TLPath: Leveraging the Digital Pathology Foundational model for Telomere Length Prediction

Anamika Yadav, Kyle Alvarez, Akanimoh Adeleye, Sanju Sinha

Keywords: Digital Patholgy, Foundational Model, Machine Learning

TL;DR: We introduce TLPath, a ML model which uses UNI features to predict bulk telomere length from histopathological images. We do the mechanistic interpretation of the TLPath as well.

Abstract: Telomere dysfunction drives aging and age-related diseases, yet its large-scale study is hindered by reliance on specialized molecular assays, limiting clinical and research advancements. Here we present TLPath,  a deep learning framework that leverages digital pathology foundational model, UNI, to predict bulk-tissue telomere length from routine H\&E-stained images. The pipeline extracts morphological features from image patches and aggregates them into a whole slide-level representations, which are then used in a supervised model to accurately predict telomere length. These extracted features can predict bulk-telomere length with significant accuracy (\(>0.51\) in well-represented tissues), outperforming chronological age as a predictor (correlation = 0.20) and identifying age-discordant cases – detecting both accelerated telomeres shortening in young individuals and preserved telomeres in older individuals. Moreover, the mechanistic interpretation of TLPath reveals that its predictions are grounded in established cellular senescence markers such as the nuclear to cytoplasmic ratio and nuclear shape variation.

Submission Number: 88