StableDNAm: towards a stable and efficient model for predicting DNA methylation based on adaptive feature correction...

Area of Science:

• Genomics
• Computational Biology
• Epigenetics

Background:

• Accurate DNA methylation prediction is crucial for understanding biological processes.
• Deep learning models offer enhanced precision but suffer from stability and generalization issues.

Purpose of the Study:

• To develop an efficient, stable, and generalizable DNA methylation prediction model.
• To address the limitations of existing deep learning approaches in DNA methylation prediction.

Main Methods:

• Developed the StableDNAm model incorporating feature fusion, adaptive feature correction, and contrastive learning.
• Encoded DNA sequences at four levels for comprehensive multi-scale feature capture.
• Implemented a sequence-specific feature correction module for adaptive weight adjustment.
• Utilized contrastive learning to mitigate data sparsity-induced instability.

Main Results:

• The model demonstrated robust stability and performance across multiple standard datasets.
• Experiments on a unified dataset confirmed the model's strong adaptability.
• The proposed methods enhanced model stability, scalability, and generalization.

Conclusions:

• The StableDNAm model is a general, stable, and effective tool for DNA methylation prediction.
• This model shows significant potential for advancing methylation-related research and analysis.