PMTIDMA: plant miRNA target interaction prediction framework based on denoising autoencoder and multi-scale spectral graph convolutional attention

Accurate prediction of plant miRNA–target interactions(MTIs) is fundamental to understanding gene regulation in plant growth, development, and stress responses. Experimental validation of MTIs is accurate but costly, time-consuming, and limited in scalability, especially in plant species. In contrast, computation-based methods can efficiently mine potential regulatory relationships, but existing models still exhibit deficiencies in feature representation capability, structural information modeling, and performance under limited-sample conditions. This paper proposes PMTIDMA, a novel plant miRNA–target gene prediction framework based on graph neural networks. It first uses a Denoising Autoencoder(DAE) to reduce dimensionality of miRNA and mRNA sequence features, enhancing robustness and reducing noise. Then, heterogeneous graphs are built from miRNA–miRNA and mRNA–mRNA sequence similarities and known miRNA–mRNA interactions. On this basis, the sequence features and heterogeneous graphs are jointly input into the multi-scale spectral graph convolutional network for joint training. This network performs weighted fusion on the convolution results of spectral graphs of different scales through the attention mechanism, and further combines the standard Graph convolutional network layer and the Jumping Knowledge module to fully capture the local and global structural information. Ultimately, the multi-layer perceptron was used to predict the interaction probability between miRNA and target genes. Experimental results on multiple plant datasets show that PMTIDMA consistently outperforms existing mainstream methods. Specifically, PMTIDMA achieves an AUPRC of 0.9302, an AUROC of 0.9369, and an F1-score of 0.9351. The ablation experiment further verified the effectiveness of the denoising autoencoder, multi-scale spectral graph convolution and attention fusion mechanism in the model. In conclusion, PMTIDMA provides an efficient and robust computational framework for predicting the interaction between plant mirnas and target genes, which is conducive to the systematic analysis of the plant miRNA regulatory network.