MiRInter-Trans: a Transformer-Based Framework for microRNA Interaction Prediction

Abstract The accurate prediction of microRNA (miRNA) interactions represents a significant challenge in computational biology, with far-reaching implications for understanding gene regulatory networks and developing RNA-based therapeutics. We present miRInter-Trans, an innovative computational framework that synergizes a pre-trained RNA foundation model (RNA-FM) with a feed-forward neural network to predict miRNA interactions using solely sequence information. Our approach harnesses the power of transformer-based embeddings to capture intricate sequence patterns and potential structural motifs without relying on handcrafted features or thermodynamic parameters. Through comprehensive evaluation across three distinct interaction types (miRNA-lncRNA, miRNA-miRNA, and miRNA-snoRNA), we demonstrate that miRInter-Trans outperforms traditional Minimum Free Energy methods and other recently proposed computational approaches, achieving AUROC above 0.9 in both the miRNA interactions datasets used in our experiments. Notably, our models can accurately perform “de novo” miRNA interaction prediction, i.e. prediction of interactions for miRNA for which no or very reduced interaction data are available.