2026-08-14
Deep learning on 25,713 GTEx slides builds 40 tissue clocks at 4.88-year mean MAE; blood gene expression then predicts organ age gaps, pinpointing affected organs across 8 diseases.
Biological age has been measurable for over a decade through DNA methylation clocks, but those read blood cells and return a single whole-body number. What clinicians actually need is organ resolution: in the same 65-year-old, the heart might behave like 75 while the liver sits at 60. No dataset previously offered a view of every organ's tissue architecture from the same individuals, and animal models translate poorly to human aging.
GTEx fills the gap: 983 donors aged 20 to 70, rapid-autopsy collection of 40 tissue types across 29 organs, 25,713 whole-slide images, with paired RNA-seq, DNA methylation and telomere data. This Nature Medicine paper turns those slides into 40 tissue-specific aging clocks.
Three stages. Each slide is tiled into roughly 112 µm patches (about 160 million across the corpus), a vision model fine-tuned on a set balanced for tissue, sex and age bracket extracts features, and patch features are mean-aggregated and concatenated across three magnifications into one vector per slide. A Ridge regression is then trained per tissue to predict chronological age, with cross-validation grouped by individual and sex, cohort and ischemic time regressed out as covariates. Predicted age minus chronological age is the age gap; a positive gap means that organ is aging faster than the calendar. For interpretation, the pathology vision-language model PLIP scores each image against 150 histological terms, translating morphological change into readable words, cross-checked with a second model (CONCH) to rule out single-model bias.
One design choice deserves mention: all downstream results come from Ridge, a regularized linear model. On independent cohorts the authors found neural-network regressors generalized substantially worse, while regularized linear models and tree ensembles held up best.
| Metric | Result | Comparison |
| Mean MAE across tissues | 4.88 years (R²=0.69) | ImageNet models 8.67; 18 pathology foundation models 5.74 |
| External-cohort Pearson | lung 0.76, brain 0.56, skin 0.46 | skin cohort n=185; all calibration slopes below 1 |
| Correlation with methylation age gap | 0.09 within GTEx | 0.30–0.47 in lung cohort with matched tissue blocks |
| Comorbidity association rate | histology significant in 89% (colon) and 79% (lung) of models | methylation clocks 7% and 50% |
Morphologically, atrophy, microvascular rarefaction and fibrosis rise across most organs while epithelial hyperplasia declines. Transcriptomically, biological age perturbs gene expression more than chronological age (average slope 1.26), with p53, senescence, SASP and hypoxia pathways upregulated more strongly. Individuals split into resilient, average, single-organ and systemic agers; renal failure associates with accelerated aging across many tissues, yet most strongly not in the kidney but in adipose tissue, pituitary, spleen and tibial nerve.
The clinical payoff: models trained on GTEx paired blood expression predict tissue-specific age gaps from peripheral blood alone, validated on 1,205 samples across nine independent ARCHS4 cohorts (577 healthy, 628 with chronic disease). Stroke samples show elevated gaps only in brain, Alzheimer's likewise hits only brain, Crohn's disease marks the entire gastrointestinal tract, and vasculitis marks kidney, liver and heart. Classification performance is moderate to strong by AUROC, with positive predictive values above 0.3 in several settings.
This is the first body-wide map of organ-level aging. For ML practitioners it lays out a deployable recipe: foundation-model features plus linear regression beat end-to-end deep regressors under domain shift, a lesson that transfers to any medical imaging task. The blood-based part swaps invasive biopsy for a blood draw; if prospective validation holds, it is a prototype for population-scale screening.
The authors list three: GTEx is two-thirds male, and 3 of the 4 underperforming clocks come from female reproductive tissues; samples are postmortem, so autolysis, however adjusted for, may reduce discovery power; cross-sectional data yields correlations, not causes. The blood predictor has a deeper problem: external cohorts lack paired tissue histology, so its absolute calibration cannot be verified, and validation used people with prevalent disease, leaving open whether elevated gaps precede onset. GTEx donors are also mostly healthy, limiting the dynamic range of age gaps, and part of the 0.09 correlation with methylation clocks may be technical, as the two data types did not always come from the same tissue blocks.