MethylGPT: a foundation model for the DNA methylome.

Area of Science:

• Epigenetics and Computational Biology
• Artificial Intelligence in Genomics
• Biomarker Discovery

Background:

• DNA methylation is a key biomarker for disease diagnosis and biological age.
• Linear models struggle to capture the complexity of methylation regulation.
• Need for advanced analytical methods to interpret epigenetic data.

Purpose of the Study:

• To introduce MethylGPT, a transformer-based foundation model for DNA methylation analysis.
• To demonstrate MethylGPT's ability to predict methylation values and capture biological context.
• To evaluate MethylGPT's performance in age, mortality, and disease prediction.

Main Methods:

• Trained MethylGPT on 154,063 human methylation profiles across diverse tissues.
• Utilized a transformer architecture trained on 7.6 billion tokens and 49,156 CpG sites.
• Applied MethylGPT to methylation value prediction, age prediction, and disease risk assessment.

Main Results:

• Achieved high accuracy in methylation value prediction (Pearson R=0.929).
• Demonstrated superior performance in age prediction compared to existing methods.
• Showed robust predictive performance in mortality and disease prediction across 60 conditions.

Conclusions:

• Transformer architectures can effectively model DNA methylation patterns with biological interpretability.
• MethylGPT shows significant potential for epigenetic analysis and clinical applications.
• The model enables systematic evaluation of intervention effects on disease risks.