MJnet: A lightweight RNN-based model for microRNA target site prediction

Abstract

Accurate prediction of microRNA (miRNA) target sites is critical for understanding post-transcriptional gene regulation. While recent deep learning models have achieved high predictive accuracy, many suffer from excessive computational complexity and limited interpretability. In this study, we propose MJnet, a lightweight and efficient deep learning model based on a Bidirectional Gated Recurrent Unit (BiGRU) architecture, integrated with simple C2 encoding, a multi-scale one-dimensional convolutional network (TextCNN), and a self-attention mechanism. This framework captures both local sequence features and global contextual dependencies while maintaining low computational cost. Extensive experiments on experimentally validated datasets demonstrate that our model outperforms several traditional and deep learning-based baselines, including Mimosa, in terms of accuracy, F1-score, and robustness across balanced gene-level test sets. Ablation studies confirm the effectiveness of each module, and attention heatmaps reveal interpretable patterns aligned with known seed regions. Our approach offers a practical, reproducible, and interpretable solution for miRNA target site prediction in biologically relevant contexts.