MTAGCN

MTAGCN predicts miRNA-target associations in Camellia sinensis var. assamica (CSA) by integrating heterogeneous biological networks with graph convolutional networks and an attention mechanism to generate embeddings for scoring potential miRNA–mRNA interactions.

Key Features:

• Heterogeneous network integration: Integrates miRNA similarity, target similarity, and miRNA-target association networks to represent CSA miRNA–mRNA relationships.
• miRNA similarity network: Captures similarities between miRNAs based on sequence or functional attributes.
• Target similarity network: Represents relationships among mRNA targets potentially based on gene expression profiles or other biological data.
• miRNA-target association network: Encodes known associations between miRNAs and their target mRNAs.
• Graph convolutional networks (GCNs): Applies graph convolution layers to the integrated network to extract node embeddings for miRNAs and targets.
• Attention mechanism across GCN layers: Employs an attention mechanism to dynamically weight outputs from multiple graph convolution layers when forming final embeddings.
• Integrated embeddings and scoring: Combines multi-layer outputs into integrated embeddings that are used to score potential, unobserved miRNA–target associations.
• Comparative performance: Experimentally validated to outperform traditional machine learning methods and other GCN-based approaches in both balanced and unbalanced task settings.

Scientific Applications:

• miRNA–target prediction in CSA: Predicts candidate miRNA–mRNA interactions specific to Camellia sinensis var. assamica.
• Regulatory mechanism analysis: Facilitates investigation of miRNA-mediated regulation of growth, development, and stress response in CSA.
• Experimental prioritization: Prioritizes candidate miRNA–mRNA pairs for experimental validation.
• Precision plant breeding: Informs precision breeding efforts by identifying regulatory interactions relevant to trait modulation.

Methodology:

Constructs a heterogeneous network integrating miRNA similarity, target similarity, and known miRNA-target associations, applies graph convolutional network layers with an attention mechanism across layers to generate integrated embeddings, and scores those embeddings to predict unobserved miRNA–target associations.