Structure-aware deep learning enhances m6A prediction and reveals cell type-associated RNA structural signatures

Area of Science:

• Epitranscriptomics
• Computational Biology
• Bioinformatics

Background:

• N6-methyladenosine (m6A) is the most prevalent mRNA modification in eukaryotes, crucial for gene regulation and disease.
• Current computational m6A site prediction primarily uses linear sequence features, neglecting RNA structural context.
• RNA structure provides complementary information potentially improving m6A prediction accuracy.

Purpose of the Study:

• To develop a novel deep learning framework, SMART-m6A, integrating both sequence and structural RNA features for m6A site prediction.
• To enhance the accuracy and interpretability of computational m6A site prediction models.
• To investigate the complementary roles of sequence and structure in m6A modification.

Main Methods:

• Proposed SMART-m6A, a deep learning framework utilizing parallel convolutional feature extraction for sequence and structure.
• Employed structure-guided attention mechanisms for multifeature fusion.
• Conducted systematic ablation studies and structural-input perturbation analyses.

Main Results:

• SMART-m6A significantly outperformed existing methods in predicting m6A sites, especially for sequence-ambiguous candidates.
• Learned attention patterns in SMART-m6A showed high concordance with known m6A-binding protein sites.
• Demonstrated that sequence and structural features offer complementary predictive information.

Conclusions:

• Sequence-derived structural features hold significant predictive value for m6A site modeling.
• SMART-m6A provides an accurate and interpretable framework for structure-aware epitranscriptomic prediction.
• The study highlights the importance of integrating structural information for understanding m6A regulation.