FSFT6mA: a feature-synthesis fine-tuning framework for DNA 6mA site prediction

Area of Science:

• Epigenetics and Genomics
• Computational Biology
• Bioinformatics

Background:

• DNA N6-methyladenine (6mA) is a crucial epigenetic modification regulating gene expression, vital for biological processes and disease understanding.
• Accurate identification of 6mA sites is critical, yet current computational methods primarily rely on sequence-derived features, limiting predictive performance.
• Exploring novel feature representations is essential to advance the accuracy of 6mA site prediction.

Purpose of the Study:

• To introduce FSFT6mA, a novel deep learning framework for enhanced DNA 6mA site prediction.
• To leverage feature synthesis using Generative Adversarial Networks (GANs) to improve prediction accuracy.
• To evaluate the performance of FSFT6mA against existing state-of-the-art predictors.

Main Methods:

• A deep convolutional neural network (CNN) was initially trained on original datasets.
• A Generative Adversarial Network (GAN) was employed to generate synthetic features from intermediate CNN layers.
• The model was fine-tuned using these GAN-generated features to enhance prediction capabilities.

Main Results:

• FSFT6mA demonstrated significant performance gains, improving the Matthews Correlation Coefficient (MCC) by 2.6% in *A. thaliana* and 1.9% in *D. melanogaster* compared to models without synthetic features.
• Independent validation confirmed FSFT6mA's superior performance, achieving high Area Under the Curve (AUC) values of 0.969 for *A. thaliana* and 0.968 for *D. melanogaster*.
• The framework outperformed existing state-of-the-art DNA 6mA site predictors.

Conclusions:

• FSFT6mA represents an accurate and effective computational tool for DNA 6mA site prediction.
• The integration of GAN-generated synthetic features significantly enhances predictive performance.
• The study provides freely accessible data and code for broader research application.