TLPath: The AI That Can Measure Your Telomere Length from a Routine Biopsy Image

Telomere length is one of the oldest biomarkers in aging research. The shorter your telomeres are relative to your age, the higher your risk for heart disease, diabetes, Alzheimer's, and cancer. The problem: to measure them, you need a complex and expensive lab test. Until now. A new study published in Cell Reports Methods in March 2026 reveals an AI model named TLPath, developed in the lab of Sanju Sinha at the Sanford Burnham Prebys Institute, which can predict tissue telomere length from routine histology images.

The Problem: Why Telomeres Are Hard to Measure

Telomeres are repetitive DNA sequences at the ends of chromosomes that shorten with each cell division. They gradually shorten with age. Telomere length is currently measured using specialized lab tests, such as:

• qPCR: Relatively cost-effective, but less accurate for certain tissues
• TRF (Terminal Restriction Fragment) or Southern blot: Accurate but expensive and requires a large amount of DNA
• FISH (Fluorescence In Situ Hybridization) or Luminex-based methods: Additional methods used in research

The telomere data used to train the model came from the GTEx database, where telomere length was measured using a Luminex-based method (Demanelis et al., Science 2020). As Sinha put it: "Direct measurement of telomere length requires complex and more expensive tests that are difficult to implement on a large scale". This is exactly where TLPath comes in.

The Idea: What If There Are Signs of Telomere Length in the Image?

The team asked a simple question: When telomeres shorten, the cell changes. It can become zombie-like (senescent), slow division, change its shape, or lose internal structures. Are these changes visible in a microscopic image of the tissue?

If so, a deep neural network could be trained to identify them. Every hospital in the world produces millions of biopsy images as part of routine care. If there is a visual signature of telomere length, it could be estimated directly from the existing clinical sample.

How the Network Was Trained

The team collected 5,263 digital histopathology images (routine H&E stained images) from 919 donors. Each image was paired with a lab measurement of telomere length from the same tissue. 18 different tissue types were included: skin, lung, kidney, liver, intestine, etc.

The network cuts each image into an average of 1,387 small patches. Each patch is examined using up to 1,024 structural features: cell shape, nucleus structure, cytoplasm color, distances between cells. The network learns which combination of features predicts short telomeres, and which predicts long ones.

The Results: Accuracy Beyond Expectations

On test samples not part of the training set, TLPath showed:

• r = 0.51 correlation between its prediction and the lab measurement across 11 tissue types. This is not as accurate as direct measurement, but it clearly outperforms estimation by chronological age alone (which achieves only r = 0.20), which is the current standard when no measurement is available
• Works on 11 different tissue types, demonstrating generality
• Model interpretation revealed it relies on senescence markers (cellular aging), such as an increased nucleus-to-cytoplasm ratio and changes in nuclear shape

"Direct measurement of telomere length requires complex and more expensive tests that are difficult to implement on a large scale," explained Sinha. TLPath is designed to bridge this gap and estimate telomere length from images that already exist.

The Significance: A Data Revolution

If TLPath is integrated into standard digital pathology software, here is what could become possible:

1. Population-scale longevity research. Instead of sampling thousands, telomere length could be estimated for many more people from existing images
2. Early identification of candidates for intervention (potential). A person coming for a pathology test, even at age 40, who is found to have a low telomere estimate for their age, could potentially start a protective lifestyle early. It is important to emphasize: this is a future research direction, not an approved clinical application

Important note: TLPath was trained to predict average telomere length in bulk tissue with only moderate accuracy (r = 0.51), and it is a research tool. It is not an approved clinical test, and is not intended for tailoring treatments like chemotherapy or for drug screening. Such uses were not tested in the study.

Why This Is Not Just Another AI Model

Many AI models in 2026 do impressive but impractical things. TLPath is different: it attempts to solve a specific problem on a large scale with existing infrastructure. Many hospitals already scan their images digitally. No new equipment, no additional procedure for the patient. Just adding a software component.

This is what scientists in the field of digital pathology call "value add": additional information extracted from a test you already performed.

Limitations to Keep in Mind

• The correlation r=0.51 means about 26% of the variance is explained. Not strong enough to serve as an accurate individual test for a single person, but useful at the statistical-population level and for research
• The model was trained on a specific population. Use on different populations (different ethnicities) requires further validation
• Telomeres are just one marker of biological age. It needs to be combined with others (epigenetic, proteomic)
• The network does not explain why the telomeres are short. Only that they appear short based on tissue characteristics

Bottom Line

TLPath is a promising research tool, not a clinical test. Its code is openly available on GitHub (Sinha-CompBio-Lab/TLPath) for researchers who want to extend the work. The broader conclusion: measuring aging may gradually move from expensive labs to tools that can be applied to samples that exist everywhere. If TLPath is the first step, it may be just the beginning of a whole wave of "image-based biomarkers": models that extract estimated value from an existing sample that was not previously visible. However, the moderate accuracy reminds us that this is still the beginning of the road.