Rprot-Vec: A deep learning approach for fast protein structure similarity calculation

The prediction of structural similarity and homology detection of protein sequences has always been a very challenging and valuable research. Based on protein structure similarity, rapid and accurate detection of homologous proteins can infer the functions of newly discovered or unannotated proteins. If we can use a method to quickly predict the structural similarity between two proteins, we can achieve the task of homologous detection of protein sequences. By using this method with partially known protein structures, we can quickly infer the functions of more unknown proteins. In this article, we propose an enhanced model Rprot-vec (Rapid Protein Vector) based on deep learning, which improves existing methods. It can quickly calculate the structural similarity of proteins and perform protein homology detection using only protein sequences. Meanwhile, our model can also be used to simulate proteins to accomplish more downstream tasks related to proteins. Our model performs well within the very important TM-score range, with an accurate similarity prediction rate of 65.3% for homologous proteins. In addition, for all protein pairs within the TM-score range, the average prediction error is 0.0561, indicating good performance. The superiority of our model can be demonstrated from various aspects. In addition, we have also supplemented the dataset in this field, providing data support for future researchers studying this field simultaneously.