Prediction of lncRNA-miRNA interaction based on sequence and structural information of potential binding site

Long non-coding RNAs (lncRNAs) act as molecular sponges for microRNAs (miRNAs) and indirectly regulate gene expression. Currently, sequence-based prediction methods for lncRNA-miRNA interactions primarily rely on extracting features from full-length sequences, which suffers from the disadvantage of information redundancy.

In this study, we proposed a machine learning method called BSILMI, which predicts lncRNA-miRNA interactions based on sequence and structural information of potential binding site. BSILMI employs XGBoost and focuses on information from potential binding sites between lncRNAs and miRNAs, including the binding free energy, binding site scores, and unpaired probability of RNA folding. BSILMI outperformed LncMirNet, which is a state-of-the-art method. Additionally, we presented a new framework for negative sampling, in which potential interaction pairs are eliminated through sequence similarity alignment. This improves the reliability of the negative sample set. Finally, the key factors influencing the predictions were analyzed using SHAP feature importance analysis.

Our results demonstrated that binding site information plays a crucial role in predicting lncRNA and miRNA interactions. This provides new insights into the research of RNA interactions.