ProStructNet: Integration of Protein Sequence and Structure for the Prediction of Multi-label Subcellular Localization

The prediction of protein subcellular localization plays a pivotal role in the field of bioinformatics. In recent years, several studies have attempted to incorporate protein structure information into the prediction framework, however, these methods do not utilize advanced feature coding techniques to fully explore the deep features of proteins, which limit the accuracy and effectiveness of protein subcellular localization prediction. In this study, we propose a novel method called ProStructNet that leverages an advanced protein language model (PLM) for precise extraction of protein sequence information. Additionally, ProStructNet independently constructs contact maps using spatial coordinates of amino acids to capture crucial details about protein structure. During the prediction phase, the graph convolutional network (GCN) and multi-head attention mechanism are used to effectively integrate sequence and structure information, leading to significant improvements in predicting accurate labels for multi-label protein subcellular localization. Through rigorous evaluation via cross-validation procedures, ProStructNet has demonstrated remarkable advantages in this domain.