Epigenetic Clock

Exploring neural architectures for DNA methylation-based age prediction.

Slides

Investigated neural network architectures for epigenetic clock models, predicting biological age from DNA methylation data (~7,900 samples across 142 datasets, 20,318 CpG sites).

Key findings:

  • Compared 5 architectures: Linear Regression, MLP, CNN, Self-Attention, and the classic Horvath elastic net baseline
  • MLP (4 layers + batch norm + dropout) achieved the best MAE, outperforming all other models
  • CNN improved over the Horvath baseline by capturing local patterns in CpG site sequences
  • Self-Attention degraded performance — methylation site data does not benefit from global attention mechanisms
  • More features did not help: using all 20,318 CpGs with linear regression led to overfitting; Horvath’s 353-CpG selection outperformed the full set

Status: Project paused — biological age predictions lack reliable ground truth for validation, making it fundamentally difficult to distinguish model improvement from overfitting to chronological age.

Technologies: PyTorch, Python