PeTMiner: Mining plant endogenous target mimics via ternary interaction modeling with hierarchical attention and adaptive weighted loss

In the plant competitive endogenous RNA (ceRNA) mechanism, endogenous target mimics (eTMs) compete with messenger RNAs (mRNAs) to bind microRNAs (miRNAs), which regulates mRNA expression. Efficiently mining eTMs is essential for exploring the functional impacts of the regulatory networks. However, effective computational methods are still lacking. To fill this gap, we propose PeTMiner, an end-to-end computational framework that mines plant eTMs by identifying eTM-miRNA-mRNA ternary interactions. PeTMiner leverages a pre-trained foundation model specialized for plant RNAs to extract deep semantic representations from raw sequences. It employs a hierarchical attention mechanism to compress long sequence representations and fuse cross-molecular representations. It also utilizes an adaptive weighted loss function to mitigate the model optimization bias caused by data imbalance across species. Extensive experiments demonstrate that PeTMiner achieves superior performance and strong generalization ability on standard, cross-species, and imbalanced cross-platform test sets. Furthermore, PeTMiner supports biological interpretability and provides reliable candidates for functional annotation and enrichment analysis.