Exploring neural network architectures for DNA methylation-based biological age prediction — comparing MLP, CNN, and Self-Attention against the classic Horvath linear model.
~7,900 samples, 142 datasets
20,318 CpG site beta values
5 architecture variants
MAE / MSE evaluation
| Rank | Model | Architecture | Result |
|---|---|---|---|
| 1 | MLP | 4 layers + batch norm + dropout | Best MAE |
| 2 | CNN | 2 conv + pooling + 2 linear | Improved over baseline |
| 3 | Horvath (baseline) | Elastic net, 353 CpGs | Classic benchmark |
| 4 | Linear (353 CpGs) | PyTorch linear regression | Below baseline |
| 5 | Self-Attention | 1 pre-linear + 3 post-linear | Degraded |
| 6 | Linear (20K CpGs) | PyTorch linear, all features | Worst (overfit) |
A 4-layer MLP with batch normalization and dropout outperformed all other architectures including the classic Horvath elastic net model.
Convolutional layers captured local patterns in CpG site sequences, beating the linear baseline while staying computationally efficient.
Self-Attention degraded performance — the sequential nature of methylation site data does not benefit from global attention mechanisms.
Using all 20,318 CpG sites with a simple linear model led to severe overfitting. Feature selection (353 CpGs) outperformed the full set.
Training on diverse tissue types was critical. Homogeneous datasets failed to generalize across the 142-dataset evaluation split.
Cancer patient samples consistently showed higher predicted age than chronological age, validating known age acceleration in tumors.