Prediction of Protein Secondary Structure Using a Hybrid Convolutional Blocks with GRU Units

In order to better understand the function of proteins, protein research in the field of bioinformatics has always been an important issue, and protein structure prediction is also one of the research topics. Starting from the one-dimensional structure (amino acid sequence), it is a good method to predict and classify the secondary structure and adopt deep neural networks to solve sequence problems. We used convolutional blocks combined with gated recurrent units (GRUs) to predict protein secondary structure. Different from previous convolutional neural networks architectures, in this study, we used a mixture of two convolutional blocks of different scales combined with GRUs for sequence labeling for protein secondary structure prediction. Additionally, considering that the coding format in previous studies was too simple, this experiment also added the physical and chemical properties and the logarithmic relative probability of 8 types of secondary structure amino acids as the feature input, which improved the accuracy after the addition of features. Experiments show that our model has good performance in Q8 accuracy.