Matrix factorization with neural network for predicting circRNA-RBP interactions.

Area of Science:

• Biochemistry and Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• Circular RNAs (circRNAs) are vital in human diseases and biological processes, often functioning as scaffolding molecules.
• Interactions between circRNAs and RNA Binding Proteins (RBPs) are crucial for circRNA function.
• High-throughput data enables large-scale prediction of circRNA-RBP interactions, overcoming experimental limitations.

Purpose of the Study:

• To develop a computational framework for predicting unknown circRNA-RBP interaction pairs.
• To utilize Positive Unlabeled (P-U) learning and a kernel model (MFNN) for accurate interaction prediction.
• To analyze circRNA-RBP interactions efficiently using high-throughput data.

Main Methods:

• A computational framework employing Positive Unlabeled (P-U) learning with a Matrix Factorization with Neural Networks (MFNN) kernel model was constructed.
• A circRNA-RBP interaction matrix was created using known interaction data from the CircRic database.
• Neural networks were used to extract latent factors, and P-U learning addressed data imbalance for predicting novel interactions.

Main Results:

• The kernel MFNN model demonstrated superior performance compared to other deep kernel models in predicting circRNA-RBP interactions.
• Deeper hidden layers in the neural network did not necessarily improve model performance.
• The P-U learning strategy effectively scored unlabeled interactions, and predicted pairs were validated against known databases, confirming the method's efficacy.

Conclusions:

• A novel prediction framework for circRNA-RBP interactions was designed, relying solely on the interaction matrix.
• The MFNN method achieves high prediction accuracy, proving effective for analyzing poorly studied circRNA-RBP interactions.