Protein Secondary Structure Prediction Based on Multi-scale Convolution and Transformer

Shiwei Yang,Xiaozhou Chen
DOI: https://doi.org/10.54097/hf4n5726
2024-01-01
Abstract:As an important field of natural science, bioinformatics reveals various complex biological facts. The task of protein structure prediction has always been an important research topic in bioinformatics, starting from the primary structure, namely the amino acid sequence, to predict the secondary structure of proteins. In this experiment, amino acid sequences were encoded in various ways, and the encoded data were input into the constructed deep learning network to predict the secondary structure of proteins. In this paper, an MCN-T network is constructed by combining CNN and Transformer. The powerful global feature capture and parallel processing capabilities of the Transformer are combined with the efficient local feature extraction and dimension reduction capabilities of CNN. This combination enables the model to perform well in protein secondary structure prediction tasks. In the experiment, the optimal form of the model is obtained after various parameter optimization experiments. The accuracy of the MCN-T network on the CB6133_filtered dataset and the CB513 dataset is 71.93% and 67.83%, respectively.
What problem does this paper attempt to address?