Fusion-m6A: A lightweight hybrid deep learning framework for RNA m6A site prediction.

Area of Science:

• Molecular Biology
• Bioinformatics
• Genomics

Background:

• N6-methyladenosine (m6A) is the most prevalent mRNA modification, crucial for RNA metabolism, gene regulation, and disease.
• Accurate m6A site identification is vital for understanding its functional and biological roles.
• Experimental methods like Nanopore direct RNA sequencing (DRS) are effective but costly and labor-intensive.

Purpose of the Study:

• To develop a computationally efficient and accurate framework for predicting m6A sites.
• To overcome limitations of existing computational methods that rely on handcrafted features or expensive models.
• To provide a scalable solution for large-scale and tissue-specific m6A site prediction.

Main Methods:

• Developed Fusion-m6A, a hybrid deep learning framework.
• Integrated Word2Vec sequence embeddings, convolutional layers for motif detection, and GRU with attention for long-range dependencies.
• Incorporated auxiliary k-mer features and fused representations for m6A site prediction.

Main Results:

• Fusion-m6A demonstrated superior accuracy and Matthews correlation coefficient compared to state-of-the-art predictors across multiple human tissues and cell lines.
• The model achieves significantly faster inference times.
• Fusion-m6A requires substantially less memory, enhancing its practicality.

Conclusions:

• Fusion-m6A offers a robust and efficient computational approach for m6A site prediction.
• The framework provides a practical solution for large-scale and tissue-specific m6A profiling.
• The publicly available implementation promotes reproducibility and further research in RNA modification analysis.