GINClus: RNA structural motif clustering using graph isomorphism network

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• RNA structural motif identification is vital for understanding RNA function but is complex and time-consuming.
• Discovering novel RNA structural motif families often relies on manual analysis, presenting a significant challenge.
• Existing methods struggle with the inherent complexity and variability of RNA 3D structures.

Purpose of the Study:

• To develop an automated tool, GINClus, for clustering RNA structural motif candidates.
• To leverage semi-supervised deep learning for accurate RNA motif clustering based on structural similarities.
• To facilitate the discovery of new RNA structural motif instances and families.

Main Methods:

• GINClus utilizes a semi-supervised deep learning model, specifically a graph isomorphism network (GIN), for clustering.
• RNA motif candidates (loop regions) are converted into graph representations capturing base interactions and 3D structures.
• Clustering is performed using a combination of GIN, K-means, and hierarchical agglomerative clustering algorithms.

Main Results:

• GINClus achieved high clustering accuracy: 87.88% for internal loop motifs and 97.69% for hairpin loop motifs.
• The tool successfully grouped known motifs into their respective families.
• GINClus identified 927 new instances of established RNA motif families and discovered 12 novel RNA structural motif families.

Conclusions:

• GINClus provides an effective computational approach for RNA structural motif clustering and discovery.
• The tool enhances the identification of known and novel RNA motifs, aiding structural and functional studies.
• This work significantly advances the automated analysis of RNA structures and the discovery of new motif families.