Knowledge reporting
No statistical strategies had been used to predetermine pattern measurement as a result of predetermined pattern availability of the GTEx venture27. Three samples had been excluded from the examine as they had been annotated with the flawed tissue label.
Knowledge acquisition
GTEx WSIs had been downloaded from the general public portal27. All slides had been acquired with a Leica Biosystems Aperio ScanScope with a ×20 goal (efficient decision of ~0.4942 µm per pixel). Demographic information (together with the precise chronological age of the people), way of life info, medical info for the individuals and gene expression information had been downloaded from dbGaP as a result of they’re beneath protected entry.
Impartial histopathological picture datasets had been used to validate the tissue clocks. Postmortem brains of people aged 35 to 99 years had been recognized from the archives of the Institute of Neuropathology and Neuromolecular Pathology on the Medical College of Innsbruck. Moral approval for the usage of formalin-fixed and paraffin-embedded (FFPE) tissues for analysis research, together with a waiver of the requirement for knowledgeable consent, was granted by the moral committee of the Medical College of Innsbruck (EK 1387/2025). FFPE tissues of frontal and cerebellar cortex had been retrieved, and hematoxylin and eosin-stained sections had been assessed for central nervous system pathology. The sections had been then digitized in ×20 magnification utilizing an Aperio GT 450 DX automated digital pathology slide scanner, and whole-slide scans had been saved in .svs format. Pores and skin samples had been obtained on the Division of Plastic and Reconstructive Surgical procedure or the Division of Dermatology of the Medical College of Vienna. The examine complied with nationwide legislation and was permitted by the Medical College of Vienna Ethics Committee (ECS 1969/2021 and ECS 1877/2024), with knowledgeable consent obtained from individuals. Pores and skin biopsies had been mounted in formalin, embedded in paraffin and saved at −20 °C to protect tissue construction and RNA for morphological and molecular analyses. For histology, FFPE blocks had been sectioned into 6-µm slices utilizing a microtome (Leica RM2235) onto SuperFrost Plus slides (VWR, 6310108) and stained with hematoxylin and eosin utilizing a Dako Cowl Stainer with commonplace reagents (Agilent). Slides had been digitized in ×40 magnification utilizing a Hamamatsu C9600-12 automated digital pathology slide scanner, and whole-slide scans had been saved in .ndpi format. Lung samples had been obtained beneath approval by the ethics committee of UZ Leuven/KU Leuven (S52174), with signed knowledgeable consent from individuals. The examine was carried out in accordance with the Declaration of Helsinki. Complete lungs had been excised, air inflated and snap frozen, with all samples derived from the identical anatomical area. Slides had been scanned on an Axioscan 7 with an EC Plan-Neofluar ×20/0.50-NA M27 goal (efficient decision of ~0.1725 µm per pixel) and saved in .czi format.
WSI processing
WSIs had been segmented into foreground tissue and background areas utilizing laptop imaginative and prescient algorithms tailored from the preprocessing utilities of the CLAM repository67. Solely the picture preprocessing parts had been reused; the CLAM multiple-instance studying and attention-based fashions weren’t used. We repurposed and modularized the preprocessing code right into a general-purpose WSI library that’s mannequin agnostic and now open sourced at https://github.com/rendeirolab/wsi. Briefly, sign from a WSI thumbnail was transformed to HED coloration house, Otsu-thresholded and dilated, and small objects (<500 pixels) had been eliminated. To provide a tissue masks, the holes had been stuffed, and a separate masks for the holes (>0.5 and <50 pixels) was used. Contours had been saved in .h5 format, and slides had been tiled within the tissue space with sq. patches of 224 pixels (efficient measurement of ~112 µm).
Imaginative and prescient mannequin fine-tuning
To seize histological options higher, we fine-tuned quite a lot of imaginative and prescient mannequin architectures in a tissue classification process. Importantly, we created a balanced dataset comprising equal numbers of slides and tile patches from completely different tissues, age brackets and sexes. We constructed upon well-established and pretrained fashions (AlexNet, VGG16, ResNet50, ResNet152, ResNet50, ConvNeXt tiny and ConvNeXt base), with a classifier head akin to the tissue lessons. To coach the brand new classifier head layer, we first froze all layers of the pretrained mannequin and educated one epoch, adopted by as much as 100 epochs of fine-tuning the entire mannequin by way of the AdamW optimizer with cosine annealing. We began with an empirically found studying fee by measuring the loss at incremental steps and selecting one-tenth of the minimal (learner.lr_find). Mannequin fine-tuning was carried out with PyTorch and Quick.ai, both on TPUs via Google Colab or on a workstation with two NVIDIA RTX A6000 GPUs.
Characteristic extraction and unsupervised evaluation
Inference was run on a high-performance computing cluster utilizing CPUs for all tiles in all slides at three ranges of magnification (~480 million tiles). We labored with three completely different tile widths (224, 448 and 896 pixels at ~0.5 µm per pixel) however centered on the identical location, such that the tile centroids throughout completely different widths had been matched and represented completely different, progressively wider views of the identical tissue location. Options had been extracted with quite a lot of fashions as described earlier, and for downstream evaluation, characteristic values had been aggregated throughout tiles by means and concatenated throughout completely different tile widths to generate a single characteristic vector for every WSI. We used AnnData and Scanpy68 to carry out dimensionality discount with principal part evaluation (PCA), compute neighbors and compute a UMAP for visualization, every step with default parameters (PCA: 50 principal parts, neighbor graph: 15 neighbors on the premise of PCA illustration, UMAP: 0.5 min_dist, 1.0 unfold).
Evaluation of whole variance defined
To offer a world measure of the significance of varied elements in explaining the noticed vary of morphological heterogeneity throughout tissue samples, we carried out linear fashions that use these elements to foretell the variability of samples within the principal parts. We collected 86 variables associated to demography, way of life, serology, morbidity and circumstances of demise for every participant. Regression was primarily based on an abnormal least squares regression mannequin (statsmodels69) and was match for every tissue and principal part at a time. The total mannequin was match first, and fashions lacking every of the 86 variables (go away one out) had been subsequently match and in comparison with the complete mannequin on the premise of the adjusted coefficient of dedication. These values had been weighted by the variance ratio of every principal part within the information and summed throughout parts to generate a worth of whole variance defined for the issue.
Tissue clocks and estimation of age gaps
We used linear fashions carried out in scikit-learn70, together with Linear Regression, Ridge, RidgeCV and RandomForestRegressor, utilizing GroupKFold (ok = 5) cross-validation, the place the group was a person, to make sure that the coaching and validation splits didn’t include the identical particular person. Fashions had been educated for every tissue individually utilizing the options from the imaginative and prescient fashions as inputs, age because the goal and intercourse, cohort kind and minutes of ischemic time as covariates. To evaluate potential overfitting, we additionally fitted the fashions on shuffled age labels as goal. We computed the coefficient of dedication and MAE as metrics. Adjustment for regression to the imply was carried out by regressing out the coefficient of age from the noticed residuals as beforehand described16. All the outcomes reported downstream had been obtained for tissue clocks educated with a Ridge mannequin, because it exhibited good efficiency and was quick.
The ‘Bladder’, ‘Cervix – Ectocervix’, ‘Cervix – Endocervix’, ‘Fallopian Tube’ and ‘Kidney – Medulla’ tissues had fewer than 100 samples accessible and total confirmed poor efficiency (for instance, MAE > 9 years) and had been excluded from additional evaluation. We additionally seen that three samples had excessive age-gap values (‘GTEX-1GMR2-0426’, ‘GTEX-1S82U-0426’ and ‘GTEX-11ZTS-0426’), which, upon inspection, had been discovered to be annotated with the flawed tissue kind.
Traditional imaginative and prescient and basis fashions
To validate and assess the generalization of findings with our fine-tuned mannequin, we used 6 traditional imaginative and prescient fashions educated on ImageNet (Supplementary Fig. 5) and 18 pathology basis fashions (Prolonged Knowledge Fig. 3)21,22,35,36,71,72,73,74,75,76,77,78,79,80,81,82,83,84,85,86,87,88,89,90,91,92,93,94. We carried out characteristic extraction utilizing LazySlide95 on NVIDIA H100 GPUs, utilizing sq. patches of 224 pixels and aggregating options inside every slide by the imply (as achieved beforehand for our fine-tuned mannequin). Age predictors had been fitted as described within the part above utilizing the Ridge mannequin with GroupKFold (ok = 5) cross-validation, and metrics had been calculated primarily based on the chronological age of the people.
GNNs
To additional validate the selection of characteristic aggregation throughout patches of the identical slide, and to allow the visualization of weights related to organic age on tissue photographs, we additionally used GNNs for organic age prediction from histopathological photographs (Supplementary Fig. 7). Right here, tissue patches had been used as nodes with options from the convNeXt fine-tuned mannequin. A ok-d tree was used for fast nearest-neighbor lookup with a radius of patch measurement × the sq. root of two to attach adjoining tissue tiles.
The mannequin structure was composed of a dropout layer (with a chance of 0.1, 0.25 or 0.5), a variable quantity (2, 4 or 8) of graph convolutional layers (torch_geometric.nn.conv.GCNConv) with batch normalization and ReLU activation. The convolutional layers had a variable variety of hidden dimensions (32, 64, 128 or 256) and had been aggregated by way of leaping data (torch_geometric.nn.fashions.JumpingKnowledge) throughout all layers, upon which an consideration aggregation layer (torch_geometric.nn.aggr.AttentionalAggregation) with sigmoid activation was added. Lastly, a completely linked linear layer was added to output a scalar worth.
The goal worth was chronological age, and GNNs had been educated per organ for 80 epochs on an NVIDIA A6000 utilizing an 80–20 dataset cut up, MAE loss, the AdamW optimizer with cosine annealing, with a beginning studying fee of 1 × 10−3, and weight decay of 1 × 10−4. We assessed efficiency within the validation set utilizing MAE, Pearson correlation and coefficient of dedication in relation to the identified chronological age of the donors.
To visualise consideration weights on tissue (Supplementary Fig. 9d), we extracted node-level consideration scores from the best-trained GNNs for every tissue. For every patch (node), we obtained the uncooked gate outputs and normalized them into per-node consideration weights utilizing scatter-based softmax over the batch dimension. The outputs had been aligned again to the spatial positions of the unique patches, and the normalized consideration weights had been overlaid on the corresponding histological tissue photographs as warmth maps. This enabled spatial visualization of areas that contributed most strongly to the graph-level age prediction.
Telomere lengths and annotated ranges of tissue pathology
We leveraged the telomere amount index values measured for tissue blocks from the identical tissues and people within the GTEx cohort34. We z scored these values per tissue to account for intertissue variability within the imply values. Solely tissues with at the very least 100 paired WSIs and telomere size samples had been used for evaluation.
We additionally leveraged the pathological notes accessible within the GTEx cohort, particularly the discretized set of 57 textual content classes describing the WSIs, which comprised 11,016 situations of a time period annotating a picture. For the aim of associating explicit pathological annotations with histological age gaps, we excluded ‘clean_specimens’, ‘no_abnormalities’ and ‘tma’. For pathologies widespread throughout multiple tissue, we additionally derived an total ‘physique pathology burden’, which is a composite measure of the variety of pathologies annotated in all tissues of a person.
Textual content time period characterization of age gaps
We used the pathology language and picture pretraining (PLIP) mannequin35 to embed 512-pixel-wide (at ~0.5 µm per pixel) tiles from each slide within the GTEx venture and used them to question the similarity to a set of histological and pathological phrases. The phrases had been chosen to signify equal elements of the histological parts and options of the varied tissues beneath examine (for instance, epithelium, muscle, neurons, adipose tissue and myofiber degeneration). The values had been aggregated by means per slide, and associations with histological age gaps had been derived by way of regularized linear regression (Ridge).
To evaluate the settlement between vision-language fashions of their interpretation of histopathological photographs of ageing, we in contrast the outputs of PLIP and CONCH utilizing slides from the colon, lung and pores and skin tissues within the GTEx dataset (n = 4,368). For every slide, we computed the cosine similarity of picture options to a curated set of 150 histological and pathological textual content phrases utilizing each fashions. The ensuing slide-level similarity matrices had been used to compute, for every textual content time period, the Pearson correlation coefficient between the PLIP- and CONCH-derived similarity values, yielding a distribution of mannequin settlement throughout phrases. To establish elements underlying this settlement, we extracted a set of text-level options describing the linguistic construction of every time period, together with character and phrase composition (for instance, variety of characters, phrase size, punctuation, uppercase and lowercase utilization, lexical range and character entropy). As well as, we included the imply and commonplace deviation of the PLIP and CONCH similarity values for every time period. These options had been standardized and used as predictors in a linear regression mannequin explaining the correlation between fashions. Lastly, we quantified, for every tissue, the affiliation of change in text-term similarity with donor age and the settlement of phrases between the vision-language fashions.
DNA methylation clocks
We used the PyAging package deal96 to foretell the organic age of the DNA methylation samples matched to the histological photographs of the tissue. DNA methylation information42 had been measured by way of the Infinium HumanMethylationEPIC bead chip, which measures roughly 850,000 CpG websites. We then derived age gaps by evaluating them to the actual chronological age of the samples and in contrast each the expected values and the DNA methylation-derived age-gap values to the histological predictions utilizing Pearson correlation coefficient.
Gene expression evaluation
We used bulk gene expression information from the GTEx venture that had been matched to the identical tissues because the histopathology photographs. We transformed gene-level counts to log counts per million (log (CPM)), and any technical replicates had been aggregated by means. To elucidate each ageing and tissue-specific charges of age acceleration (age gaps) with gene expression of the corresponding tissues, we carried out Ridge regression utilizing, as beforehand described, intercourse, cohort and minutes of ischemic time as covariates. Genes with an absolute coefficient worth above 0.005 (> 5% change per decade) had been chosen. Lastly, we used Enrichr97 via the GSEApy package deal98 for gene set enrichment with the MSigDB database individually for the up- and downregulated gene units.
Equally, we additionally used the 50 MSigDB pathways along with a set of fifty age-related signatures beforehand compiled (accessible at https://github.com/WJPina/HUSI/), in addition to a set of genes associated to extracellular matrix biology, and quantified them within the transcriptomes utilizing the Scanpy operate sc.tl.score_genes. This allowed us to acquire a decreased set of broadly and particularly related transcriptional options for every pattern. We then calculated their affiliation with chronological age or histologically derived organic age per organ and in contrast the coefficients between the 2 per organ, in addition to the imply throughout organs, and assessed whether or not their signed deviation differed utilizing a Wilcoxon check.
Affiliation of tissue-specific age gaps with particular person elements
To visualise people in teams with differential patterns of organic ageing throughout the varied tissues profiled per particular person (Fig. 4b), we chosen the tissues with abnormally massive age gaps (commonplace deviation of ≥3), set the remaining tissues to 0 and carried out PCA, computed neighbors and carried out a UMAP as described beforehand.
To statistically affiliate particular person age gaps with individual-specific elements, we fitted regularized linear fashions (Ridge) per tissue and intercourse with elements explaining tissue-specific age gaps. To keep away from collinearity, we excluded the next variables: ‘Irregular WBC’, ‘Medicine For Non Medical Use In 5 y’, ‘HBcAb IgM’, ‘HIV 1 NAT’, ‘HIV I II Ab’, ‘HIV I II Plus O Antibody’, ‘Nephritis, Nephrotic Syndrome and/or Nephrosis’, ‘Night time Sweats’, ‘Open Wounds’, ‘Acquired Human Progress Hormone’ and ‘Tattoos Executed In 12 m’. Moreover, we retained solely elements with greater than three people per tissue and intercourse.
Tissue clock software to exterior cohorts
To validate and assess the generalization of the tissue clocks, we investigated extra cohorts for which histopathological photographs had been collected (from mind, lung or pores and skin tissue). In these cohorts, WSI processing was carried out in the identical method as within the GTEx cohort, however we ran each our personal fine-tuned mannequin in addition to foundational fashions for characteristic extraction and aggregated the options throughout patches of every picture by the imply. To establish whether or not the impact of ageing could possibly be detected in every cohort, we carried out PCA, in addition to fitted regression fashions utilizing the PCA house as a predictor of age to evaluate the variance defined by donor age.
We then utilized the educated tissue clocks from GTEx to the brand new cohorts, with out including covariates or adjustment for regression to the imply. We carried out this not just for fashions matched by tissue/organ (that’s, making use of the mannequin educated on GTEx pores and skin samples to the pores and skin cohort) but additionally throughout tissue/organ (that’s, making use of the mannequin educated on GTEx pores and skin samples to the mind cohort).
For the cohort of lung tissue donors, we additionally used DNA methylation information, which had been collected from the identical tissue blocks because the histopathological photographs. Three well-established DNA methylation clocks had been used (Horvath, Hannum and PhenoAge) to generate organic age and age-gap predictions. These DNA methylation predictions had been contrasted with the predictions from histological photographs to quantify the settlement between assays and organic age predictors.
To analyze how completely different regression approaches generalize within the process of cross-cohort prediction, we systematically evaluated a various set of regression fashions spanning linear, nonlinear and ensemble approaches. Linear fashions included Ridge (L2 regularization), Lasso (L1 regularization) and Elastic Internet (mixed L1/L2 penalties by way of ElasticNetCV), in addition to Bayesian Ridge regression for computerized regularization tuning. To account for potential outliers amongst donors, we examined strong regression approaches (Huber and Quantile regressors). Generalized linear fashions (Gamma and Tweedie regressors) had been included to explicitly mannequin the optimistic, probably skewed distribution of age. We additional evaluated assist vector regression (SVR and LinearSVR), random forest and gradient-boosted timber (LightGBM). Lastly, we examined multilayer perceptron neural networks in three configurations: commonplace, with L2 regularization and early stopping and with a customized L1 regularization to advertise sparse weight matrices, additionally with early stopping. All fashions had been in contrast on their skill to generalize age predictions to impartial held-out cohorts by way of MAE to chronological age.
Settlement and calibration evaluation
To quantify the efficiency of the GTEx-trained tissue clocks on exterior cohorts, we used Bland–Altman evaluation to evaluate settlement between predicted organic age and chronological age. For every tissue-specific clock utilized to every exterior cohort, we computed (1) the bias, outlined because the imply distinction between predicted and chronological age, with 95% confidence interval; (2) the bounds of settlement, computed as bias ± 1.96 × the usual deviation of variations; and (3) the usual deviation of variations. Moreover, we assessed calibration by becoming a linear regression of predicted age on chronological age for every cohort and tissue, extracting the slope (ideally = 1.0) and intercept (ideally = 0), each with 95% confidence intervals, together with the coefficient of dedication (R2). Outcomes had been visualized utilizing Bland–Altman plots exhibiting the imply of predicted and chronological ages (x axis) versus their distinction (y axis), with reference strains for bias and limits of settlement.
Prediction of age gaps from blood gene expression
We used bulk gene expression information from the GTEx venture that had been matched to the identical tissues because the histopathology photographs. Blood gene expression profiles had been log reworked and transformed to CPM (log (CPM)). We filtered genes with low expression by eradicating these with a imply log(CPM) of <1 (leaving 11,859 genes) and aggregated the technical replicates by means. As a result of we noticed an impact of the ageing course of on gene expression usually, we opted to first regress out the impact of age from the blood gene expression profiles utilizing regression. We then fitted regularized regression fashions to foretell tissue-specific age gaps from blood gene expression utilizing Ridge fashions with two cross-validation loops (RidgeCV): one exterior Kfold (ok = 5) that separates completely different teams of people and the place commonplace scaling is utilized and one inner Kfold (ok = 5) for α-hyperparameter optimization (chosen from 10−1 to 107 in 20 uniform steps). This course of was carried out independently for every tissue, and for the imply age hole throughout the tissues of a person (systemic clock).
For validation, we obtained a complete set of bulk gene expression profiles from samples of peripheral blood mononuclear cells, leveraging the ARCHS4 database52, model 2.2. Gene expression counts had been log reworked, normalized in relation to the overall, standardized and scaled for every gene, and the coefficients derived for every predictor within the GTEx cohort had been utilized in a linear mannequin to make age hole predictions for the majority RNA-seq samples. The first metric chosen to judge efficiency between every illness and wholesome samples was a two-tailed t-test with Benjamini–Hochberg false discovery fee correction. This was used because of the truth that our regression fashions output steady values (age hole), that are Gaussian and centered at 0, as these had been realized from the residuals of our tissue-clock fashions. We additionally evaluated the utility of the blood-based predictors in a binary setting by thresholding the age gaps at 1 and calculating the realm beneath the receiver operator curve and the optimistic predictive worth/precision for separating wholesome people from these with a illness.
Reporting abstract
Additional info on analysis design is obtainable within the Nature Portfolio Reporting Summary linked to this text.
