LncMiRPath: A Transformer-based Deep Learning Framework forlncRNAs and miRNAs Interaction Prediction

Introduction/Objective: Interactions between long non-coding RNAs (lncRNAs) and microRNAs (miRNAs) play a critical role in gene regulation and disease mechanisms. However, most existing prediction models rely solely on sequence features, overlooking RNA secondary structures that are essential for accurate interaction prediction. This study introduces LncMiRPath, a Transformer-based framework that integrates both sequence and structural information to enhance predictive performance. Methods: We developed LncMiRPath using a dual-input Transformer architecture that incorporates lncRNA and miRNA sequences alongside their predicted secondary structures. Datasets were obtained from LncBase v3, ENCORI, and miRcode. Secondary structures were inferred using IPknot and represented in dot-bracket notation. We compared three model variants—sequence-only, structure-only, and combined models—using accuracy, precision, recall, and area under the curve (AUC) as performance metrics. Results: LncMiRPath outperformed all baseline models, achieving an AUC of 95% on the curated dataset, demonstrating the effectiveness of integrating structural features. On the independent LncRNASNPv2 dataset, the model maintained strong generalization capability with an AUC of ~91%. Discussion: These results underscore the importance of incorporating RNA secondary structure, a factor often neglected in previous studies. By capturing complementary sequence and structural signals, LncMiRPath not only improves prediction accuracy but also enhances biological interpretability. Although structure inference relies on computational tools such as IPknot, consistent performance across multiple datasets supports the robustness and translational potential of the proposed approach. Future validation with experimental structure data may further strengthen the model. Conclusion: LncMiRPath represents a robust and biologically informed framework for predicting lncRNA–miRNA interactions by jointly leveraging sequence and structural features. This approach advances RNA computational biology and provides a promising tool for RNA-based therapeutic research.