Predicting LncRNA-Protein Interactions Through Global and Local Features based on Cross Attention Mechanism

In various biological processes, long non-coding RNAs (lncRNAs) have been recognized for their essential roles. Investigating lncRNA-protein interactions (LPIs) is crucial for unveiling the yet undiscovered molecular functions of lncRNAs. Over recent years, computational approaches have gained prominence as efficient alternatives to traditional, time-consuming experiments for exploring potential unknown associations. However, prevailing methods often oversimplify by merely concatenating information from lncRNA and protein sequences, falling short of fully capturing the intricate interaction mechanisms. In this study, we propose an end-to-end deep learning approach called LPIGLAM for predicting LPIs based on the global features and local interaction features of lncRNA and protein. First, the raw sequence features are extracted using dilated inception modules. These modules help capture information at different scales and improve the representation of the input sequences. Then, a shared-weight-based CrossAttention is used to extract the local structural interaction features between lncRNA and protein. The global features of lncRNA and protein are extracted using a two layer fully connected network. Finally, all the extracted features are fused and input into fully connected dense layers in the downstream task for predicting LPIs. LPIGLAM considers the interaction features of lncRNA and protein, simulating the local structural interaction characteristics from a biological perspective, while utilizing global features to enhance the prediction performance. The experimental results on three datasets of different scales and sizes indicate a significant improvement of our method compared to the latest approaches. Furthermore, we conducted cross validation experiments by randomly masking lncRNA or protein to assess the model’s capability to capture substructures when confronted with unknown lncRNA (protein).