Tissue morphology predicts telomere shortening in human tissues.

We present TLPath, a deep learning framework that predicts bulk-tissue telomere length from tissue morphology extracted from routine histopathology images. The central thesis of TLPath is that telomere length can be determined from tissue morphology. Trained on >5,000 whole-slide images across 919 individuals and 18 organs, TLPath discovers that extracted morphological features naturally separate young, middle-aged, and elderly individuals, demonstrating that aging causes architectural changes in tissues detectable without explicit age supervision. These features accurately predict telomere length (correlation r = 0.51) in 11 tissues, outperforming chronological age. This enabled us to identify young tissues with considerably shortened telomeres than expected and old tissues with preserved telomeres based on their morphology. Model interpretation revealed that TLPath leverages senescence markers, including increased nuclear-to-cytoplasmic ratio and altered nuclear shape heterogeneity, resulting from telomere shortening. In ∼2,800 GTEx biopsies, TLPath detected shortened telomeres across multiple tissues in individuals with type 1 and type 2 diabetes, validated experimentally. H&E imaging-based tissue morphology can determine bulk-telomere shortening, enabling large-scale telomere biology studies.