RNA knowledge-graph analysis through homogeneous embedding methods

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• The RNA-knowledge graph (RNA-KG) integrates diverse RNA data from over 60 public databases.
• RNA-KG represents functional relationships and interactions between RNA molecules, biomolecules, chemicals, and biomedical concepts within graph-structured bio-ontologies.

Purpose of the Study:

• To perform the first comprehensive computational analysis of RNA-KG.
• To evaluate the potential of graph representation learning and machine learning models for predicting node types and edges within RNA-KG.

Main Methods:

• Node classification experiments were conducted to predict up to 81 distinct node types.
• Both generic-edge prediction (presence of an edge) and specific-edge prediction (e.g., miRNA-miRNA, siRNA-mRNA, miRNA-disease) were performed.
• Homogeneous graph embedding methods (LINE, node2vec) combined with machine learning models (decision trees, random forests) were utilized.

Main Results:

• Balanced accuracy exceeded 90% for predicting the 20 most common node types.
• Over 80% accuracy was achieved for most specific-edge prediction tasks.
• Simple embedding methods for homogeneous graphs successfully predicted nodes and edges within RNA-KG.

Conclusions:

• Computational analysis validates the predictive power of graph representation learning on RNA-KG.
• These findings pave the way for discovering novel non-coding RNA (ncRNA) interactions.
• The study lays the foundation for enhanced prediction accuracy and further research into the 'RNA world'.