A hybrid deep learning framework for accurate N6,2'-O-Dimethyladenosine site prediction

Area of Science:

• Computational Biology
• Epitranscriptomics
• Bioinformatics

Background:

• N6,2'-O-dimethyladenosine (m⁶Am) is a key RNA modification influencing RNA stability, translation, and gene expression.
• Accurate m⁶Am site identification is vital for epitranscriptomics and understanding disease pathogenesis.
• Existing computational methods lack comprehensive feature representation and interpretability.

Purpose of the Study:

• To develop a novel deep learning framework for accurate and interpretable m⁶Am site prediction.
• To improve the predictive performance and generalizability of m⁶Am site identification across diverse transcriptomic contexts.

Main Methods:

• Integrated heterogeneous sequence-derived features: Accumulated Nucleotide Frequency (ANF), Enhanced Nucleic Acid Composition (ENAC), Dinucleotide-based Auto-Cross Covariance (DACC), Trinucleotide-based Auto-Cross Covariance (TACC), and Pseudo Dinucleotide Composition (PseDNC).
• Employed SHapley Additive exPlanations (SHAP) for feature importance analysis and selection.
• Utilized a deep neural network (DNN) for m⁶Am site classification.

Main Results:

• Achieved 86.55% accuracy using 10-fold cross-validation.
• Demonstrated 85.23% accuracy on an independent test dataset, confirming robust generalizability.
• The SHAP analysis identified and retained the most discriminative features, enhancing model interpretability.

Conclusions:

• The proposed deep learning framework provides an accurate, reliable, and interpretable computational tool for m⁶Am site identification.
• This approach advances epitranscriptomic analysis and the study of RNA modification functions in biological processes and diseases.
• The method offers broad applicability for investigating RNA modification roles in health and disease.