prediction with specialized multi-network and self-attention-based deep learning model

Abstract

Protein structure prediction (PSP) is a vital challenge in bioinformatics, structural biology and drug discovery. Protein secondary structure (SS) prediction is critical since three-dimensional (3D) structures are primarily made up of secondary structures. With the advancement of deep learning approaches, SS classification accuracy has been significantly improved. Many existing methods use an ensemble of complex neural networks to improve SS prediction. Because of the high dimensionality of the hyperparam eter space, deep neural networks with complex architectures are typically challenging to train effectively. Also, predicting secondary structures in the boundary regions between different types of SS is challenging. This study presents Multi-S3P, which employs bidirectional Long-Short-Term-Memory (BILSTM) and Convolutional Neural Networks (CNN) with a self-attention mechanism to improve the secondary structure prediction using an effective training strategy to capture the unique characteristics of each type of secondary structure and combine them more effectively. The ensemble of CNN and BILSTM can learn both contextual information and long-range interactions between the residues. In addition, using a self-attention mechanism allows the model to focus on the most important features for improving performance. We used the SPOT 1D dataset for the training and validation of our model using a set of four input features derived from amino acid sequences. Further, the model was tested on four popular independent test datasets and compared with various state-of-the-art predictors. The presented results show that Multi-S3P outperformed the other methods in terms of Q3, Q8 accuracy and other performance metrics, achieving the highest Q3 accuracy of 87.57% and a Q8 accuracy of 77.56% on the TEST2016 test set. More importantly, Multi-S3P demonstrates high performance in SS boundary regions. Our experiment also demonstrates that the combination of different input features and a multi-network-based training strategy significantly improved the performance.