Fig. 4. MitraCluster generalises across body sites, ectodermal tissues and captures rejuvenation.

Spatial generalisability of MitraCluster was assessed using matched tape-strip samples from the face and upper back of 85 individuals. 61 out of 85 of these pairs shown here in teal have had the facial skin sample included in the training of the model, while the remaining 24 indicated in orange had not. MitraCluster accurately predicted age in upper back skin (all samples: R² = 0.75, MAE = 4.90; teal only: n = 61, R² = 0.71, MAE = 5.19; orange only:, n = 24, R² = 0.84, MAE = 4.13) (a), with high concordance between predictions from both body sites (all samples: R² = 0.71, MAE = 5.53; teal only: n = 61, R² = 0.75, MAE = 4.93; orange only:, n = 24, R² = 0.61, MAE = 7.04) (b), demonstrating reliable intra-individual consistency across sampling locations. Cross-tissue performance of MitraCluster (top row) and BS-Clock (bottom row) was evaluated using publicly available bisulfite sequencing datasets from lung (c, d), blood (e, f), and brain (g, h) tissue. For the BS-clock, the predictions presented correspond to the version specific to the corresponding tissue shown. BS-Clock performed well in blood and lung (R² = 0.90 and 0.79; MAEs = 3.34 and 3.78), whereas MitraCluster underperformed in these tissues (R² ≤ 0.05; MAEs > 50), consistent with its epidermis-specific training. However, in brain tissue, MitraCluster achieved moderate predictive performance (g MAE = 15.00, R² = 0.16), approaching that of the brain-specific BS-Clock (h MAE = 10.86, R² = 0.61). 59.1% of MitraCluster CpGs were covered on average by Blood samples, 75.2% by Lung and 83.7% by Brain samples. i Plot of age predictions from in vitro grown Normal Human Epidermal Keratinocytes under control conditions, or after two types of Yamanaka factor treatment supplementation that generate keratinocyte-derived iPSCs.