Tissue Clocks

A person's organs do not age at the same rate. An aorta can thicken and calcify while the skin of the same donor still looks a decade younger, and the record of that divergence is already sitting in the tissue: the H&E section a pathologist reads carries the myelin loss, atrophy, microvascular rarefaction and fibrosis that accumulate over a lifetime. Aging clocks have overwhelmingly been molecular — a DNA methylation array read from one sample — while the morphological evidence sat unquantified in archives holding whole-slide images by the million.

Tissue clocks are a deliberately thin recipe for turning that morphology into a number. A whole-slide image is tiled into 224-pixel patches, each patch is embedded by a frozen pathology vision encoder, the tile embeddings are mean-pooled into a single slide-level feature vector, and a regularized linear regression fit separately for each tissue predicts the donor's chronological age from it. The quantity of interest is the residual — predicted minus chronological age, the age gap — because that is what tracks pathology rather than the calendar.

The clocks were built by the Rendeiro lab at CeMM with collaborators across Austria, Germany, Belgium and Australia, from 25,712 whole-slide images spanning 40 tissues in 29 organs from 983 GTEx donors. A parallel set of predictors maps bulk blood gene expression onto the same histology-derived age gaps, so an organ's morphological age can be estimated from a blood draw.

#Key Features

• One clock per tissue: Roughly 40 tissue-specific regressors give an independent age gap for each organ sampled, distinguishing systemic aging from single-organ aging within one donor.
• Encoder-agnostic: The same regression head works over 18 pathology foundation models as well as the authors' fine-tuned encoder, making the clock a portable recipe rather than one monolithic checkpoint.
• Transfers without refitting: Coefficients trained on GTEx are loaded and applied to external cohorts imaged at other institutions on other scanners; target ages are used only to score predictions.
• Blood-based organ readout: Coefficients learned on matched GTEx blood transcriptomes predict tissue-specific age gaps from blood alone.
• Interpretable through language: Slide embeddings are queried against 150 histological and pathological text terms with PLIP and CONCH, naming atrophy, fibrosis and microvascular rarefaction as the recurring aging features.

#Technical Details

The headline clocks use a ConvNeXt-Base encoder fine-tuned on a tissue-classification task over a GTEx tile set balanced for tissue, sex and age bracket, with features extracted at three tile widths and concatenated. Age is fit with Ridge regression under five-fold GroupKFold grouped on donor, with sex, cohort and ischemic time as covariates. Across all tissues this gives a mean absolute error of 4.88 years and an R² of 0.69; substituting 18 pathology foundation models — among them UNI, Prov-GigaPath, H-optimus-1, TITAN and Midnight — gives a mean MAE of 5.74, while seven ImageNet-era vision models reach only 8.67. On external cohorts of brain (n = 70), lung (n = 40) and skin (n = 185), GTEx-trained clocks reach Pearson correlations of 0.56, 0.76 and 0.46, with calibration slopes below 1.0 reflecting compression under domain shift. Age gaps correlate with telomere attrition across 6,197 samples and with annotated subclinical pathology, and recover transcriptional aging signatures more strongly than chronological age. The blood arm was validated on 1,205 samples across nine cohorts covering eight diseases. One deployment detail matters: the released code re-estimates feature standardization per target cohort, so clocks apply to a batch of slides rather than one slide in isolation.

#Applications

Because the input is a routine stained section, the clocks run on material that already exists — biobank archives, diagnostic blocks, tissue taken during colonoscopy or biopsy — without a new assay. Aging researchers get a per-organ phenotype to associate with lifestyle, comorbidity and demographic variables; pathologists get an age-referenced measure of structural decline that flags organs deviating from a donor's chronological age. The blood arm is the more clinically tractable half, offering organ-level aging estimates from a minimally invasive sample for cohort screening and for stratifying participants in intervention studies.

#Impact

Tissue clocks establish tissue architecture as a measurable aging readout alongside DNA methylation, and the two agree only loosely, suggesting they capture partly distinct dimensions of aging. Several caveats bound the result. Four of the 40 clocks are underpowered with fewer than 100 samples, three of them female reproductive tissues in a cohort skewed two-to-one male; GTEx donors span only ages 20 to 70, limiting age-gap dynamic range; and the validation is cross-sectional, so whether an elevated age gap precedes disease remains untested. The analysis code is public under a noncommercial license, with the fine-tuning image crops and extracted GTEx feature matrices, but the fitted clock coefficients and encoder weights have not been deposited — reproducing a clock means re-running the pipeline rather than loading a checkpoint.