A combined deep learning framework for mammalian m6A site prediction.

Area of Science:

• Epigenetics and RNA biology
• Computational biology and bioinformatics
• Molecular mechanisms of gene regulation

Background:

• N6-methyladenosine (m6A) is the most abundant mRNA modification in eukaryotes, crucial for various cellular processes.
• Accurate m6A site identification is essential for understanding its functional roles and regulatory mechanisms.
• Existing m6A prediction tools require improvement in accuracy and performance.

Purpose of the Study:

• To develop a novel computational framework, deepSRAMP, for precise identification of m6A sites.
• To leverage sequence-based and genome-derived features for enhanced m6A prediction accuracy.
• To compare deepSRAMP's performance against existing state-of-the-art m6A predictors.

Main Methods:

• Designed a hybrid deep learning model integrating Transformer architecture and recurrent neural networks.
• Utilized sequence-based and genome-derived features for m6A site prediction.
• Evaluated performance using benchmark datasets and compared with SRAMP, WHISTLE, and DeepPromise.

Main Results:

• deepSRAMP significantly outperforms its predecessor SRAMP and other leading predictors like WHISTLE and DeepPromise.
• Achieved an average 16.1% and 18.3% increase in AUROC, and 43.9% and 46.4% increase in AUPRC, respectively.
• Demonstrated successful application in mammalian m6A epitranscriptome mapping and identification of differential m6A sites.

Conclusions:

• deepSRAMP represents a significant advancement in m6A site identification tools.
• The enhanced accuracy facilitates deeper investigation into m6A functions across diverse cellular conditions.
• This tool can aid in mapping epitranscriptomes and uncovering isoform-specific m6A modifications.