Graph-RPI: predicting RNA-protein interactions via graph autoencoder and self-supervised learning strategies.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• RNA-protein interactions (RPIs) are crucial for cellular functions and disease pathogenesis.
• Existing experimental methods for RPI detection are resource-intensive.
• Computational approaches are needed for efficient RPI prediction.

Purpose of the Study:

• To develop a novel, sequence-based computational framework for predicting RNA-protein interactions (RPIs).
• To address limitations in existing RPI prediction methods, including feature integration and negative sample construction.
• To improve the accuracy and robustness of RPI prediction using graph neural networks (GNNs).

Main Methods:

• Representing RNAs and proteins as nodes in a unified interaction graph.
• Employing multi-feature fusion to enhance the representation of RPI pairs.
• Utilizing self-supervised learning strategies for model training.
• Validating performance using five-fold cross-validation on multiple RPI datasets.

Main Results:

• Achieved high accuracy across various RPI datasets (e.g., 0.979 on RPI1807).
• Demonstrated superior performance in cross-species generalization tests with 0.989 overall accuracy.
• Outperformed existing state-of-the-art RPI prediction methods in robustness and stability.

Conclusions:

• The proposed GNN-based framework offers a robust and stable method for sequence-based RPI prediction.
• This approach has significant potential for broad biological applications and large-scale RPI analysis.
• The method provides an efficient alternative to traditional labor-intensive RPI detection techniques.