Graph-based RNA structural representation reveals determinants of subcellular localization.

Area of Science:

• Computational Biology
• Molecular Biology
• Bioinformatics

Background:

• RNA subcellular localization is crucial for RNA function and regulation.
• Current computational methods for predicting RNA localization are limited by sequence-based or simplified structural features, hindering scalability and applicability across RNA types.
• Existing methods struggle to model inter-label dependencies and regional structural context.

Purpose of the Study:

• To develop a unified computational framework for predicting RNA subcellular localization.
• To improve the accuracy and scalability of RNA localization prediction by incorporating RNA substructure information.
• To capture multi-label dependencies for co-localization patterns across cellular compartments.

Main Methods:

• Developed Graph-based RNA Substructure-Aware Subcellular localization Prediction (GRASP), a graph neural network framework.
• Represented RNAs as heterogeneous, multi-scale graphs with nucleotide and substructure nodes.
• Incorporated multi-label dependency learning to model co-localization patterns.

Main Results:

• GRASP significantly outperforms state-of-the-art sequence-based and structure-informed methods on benchmark datasets.
• Achieved substantial improvements in accuracy, F1-score, and AUC across diverse RNA types.
• Demonstrated strong scalability for long RNA transcripts and provided biologically interpretable insights.

Conclusions:

• GRASP offers a powerful and scalable approach for predicting RNA subcellular localization.
• The substructure-aware graph representation enhances predictive performance and biological interpretability.
• GRASP advances computational methods in RNA biology and localization prediction.