Photographic age gap is associated with frailty and predicts survival in laboratory mice

Abstract

Ageing studies in mice need biomarkers that are cheap, non-invasive and repeatable. We asked how much of a mouse's age can be read from an ordinary photograph, and whether the part that cannot be explained by age carries biological information. We assembled 2,846 photographs of 1,292 mice (four strains, 1 to 34 months old) from seven facilities and predicted age with a frozen self-supervised vision model (DINOv2) and a ridge regression, always tested on facilities the model had never seen. Absolute age transferred poorly between facilities (median absolute error 138 days), because each facility's photographs carried an offset. Relative age transferred well: among mice photographed in the same session, the older mouse was predicted older in 65% of pairs, rising above 95% for age differences over 500 days, and the later photograph of the same mouse was predicted older in 63% of repeat pairs. Sex, in comparison, was inferred more reliably (87%). Mice predicted older than their age were frailer (+0.12 SD per 100 days) and had shorter survival (hazard ratio 1.14 per 100 days), and fourth generation (G4) telomerase-deficient mice looked 151 days older than their wild-type controls. Routine photographs therefore carry a real biological age signal that can be read at no extra cost, best used to compare mice photographed together.