MCAMEF-BERT: an efficient deep learning method for RNA N7-methylguanosine site prediction via multi-branch feature...

Area of Science:

• Bioinformatics
• Molecular Biology
• Computational Biology

Background:

• N7-methylguanosine (m7G) modifications are crucial regulators in human development and diseases like cancer.
• Current prediction methods for m7G sites lack representational power, feature fusion efficiency, and biological insight.

Purpose of the Study:

• To develop a novel deep learning model, MCAMEF-BERT, for accurate and interpretable prediction of m7G modification sites.
• To overcome limitations of existing methods in feature extraction, fusion, and biological knowledge integration.

Main Methods:

• Proposed MCAMEF-BERT, a parallel deep learning architecture integrating DNABERT-2 and traditional feature encoding.
• Implemented a multi-channel attention module to mitigate redundant feature fusion.
• Utilized m7GHub datasets for model training and evaluation.

Main Results:

• MCAMEF-BERT achieved superior accuracy and effectiveness compared to state-of-the-art classifiers on m7GHub datasets.
• Demonstrated model interpretability via in silico saturation mutagenesis experiments.
• Confirmed robustness in motif recognition and generalization across diverse RNA modification prediction tasks.

Conclusions:

• MCAMEF-BERT offers a powerful and interpretable approach for m7G site prediction.
• The model advances RNA modification analysis, aiding research in human development and cancer.
• Highlights the potential of integrating pre-trained models with attention mechanisms for biological sequence analysis.