Prediction of DNA Methylation With Long-Range State-Space Models.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• DNA methylation is crucial for gene regulation and can be predicted from DNA sequence.
• Deep learning models, including transformers and CNNs, are used for DNA methylation prediction.
• Accurate prediction helps impute methylation status for cytosines with limited sequencing data.

Purpose of the Study:

• To evaluate the performance of long-range state-space models, specifically the Hyena architecture, for DNA methylation prediction in six plant species.
• To develop and fine-tune a HyenaDNA framework for accurate genome-wide DNA methylation prediction.

Main Methods:

• Training a HyenaDNA genome-wide foundation model for each of the six plant species.
• Fine-tuning the foundation models using sequence data around methylated and unmethylated cytosines.
• Utilizing long-range state-space models based on the Hyena architecture.

Main Results:

• The HyenaDNA framework achieved high accuracy in predicting DNA methylation across six plant species.
• Fine-tuning improved the model's performance on specific cytosine methylation prediction tasks.
• The developed model outperformed existing state-of-the-art methods in DNA methylation prediction accuracy.

Conclusions:

• Long-range state-space models, exemplified by HyenaDNA, show significant promise for accurate DNA methylation prediction in plants.
• The HyenaDNA framework provides a robust tool for imputing DNA methylation status, especially in regions with insufficient sequencing coverage.
• This approach advances the field of plant epigenomics and computational biology.