LPI-FKLGCN: Predicting LncRNA-Protein Interactions Through Fast Kernel Learning and Graph Convolutional Network

Abstract

Predicting lncRNA-protein interactions (LPIs) through computational models can not only help to identify the function of lncRNAs, but also help to solve the problem of huge cost of materials and time. In this study, we develop a novel computational model combining fast kernel learning (FKL) and multi-layer graph convolution network (GCN) to identify potential lncRNA-protein interaction (LPI-FKLGCN). The LPI-FKLGCN model can fuse the multi-source features and similarities by the FKL technique and code the embedding representive vectors by the multi-layer graph convolution network. Through 5-fold cross-validation, the LPI-FKLGCN obtains an AUPR value of 0.52 and an AUC value of 0.96, which is superior to other methods. In case studies, most of the predicted LPIs are confirmed by the newly published biological experiments. It can be seen that the fusion of multi-source similarities and features, combined with multi-layer embedding vectors from graph convolution network can improve the accuracy of LPI prediction and the model of LPI-FKLGCN is an efficient and accurate tool for LPI prediction.