MCLCBA: multi-view contrastive learning network for RNA methylation site prediction.

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• RNA methylation (RM) is crucial for gene regulation, RNA stability, and protein translation.
• Accurate prediction of RM sites is vital but challenged by complex, costly wet-lab methods.
• Existing deep learning models struggle with performance degradation on smaller datasets.

Purpose of the Study:

• To develop an effective deep learning framework for predicting RNA methylation modification sites, particularly in sample-limited scenarios.
• To overcome the limitations of existing methods that exhibit performance degradation with reduced training data.
• To improve the accuracy and generalization of RNA methylation site prediction models.

Main Methods:

• Proposed a Multi-view Contrastive Learning with CNN-BiLSTM-Attention (MCLCBA) framework.
• Utilized a multi-view approach with DNA Bidirectional Encoder Representations from Transformers (DNABERT) for sequence features and Chaos Game Representation (CGR) for structural features.
• Implemented dual differential data augmentation, multi-view encoders, projection heads, and contrastive loss functions for robust feature learning.

Main Results:

• The MCLCBA framework demonstrated superior performance on a sample-limited m⁷G dataset.
• Achieved Area Under the Receiver Operating Characteristic Curve (AUROC) of 85.64% and Area Under the Precision-Recall Curve (AUPRC) of 86.94%.
• Outperformed existing methods by 5-6% in both AUROC and AUPRC, effectively addressing sample-limited feature learning.

Conclusions:

• Multi-view contrastive learning offers a promising approach for RNA methylation site prediction in data-scarce environments.
• The MCLCBA framework effectively learns discriminative and generalizable features from limited data.
• This study provides a valuable tool for advancing the understanding of RNA methylation's biological roles.