BioData Mining · Nov 19, 2026
Methylation age-associated signature remodelling based on vae-diffusion model: algorithm development and multi-dimensional validation
By Wang T, Wang J, Wang T et al.
Background: Whether age-associated methylation changes can be remodelled while preserving individual identity remains a key challenge. Methods: We applied a conditional diffusion model for methylome age-associated signature remodelling. A VAE compressed 15,000 CpGs into 64 latent dimensions, followed by diffusion training under cell-type constraints. Ten-fold cross-validation compared six methods using identity correlation, Clock MAE, and cell-type correlation. Results: Our method significantly outperformed linear interpolation, nearest neighbor, and PCA regression in identity preservation and Clock MAE ( p < 0.001, d = 4.04–5.68), matching cVAE and DDPM. Fifteen driver dimensions were identified. Driver CpGs significantly enriched at POLR2A binding sites (fold = 3.13, p = 0.0044). Metformin-induced changes correlated with remodelling direction ( r = 0.518, p = 0.048). Six CMap-predicted drugs overlapped with DrugAge, including canagliflozin (9.0% lifespan extension in mice). Saliva validation supported cross-tissue generalizability ( r =-0.406, p = 1.47 × 10⁻⁶). Conclusions: Age-associated methylation signature remodelling proceeds via targeted reprogramming along 15 driver dimensions, preferentially modulating POLR2A-bound regions while preserving individual identity, providing a framework for epigenomic studies of aging. Clinical trial number: Not applicable.