Telomere length is one of the most widely studied biomarkers of biological aging, yet measuring it at scale remains technically challenging and costly. Now researchers have developed a computational approach that may allow scientists to estimate telomere length directly from routine histopathology slides—potentially enabling large-scale studies of aging and age-related disease using existing clinical biopsy data.
The method, described in Cell Reports Methods, uses a deep learning framework called TLPath to infer telomere length from subtle structural features present in standard tissue images.
Developed by investigators at Sanford Burnham Prebys, the model demonstrates how computational pathology may unlock new biological insights from imaging data already generated in clinical care.
Linking tissue structure to a key aging biomarker
Telomeres are repetitive DNA sequences that cap the ends of chromosomes, protecting genetic material during cell division. Over time, these protective caps shorten as cells replicate, making telomere length an important indicator of cellular aging.
“Whenever DNA gets replicated as our cells grow and divide, the part at the end of the DNA cannot be replicated,” said senior author Sanju Sinha, PhD, assistant professor in the Cancer Metabolism and Microenvironment Program at Sanford Burnham Prebys.
Cells address this limitation by protecting chromosome ends with telomeres, which gradually shorten over a lifetime. Numerous studies have linked telomere length to aging and to the risk of age-related diseases including cardiovascular disease, cancer, and neurodegeneration.
However, direct measurement of telomeres typically requires specialized laboratory assays such as quantitative PCR, fluorescence in situ hybridization, or sequencing-based approaches—methods that are not routinely applied to large clinical cohorts.
The researchers hypothesized that morphological changes associated with cellular aging might be detectable in tissue architecture, allowing telomere length to be inferred from histopathology images.
Training a model on thousands of biopsy slides
To test this idea, the investigators leveraged data from the Genotype-Tissue Expression (GTEx) Project, a large National Institutes of Health initiative that pairs genomic measurements with tissue samples collected from hundreds of individuals.
The team trained TLPath using 5,263 digitized histopathology slides representing 18 tissue types from 919 donors. Each slide was paired with laboratory measurements of telomere length, enabling the algorithm to learn relationships between tissue morphology and telomere biology.
The model processes each slide by dividing it into small image fragments, or “patches.” On average, a slide is segmented into about 1,387 patches, each analyzed for up to 1,024 structural features.
By weighting these features and integrating information across patches, TLPath produces a prediction of telomere length for the entire tissue sample.
The model was trained separately for different tissue types to account for organ-specific differences in cellular structure.
A foundation-model approach to computational pathology
TLPath builds on recent advances in foundation models for digital pathology, which learn higher-level visual features beyond simple pixel patterns.
“These models don’t look at discrete pixels, but instead define higher-order features,” Sinha explained. “Only some of these features can be interpreted by humans, yet they can be validated for their predictive power.”
This approach has rapidly gained traction across biomedical imaging, where large neural networks trained on vast datasets can uncover patterns invisible to conventional analysis.
In validation experiments, TLPath successfully predicted telomere length in biopsy samples that were not included in the training dataset.
Notably, the model outperformed predictions based solely on chronological age, suggesting it captures biological variation in aging processes that age alone cannot explain.
The researchers also demonstrated that TLPath could distinguish differences in telomere length between individuals of the same age, indicating that tissue architecture may encode information about biological aging beyond simple time-dependent changes.
Enabling larger studies of aging biology
A key advantage of the new approach is scalability.
Direct laboratory measurement of telomere length can be expensive and difficult to perform on large cohorts. In contrast, histopathology slides are routinely generated in clinical practice.
“Directly measuring telomere length requires more complicated and costly tests that are difficult to scale,” Sinha said.
By extracting telomere-related signals from digital tissue images, TLPath could enable researchers to analyze aging biomarkers across massive pathology datasets.
Such datasets already exist in biobanks and clinical archives, many of which contain decades of tissue samples.
The primary barrier to broader use, the researchers note, is that many histology slides remain undigitized.
“Whether it is new slides being developed today or those preserved in biobanks, all we need is for them to be properly scanned, stored and shared,” Sinha said.
Further implications for aging research
The study highlights the growing convergence of computational pathology, machine learning, and aging biology.
If validated in additional datasets, models like TLPath could support large population studies examining how telomere dynamics influence disease risk, treatment responses, and longevity.
More broadly, the work demonstrates that biological aging markers may be encoded in tissue architecture itself, potentially allowing clinicians and researchers to extract molecular insights from routine pathology images.
As digital pathology adoption accelerates worldwide, tools capable of linking histologic features to molecular aging processes may become an increasingly powerful resource for precision medicine research.
