AI Models Predict Biological Age of Organs from Tissue Images with 4.9-Year Accuracy

Histological aging signatures for monitoring tissue-specific aging and disease - Nature Medicine

To address these challenges, it is imperative to move beyond a gene or cell focus, to a holistic view of the aging organism19. Deep learning20 has greatly advanced our ability to quantify and understand tissue morphological and architectural patterns21,22 and detect and classify diseases23,24. Here, we use a systematic approach to probe the interplay between aging and tissue architecture in humans and its connection to pathology, illuminating the complex relationship between tissue-specific aging patterns and their implications across the human body. Aging strongly affects healthy tissue architecture in humans To study the impact of aging on tissue structure in humans, we used large-scale whole-slide histopathological images (WSIs) from 983 individuals in the Genotype-Tissue Expression (GTEx) project (Fig. 1a). The individuals profiled in this cohort died primarily from accidents, suicide or natural death, with ages ranging from 20 to 70 years (mean of 52.77 years; Supplementary Fig. 1a and Supplementary Table 1). Tissue was collected under a rapid autopsy protocol from a total of 40 different tissues in 29 organs, making up 25,713 WSIs (Supplementary Fig. 1b-e), which have been reviewed by pathologists and include a description of subclinical levels of pathology (Supplementary Fig. 1f-h), with demographic, lifestyle and clinical information also available (Supplementary Fig. 1i). Modern artificial intelligence vision models are powerful extractors of deep features from images (Fig. 1b) that can then be linked to information such as clinical outcomes. For our aim, to numerically quantify the morphological features in these WSIs, we first fine-tuned vision models on a set balanced for tissue, sex and age brackets (Extended Data Fig. 1a, b) and then used them to extract a feature vector representing the morphological features of the WSI (Extended Data Fig. 1c). We observed that these features were well reflective of WSI tissue identity even in pretrained models without fine-tuning (Extended Data Fig. 1c), but fine-tuning provided better separation (Extended Data Fig. 1d). More importantly, we found that deep learning morphological features also provided information on individual-level covariates, such as age, even though they were not fine-tuned for that task (Fig. 1c and Extended Data Fig. 1d-h). Indeed, we found that age was the strongest factor driving variation across tissues (Supplementary Fig. 2), highlighting the influence of aging on tissue structure throughout the human lifespan. Tissue clocks predict biological, pathological and clinical factors associated with age Building on the observation that age has a marked imprint on tissue structure, we sought to develop 'tissue clocks', which are tissue image-based predictors of biological age, building on the concept first introduced with DNA methylation. To accomplish this, we trained regression models to predict the chronological age of each individual from the morphological features of the WSI in a cross-validated manner and observed the deviation of the predicted age from the known chronological age, that is, an age gap (Fig. 1d). This gap, in DNA methylation data, has been associated with several age-related diseases, such as Alzheimer's, Parkinson's and Huntington's diseases, amyotrophic lateral sclerosis and all-cause mortality. We used the model predictions to assess their performance and found an average mean absolute error (MAE) of 4.88 years across all tissues and a coefficient of determination of 0.69 (Extended Data Fig. 2 and Supplementary Table 2). We also assessed the model robustness by shuffling the ages across individuals and observed no evidence of overfitting; however, 4 of 40 clocks were underpowered, as fewer than 100 samples were available for these tissues (Extended Data Fig. 2a-c and Supplementary Table 2). To understand the biological relevance of the tissue-based prediction of age and the value of the inferred age gaps, we leveraged telomere length measurements from the same tissues and individuals, available for 6,197 samples (25.2%). We observed that, as with chronological age, there was an overall decrease in telomere length with biological age (Supplementary Fig. 3a,b). However, this trend was stronger in terms of biological age, as determined through tissue analysis, than for chronological age and was still present in the estimated age gaps (Supplementary Fig. 3a-e). This correlation was particularly pronounced for tissues such as the esophagus, stomach, kidney, prostate and pancreas (Supplementary Fig. 3c,d). Furthermore, by leveraging the annotations of subclinical levels of pathology from the same samples, we observed an enrichment of greater age gaps in samples with more pathology, which was concordant across most pathologies and, importantly, was greater in tissue-driven estimation of biological age than in chronological age alone (Supplementary Fig. 4a,b); for example, both were enriched in artery calcification. However, age gaps also recovered enrichment in pathological features not particularly present in chronological age, such as muscle atrophy (Supplementary Fig. 4c,d). Confirming the observations that higher age gaps are associated with tissue-specific pathology, we visually confirmed that WSIs with higher age gaps acquired noticeable changes in tissue-specific pathology (Fig. 1e,f). For example, independent of chronological age, cerebellar samples with greater age gaps presented discoloration associated with the loss of myelin and ischemic changes (Fig. 1e), whereas in the aorta, samples with higher age gaps presented thickening of the wall and disruption of integrity associated with atherosclerosis and other vascular disorders (Fig. 1f). To assess the generalization of our tissue-based predictors of biological age, we used 7 classical deep learning models and 18 foundation models for pathology in feature extraction, building predictors from them (Fig. 1g, Supplementary Fig. 5 and Extended Data Fig. 3). The classical models, which were never trained on tissue images, showed predictive power but were not on par with our model (Extended Data Fig. 3a; mean MAE across organs and models of 8.67), whereas the performance of foundation models was more comparable to our own model (Extended Data Fig. 3b; mean MAE across organs and models of 5.74). However, the age gaps inferred by all types of models also showed significant associations with shorter telomeres and increases in comorbidities (Extended Data Fig. 3b-d). We also showed that models with markedly different architectures, such as graph neural networks (GNNs) with attention mechanisms, can learn to predict biological age (Supplementary Fig. 6a), with the advantage that their attention maps can be used to further locate areas in tissue most associated with biological aging (Supplementary Fig. 6b). We noted that a subset of the foundation models were pretrained on datasets that include GTEx histopathological images, which could favor their internal performance. To address this, we validated the histological predictors in external cohorts of brain (n = 70), lung (n = 40) and skin (n = 185) donors (Supplementary Table 3) by applying tissue clocks trained on the GTEx cohort to their histopathological images. In these cohorts, GTEx clocks predicted biological age with specificity depending on the tissue they were trained in (Pearson correlation coefficients of 0.56, 0.76 and 0.46 for brain, lung and skin, respectively; Fig. 1h, Extended Data Fig. 4 and Supplementary Table 4). In these external cohorts, we also evaluated the ability of multiple regressors to generalize (Supplementary Fig. 7) and found that both regularized linear models, as well as ensemble methods, performed comparably, but were superior to deep learning or unregularized models. Finally, to assess whether the estimated age gaps were meaningful, we compared them to biological age and age gap estimates derived from DNA methylation in the lung cohort and observed good agreement (Pearson correlation coefficient from 0.3 to 0.47 depending on the DNA methylation clock; Extended Data Fig. 5). Histological and molecular interpretation of tissue aging To characterize the histopathological features associated with the aging of tissues more systematically, we leveraged a pretrained self-contrasting multimodal model trained on many pathological images and their natural language descriptions (Fig. 2a). With this model, we queried the similarity of each image to a set of histological and pathological terms (n = 150), generating a highly interpretable numerical vector to further describe the images. The histological text terms showed high organ- and tissue-specific enrichment, with terms such as 'adipose tissue' high in breast tissue and visceral and subcutaneous adipose tissue or 'neurons' in the brain (Fig. 2b and Extended Data Fig. 6a). We then characterized the association of each term with aging in each organ (Extended Data Fig. 6b,c) and found that terms such as atrophy, microvascular rarefaction and fibrosis increased across multiple organs, whereas the most common reduction in text terms across organs was related to the epithelium and hyperplasia (Fig. 2c). In addition to these common trends across organs, we also found some more specific trends in certain organs, such as fat infiltration in skeletal muscle (Fig. 2d), which we could confirm visually. Similarly, we also visually confirmed additional associations, such as microvascular rarefaction in the uterus (Fig. 2e) and peripheral nerve atrophy (Fig. 2f). Finally, we verified using another vision-language model that there was no systematic bias for terms that were changing with age to be specific to one model or the other (Supplementary Fig. 8). These findings support our deep learning-driven analysis of tissue and enable a histological interpretation of age-associated changes in healthy tissue at a large scale. To further interpret the observed morphological insights, we leveraged paired samples from the same tissues in the same donors in the GTEx project. First, we sought to understand the relationship between estimates of biological age derived from DNA methylation and those derived from histological images (tissue clocks; Extended Data Fig. 7). In the GTEx cohort, 987 samples from 8 tissues had available DNA methylation data, which we used to derive a range of biological age estimates via established biological age clocks (Extended Data Fig. 7a,b). In a systematic comparison of 28 DNA methylation clocks and 19 histological clocks on the same GTEx donors (Extended Data Fig. 8 and Supplementary Table 5), we found that neither modality was uniformly superior across biologically relevant outcomes. For comorbidity burden, histological age gaps showed more consistent associations (89% and 79% of models significant in the colon and lung, respectively) than DNA methylation clocks (7% and 50%), with greater directional consistency and narrower confidence intervals. For telomere length, DNA methylation clocks performed comparably or better (18% and 43% significant in the colon and lung, respectively) than histological clocks (0% and 37%). For tissue pathologies, both modalities performed comparably in the lung (47% versus 25%), whereas neither achieved robust associations in the colon. Histology- and DNA methylation-derived age gap estimates showed no direct agreement (Pearson correlation coefficient of 0.09; Extended Data Fig. 7e-g), consistent with the two modalities capturing partially overlapping but distinct dimensions of the aging process. Second, leveraging over 17,000 matched bulk RNA-sequencing (RNA-seq) samples, we investigated how gene expression changes with chronological age and histologically derived biological age (Fig. 3). We used regularized regression models to associate the expression of all genes with varying degrees of deviation from chronological age in the same samples (Fig. 3a). We observed that many more genes were associated with changes in chronological age than with age gaps (thousands versus hundreds; Supplementary Fig. 9c-e) but that the degree of association was not necessarily driven by the performance of the tissue clock or sample size (Supplementary Fig. 9c), suggesting that tissue-specific factors may also affect biological age. We found that genes associated with tissue-specific age gaps were overall also tissue-specific (Fig. 3b and Supplementary Fig. 9f) but also included genes not usually expressed in the tissue. For example, we detected the upregulation of EYA transcriptional coactivator and phosphatase 4 (EYA4) in adipose tissue during aging (Fig. 3c), which is usually restricted to the muscle, tongue and brain and was previously associated with hearing loss. Insulin-like growth factor binding protein 2 (IGFBP2), which is usually restricted to the liver, pancreas and stomach and promotes oncogenic processes, was also found to be upregulated in the adrenal gland with increasing age (Fig. 3c). These gene expression profiles offer potential biomarkers for aging and highlight the kind of gene-specific shifts that may contribute to the aging process. To study how transcriptional pathways and signatures are altered with aging, we leveraged the Hallmark pathways from the molecular signatures database (MSigDB; Supplementary Fig. 10 and Supplementary Table 6). The most common significantly upregulated pathways were related to apoptosis and inflammation, whereas the downregulated pathways were most related to metabolism (Supplementary Fig. 10b,c). We also observed that among extracellular matrix/matrisome genes, the group of structural collagens was the most downregulated, followed by the basement membrane proteins, whereas elastic fiber components were the most upregulated (Supplementary Fig. 11). However, each organ showed specific changes, likely related to its specific physiological function (Fig. 3d and Supplementary Figs. 10b,c and 11d,e). When analyzing these data, we observed that changes with biological age were consistently stronger than those with chronological age, regardless of direction (Fig. 3e). Pathways such as p53 signaling, cellular senescence, senescence-associated secretory phenotype, hypoxia, genomic instability and TNF signaling were upregulated with both measures of age but more strongly with biological age. Conversely, the mitotic spindle, G2M checkpoint, adipogenesis, peroxisome and oxidative phosphorylation pathways showed greater downregulation. On average, the slope between biological and chronological age was 1.26 (compared to the expectation of 0), indicating that histologically derived biological age captured a greater degree of transcriptional dysregulation and provided a refined view of the molecular-morphological interplay in aging. Demographic and clinical factors modulating tissue-specific aging The morphological and molecular analyses above captured the general landscape of aging across human tissues, yet aging unfolds differently across individuals. To understand what drives this heterogeneity, we investigated patterns of coordinated and discordant aging across tissues and their association with demographic and clinical variables. By leveraging age gaps estimated for various tissues from the same individual (Fig. 4a), we identified several patterns of cross-tissue aging within individuals: consistently below-average age gaps across tissues ('resilient agers'), consistently average age gaps ('average agers'), single- and multiple-tissue agers (one or a few tissues with greater age than others) and a group of individuals who consistently demonstrated greater predicted age gaps in most tissues ('systemic agers'). To visualize the pattern of aging across many individuals, we focused on a set of 303 individuals with greater than average age gaps in at least one tissue and identified various groups of individuals on the basis of their internal organ synchronicity in aging rates (Fig. 4b). We observed that most individuals with at least one high age gap tended to have several organs with large age gaps (Fig. 4b, pan-tissue cluster). However, we detected several individuals with one particularly prominently aged organ, such as the heart, lung or stomach (Fig. 4c). To provide an overview of the landscape of factors potentially underlying a specific deviation in the aging of one organ compared to others among the 983 individuals, we devised a strategy to associate tissue-specific age gaps with demographic, clinical and lifestyle variables. This analysis sought to characterize broad association patterns across tissues, serving as an exploratory framework rather than a formal test of specific hypotheses. These included 184 variables, from which we identified 80 that were recurrent enough to assess in light of the tissue-specific age gaps (Supplementary Fig. 1i). We discovered that, by and large, most factors tended to affect several or most tissues (Fig. 4d and Extended Data Fig. 9a,b) but also revealed several associations that were tissue and sex specific. One of the most prominent observations was the association of renal failure with accelerated aging across several tissues (Fig. 4d). Interestingly, the strongest association was not found in the kidney but was detected in tissues such as adipose tissue, pituitary, spleen and tibial nerve tissue (Fig. 4e). In women, the association between tibial nerve aging and kidney failure could be related to the fact that more than 90% of individuals with chronic kidney disease who undergo dialysis suffer from peripheral neuropathy. With respect to variations across tissues, we confirmed several known associations, such as chronic respiratory disease most prominently affecting deviations in lung tissue biological age from chronological age (Fig. 4d), type 2 diabetes associated with pancreatic aging (Fig. 4d,e), heart disease associated with adipose tissue aging (Fig. 4d) and the effects of steroid use on the spleen (Fig. 4d). However, we also found other less well-known associations that have not yet been linked to an explicit morphological manifestation of the aging process, such as higher rates of aging in the cerebellum associated with unexplained seizures (Fig. 4d) or accelerated prostate aging and hypertension in men (Extended Data Fig. 9b), even though there are reports that link hypertension with prostate hyperplasia. Blood-based prediction of tissue-specific age gaps Given that gene expression could be associated with tissue-specific biological age derived from images (Fig. 2) and that aging is reflected in a systemic manner in many individuals (Fig. 4), we aimed to develop a predictor of tissue-specific biological age (driven by morphological features in tissue) from gene expression samples of blood in the GTEx cohort (Fig. 5a). We also derived a 'systemic' age gap for each donor, which is the mean of all age gaps across tissues. By linking blood-based gene expression profiles with the histology-derived tissue-specific age gaps of the same individuals, we were able to derive blood-based predictors of tissue age gaps across tissues. We observed good performance in the prediction of age gaps from blood (Extended Data Fig. 10a,b), particularly for the systemic age gap, gastrointestinal tract and spleen. Finally, we observed that the predicted age gaps from blood were negatively associated with telomere length in tissue and highly enriched at the level of pathologies across all tissues of an individual (Extended Data Fig. 10c-e). When we analyzed the genes underlying the predictive process, we found several genes not known to be expressed in blood to be among the top predictors of tissue-specific age gaps (Supplementary Fig. 12). For example, junction plakoglobin (JUP) is expressed in glandular and epithelial cells and has the strongest association with the small intestine, skin and adipose tissue, as well as with systemic predictors. Similarly, the expression of keratin 5 (KRT5) in the blood, which is traditionally a marker of basal epithelial cells, had the strongest association with all genes predicting pancreas age gaps. To validate the blood-based tissue-specific predictors, we leveraged the large-scale repository of harmonized gene expression and sample metadata ARCHS4 (ref. ) to find blood samples of healthy individuals (n = 577) or those with one of seven chronic diseases (n = 628): systemic lupus erythematosus, Crohn's disease, diabetes, cystic fibrosis, ulcerative colitis, vasculitis or Alzheimer's disease (Fig. 5a). As a control, we also included individuals with stroke, representing an acute medical event specifically affecting the brain. We then predicted tissue-specific age gaps based on the coefficients derived from the GTEx cohort and observed that although there was no relationship with sample size, they were distinct for each disease group (Supplementary Fig. 13a-c). Joint analysis of age-gap predictions across all organs revealed that samples were distributed along two axes, with disease entities forming distinct zones (Fig. 5b and Supplementary Fig. 13d,e). One axis showed a transition from healthy samples with near-zero age gaps to samples from various diseases with higher gaps, whereas the second axis reflected increasing age gaps in multiple organs, particularly in the heart, adipose tissue and gastrointestinal tract. Samples from individuals with stroke and individuals with cystic fibrosis had especially high age gaps in the brain and kidney, respectively (Fig. 5b). Statistical testing confirmed significant associations between diseases and elevated tissue-specific age gaps (Fig. 5c,d and Supplementary Fig. 13f). Notably, individuals with stroke showed the most pronounced age-gap deviation specifically in the brain (Fig. 5c-e), in alignment with the affected tissue in this acute event. Moreover, all seven chronic diseases were significantly associated with tissue-specific accelerated aging in at least one organ (Fig. 5d). For example, Crohn's disease was associated with increased predicted age gaps in the gastrointestinal tract (esophagus, stomach, small intestine and colon), whereas vasculitis exhibited higher age gaps in the kidney, liver and heart (Fig. 5e), which are highly vascularized organs and primary targets of inflammatory processes in this disease. In some diseases, fewer and milder but specific associations were detected; for example, the brain was the only organ with significantly higher age gaps in Alzheimer's disease (Fig. 5d). Interestingly, we also found associations of age gaps with organs that are not the primary site of disease action. These associations could be revealing of how treatment of a primary condition may affect other organs, as detected by significantly higher age gaps for individuals with stroke in the kidney and liver, or of secondary effects of disease, such as changes in the brain due to chronic hypoxia in cystic fibrosis, for which we also detected significantly higher age gaps than healthy individuals (Fig. 5e). Leveraging a subset of blood samples for which the chronological age of the donors was available, we could also verify that the inferred age gaps from blood were not associated with chronological age (Supplementary Fig. 14). Finally, to assess the potential clinical utility of blood-based tissue-specific age gaps in detecting disease, we evaluated the performance of thresholded age gaps (Supplementary Fig. 15). The models demonstrated moderate to strong classification performance, as assessed by the area under the receiver operating characteristic curve, and positive predictive values above 0.3 in several cases suggested potential applicability for population-level screening. Overall, this approach demonstrates the feasibility of inferring tissue-specific biological age and disease-related aging signatures from minimally invasive blood samples, underscoring its potential utility in early disease detection and monitoring aging in clinical settings.

AI-based "tissue clocks" measure the biological age of human organs

CeMM Research Center for Molecular Medicine of the Austrian Academy of SciencesAug 14 2026Reviewed AI-based "tissue clocks" can estimate the biological age of human organs from histological images, researchers at the CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences and the Ludwig Boltzmann Institute for Network Medicine (LBI-NetMed) at the University of Vienna showed. By analyzing more than 25,000 tissue samples across 40 tissue types, their study reveals that organs age at different rates throughout life and that these changes can even be detected from blood samples. The findings, published in Nature Medicine (DOI: 10.1038/s41591-026-04566-5), provide a new framework for understanding aging and may open new avenues for disease monitoring and early diagnosis. Some people seem to age slower than others, looking and acting like 45 at 60. Others appear to have gotten ahead of the calendar. But why is that, and what is actually happening inside the body? Does a liver age differently from a brain? And is it possible to measure the gap between the age on a passport and the biological age of each organ? By combining artificial intelligence with one of the world's largest collections of human tissue images, a new study led by CeMM and LBI-NetMed Principal Investigator André Rendeiro and co-first authored by Ernesto Abila, Iva Buljan, and Yimin Zheng, takes a large step towards answering these questions. While previous studies focused mainly on molecular changes such as DNA methylation or gene expression, the team examines how the architecture of tissues themselves changes over time. A silent diary of time To do this, the researchers turned to the Genotype-Tissue Expression Project (GTEx), which collected tissue samples from 983 individuals across 40 different tissue types, ranging from the brain and heart to the lung, pancreas, skin, and intestine. These were transformed into high-resolution digital photographs of tissue slices, each revealing the microscopic architecture of the organ in question. The scale is staggering: 25,712 images, representing ~480 million individual image tiles, analyzed with state-of-the-art vision models. They found that the architecture of organs keeps a silent diary of time: Even without explicitly teaching the AI about it, age turned out to be the single strongest factor shaping tissue appearance across all 40 tissue types. Building on this, the research team developed so-called 'tissue clocks' - predictive models that estimate a person's biological age from the appearance of their tissue, for each organ independently. These clocks achieved a mean prediction error of just 4.9 years and outperformed existing DNA-based aging estimates in capturing tissue-specific pathology. Importantly, the predicted biological age was strongly linked to known hallmarks of aging, including telomere shortening, tissue pathology, and the number of chronic diseases an individual had. Our tissues carry a remarkably detailed record of the aging process. By combining histology images with artificial intelligence, we can detect patterns of biological aging that are invisible to the human eye and begin to understand how aging unfolds differently across the body." André Rendeiro, Principal Investigator at CeMM and corresponding author of the study Different schedule for every organ The analysis revealed that aging does not occur uniformly: Some tissues, such as the lung, kidney, pancreas, and adrenal gland, showed signs of accelerated aging already between the ages of 20 and 40. Others followed more complex trajectories, with peaks of accelerated aging appearing later in life. The uterus displayed a particularly striking shift around the age of menopause. The researchers also identified strong links between tissue-specific aging and medical conditions or lifestyle-associated factors. For example, kidney failure was associated with accelerated aging signals in multiple tissues, while diabetes showed pronounced effects in the pancreas. "What stands out is how differently each organ ages, and how that shows up in tissue architecture," says Ernesto Abila, co-first author of the study. "Deep learning lets us read these spatial patterns, capturing aging as architectural remodeling, not just molecular drift." While the tissue clocks captured the normal pace of aging across organs, they also highlighted outliers - individuals whose tissues showed pronounced structural shifts ahead of their chronological age. However, tissue samples cannot always be collected. By linking blood-based gene expression profiles with the histologically derived tissue age gaps of the same individuals, the researchers built predictors of tissue-specific biological age from blood samples alone. "This is a conceptual leap: using the language of tissue aging, learned from images, and translating it into something readable from a routine blood draw," explains co-first author Iva Buljan. Blood samples show aging patterns These blood-based predictors successfully identified aging patterns linked to several diseases, including Alzheimer's disease, Crohn's disease, cystic fibrosis, vasculitis, diabetes, and stroke. In Alzheimer's disease, for example, the strongest aging signal was detected specifically in the brain, whereas Crohn's disease showed accelerated aging across the gastrointestinal tract. "This study highlights that aging is not simply a matter of chronological time," says Yimin Zheng, the third co-first author of the study. "Different organs age in different ways, and these processes appear to be shaped by both systemic and tissue-specific factors." The findings suggest that tissue architecture integrates many of the molecular and physiological changes associated with aging and disease. In the future, such approaches could contribute to minimally invasive diagnostics that monitor organ health and disease progression through blood tests. The study also demonstrates the growing potential of artificial intelligence in pathology and aging research. By connecting tissue imaging, gene expression, and clinical data at large scale, the work provides a comprehensive view of how aging manifests throughout the human body. Source: CeMM Research Center for Molecular Medicine of the Austrian Academy of Sciences Journal reference: Abila, E., et al. (2026). Histological aging signatures for monitoring tissue-specific aging and disease. Nature Medicine. https://doi.org/10.1038/s41591-026-04566-5. https://www.nature.com/articles/s41591-026-04566-5