AttMVGraph: Attention-Based Multimodal Fusion and Variational Graph Learning for SM-miRNA Association Prediction.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• MicroRNAs (miRNA) are crucial noncoding RNAs regulating gene expression and are therapeutic targets.
• Experimental identification of small molecule (SM)-miRNA interactions is costly and inefficient.
• Developing accurate predictive models is essential for advancing SM-miRNA-based therapeutics.

Purpose of the Study:

• To propose an effective computational method, AttMVGraph, for predicting SM-miRNA associations.
• To leverage attention-based multimodal fusion and variational graph learning for improved prediction accuracy.
• To address the challenges of data imbalance and enhance model robustness.

Main Methods:

• Utilized Random Walk with Restarts (RWR) for topological features and SM/miRNA similarity as multimodal inputs.
• Employed feature-enhanced channel attention (FECA) for adaptive weighted fusion and discriminative graph embeddings.
• Integrated a Variational Graph Autoencoder (VGAE) for uncertainty modeling and representation learning.
• Incorporated dynamic hard negative mining (DHNM) to handle sample imbalance and strengthen decision boundaries.

Main Results:

• AttMVGraph achieved excellent performance in 5-cross-validation.
• Achieved AUC values of 0.9937 ± 0.0061 on Dataset 1 and 0.9727 ± 0.0001 on Dataset 2.
• Achieved AUPR values of 0.9397 ± 0.0753 on Dataset 1 and 0.8807 ± 0.0589 on Dataset 2.
• Demonstrated the effectiveness and superiority of the proposed AttMVGraph model.

Conclusions:

• AttMVGraph significantly improves the prediction of SM-miRNA associations.
• The model's architecture effectively handles multimodal data and complex relationships.
• The findings support the utility of AttMVGraph in accelerating the discovery of SM-miRNA therapeutics.