A computational tool predicts telomere length from routine biopsy slides - no DNA test needed

Every time a cell divides, its chromosomes get a little shorter. Not the important parts - evolution solved that problem by capping chromosome ends with repetitive DNA sequences called telomeres, which absorb the shortening so that essential genes are preserved. But telomeres are not expendable buffers. Their length correlates with chronological age, predicts risk for chronic diseases, and has become one of the most studied metrics in aging biology.

The problem is measuring them. Direct telomere length assays require specialized laboratory techniques that are expensive, time-consuming, and difficult to scale. A team at Sanford Burnham Prebys Medical Discovery Institute has now shown that a computational model can infer telomere length from something far simpler: routine biopsy slides that are already being made in hospitals every day.

Training a model on tissue architecture

The tool, called TLPath, works on the hypothesis that telomere shortening leaves visible traces in cell and tissue structure - changes in morphology that a sufficiently sensitive model can detect even though a human pathologist looking at the same slide would not consciously register them.

The research team, led by assistant professor Sanju Sinha, trained TLPath using data from the NIH-funded Genotype-Tissue Expression (GTEx) Project. This large resource pairs high-resolution histopathology images with laboratory measurements of telomere length for the same tissue samples. The training set comprised 5,263 slides from 18 tissue types donated by 919 individuals.

The model works by segmenting each slide into an average of 1,387 square fragments, or patches. For each patch, it identifies up to 1,024 structural features using computer vision foundation models - higher-order patterns in cell and tissue architecture that go beyond individual pixels. A statistical weighting across all patches produces an overall score for each slide, which the model learns to map to telomere length.

What TLPath can and cannot do

After training separately on each tissue type, TLPath successfully predicted telomere length on held-out GTEx samples that were not part of the training data. It outperformed the simpler approach of predicting telomere length based solely on the donor's age - a meaningful bar to clear, since age is the strongest single predictor of telomere length in the general population.

More impressively, the model could distinguish telomere length differences between individuals of the same chronological age. This is where the tool's potential value becomes clear: two 50-year-olds may have very different biological ages as reflected by their telomeres, and TLPath can pick up on structural differences in their tissues that correlate with that variation.

The findings were published in Cell Reports Methods.

The scalability argument

The practical appeal of TLPath is not accuracy alone - it is throughput. Direct telomere measurement requires DNA extraction, specialized assays, and significant per-sample cost. Histopathology slides, by contrast, are already produced in enormous numbers as part of routine clinical care. They just need to be digitized.

Sinha emphasized this point: the only limit to using TLPath at scale is the availability of scanned histopathology slides. Whether those slides are generated fresh or pulled from biobank archives, digitization is the sole bottleneck. If hospitals and biobanks invested in systematic slide scanning and sharing, the approach could enable population-scale telomere studies that would be prohibitively expensive using direct measurement.

Caveats and open questions

TLPath was developed and validated on GTEx data, which skews toward deceased donors (most GTEx tissues come from post-mortem collection). How well the model generalizes to biopsies from living patients - which differ in tissue handling, fixation quality, and clinical context - remains to be tested.

The model predicts a correlate of telomere length, not the thing itself. The structural features it detects in tissues are associated with telomere status but may also reflect other biological processes - cellular stress, inflammation, or tissue remodeling - that happen to co-vary with telomere shortening. Disentangling these signals will require validation against diverse datasets and direct comparison with established telomere measurement methods.

The 18 tissue types in the training set cover many major organs but not all. Telomere dynamics differ across tissues, and the model would need separate training and validation for tissue types not represented in GTEx.

Still, the core conceptual advance is solid: measurable structural changes in cells and tissues can predict telomere length. That opens the door to studying telomere biology at a scale that molecular assays alone cannot reach - and potentially to integrating aging metrics into routine pathology workflows where biopsy slides are already in hand.