RPI-WGNN: ncRNA-Protein Interaction Prediction based on Weighted Graph Neural Networks

Non-coding RNA (ncRNA) is functional RNA derived from DNA transcription. Although ncRNA does not directly participate in protein translation, it often plays crucial roles in biological processes via interacting with specific proteins. We proposed a novel model for predicting ncRNA-protein interactions (RPI), which takes into account weighted graphs of molecular internal structures in proteins and comprehensive sequence information of ncRNA. Using a reliable negative sample partitioning method, we feed the predicted protein graphs and ncRNA numerical vectors to learn topology structure information and global features. Subsequently, we accurately and efficiently extract comprehensive high-order intrinsic feature representations of proteins using weighted graph neural networks, and concatenate them with embedding representations of global information encoded by ncRNA sequences. Finally, the molecular structure information and sequence feature information are combined as the ultimate representation, which is then fed into the classifier to determine the presence of interactions. The results of the five-fold cross-validation experiments demonstrate that our model achieved accuracy of 99.94% and 98.92% in RPI10412 and RPI369. Moreover, it performs better than state-of-the-art methods on the benchmark datasets.