Enhanced hierarchical attention networks for predictive interactome analysis of LncRNA and CircRNA in oral herpes virus

Background

Non-coding RNAs, including lncRNAs, circRNAs, and microRNAs, constitute 98 % of the human transcriptome and are vital regulators of gene expression, cellular processes, and host-pathogen interactions, particularly in viral infections. This study explores lncRNA-circRNA interactions and their biological significance in oral viral infections.

Methods

ViRBase, a database with over 820,000 interactions involving 50,000 RNAs from 116 viruses and 36 host organisms, was used to analyze herpesvirus datasets. The study employed hierarchical attention and knowledge graph embeddings to represent nodes and edges in the knowledge graph. These served as input features for a hierarchical attention model trained over 100 epochs. Model performance was evaluated based on loss calculation, optimization, and attention weight stability.

Results

The model achieved a final loss of 0.000180 at Epoch 100, with stable attention weights confirming reliability. Node embedding statistics showed a mean of 0.005110 and a standard deviation of 0.013370, while attention weights had a high mean of 0.997178, emphasizing model robustness.

Conclusion

This study provides insights into lncRNA-circRNA interactions in herpes viral infections, enhancing therapeutic development, disease progression monitoring, and understanding host-pathogen interactions, paving the way for targeted interventions and improved outcomes.