FourierAN-m¹A: A Novel Fourier-Domain Adaptive Feature Extraction Framework for m¹A RNA Modification Prediction.

Area of Science:

• Molecular Biology
• Bioinformatics
• Computational Biology

Background:

• RNA modifications regulate gene expression, with N1-methyladenosine (m1A) being a key post-transcriptional modification.
• Accurate identification of m1A sites is essential for understanding its regulatory roles.
• Current prediction models face challenges in feature extraction, long-range dependency modeling, and interpretability, with limited investigation into modification-loss variants.

Purpose of the Study:

• To address limitations in existing m1A prediction models.
• To develop a high-performance deep learning framework for precise m1A site identification.
• To systematically investigate modification-loss variants.

Main Methods:

• Developed a high-quality dataset of m1A modifications and loss-of-modification variants.
• Proposed FourierAN-m1A, a novel deep learning framework integrating CNNs, AFNO, and Transformer encoder.
• Implemented a dynamic filtering mechanism using complex-valued neural networks and a learnable adaptive mask for joint frequency- and spatial-domain feature learning.

Main Results:

• The proposed FourierAN-m1A model significantly outperforms existing methods in prediction accuracy.
• The model demonstrates biological interpretability, offering insights into m1A modification mechanisms.
• Successfully identified m1A sites with high precision.

Conclusions:

• FourierAN-m1A provides a high-performance and interpretable approach for precise m1A site identification.
• The developed framework advances the study of RNA modifications and their regulatory functions.
• Highlights the importance of considering modification-loss variants in RNA modification analysis.