ChromAgeNet reads hematopoietic stem cell age from a nucleus image, yet on its scale compounds with no proven benefit outscored one whose benefit is already established
ChromAgeNet reads hematopoietic stem cell age from a nucleus image, yet on its scale compounds with no proven benefit outscored one whose benefit is already established
On September 27, researchers from IDIBELL (the Barcelona Biomedical Research Institute), the Barcelona Supercomputing Center (BSC), and ISGlobal (the Institute for Global Health) published ChromAgeNet in Aging Cell: a neural network that distinguishes young from old hematopoietic stem cells (the cells that produce all blood cells) from a 3D image of the cell nucleus, achieving a classification score of 0.77 out of a maximum of 1.0 (random guessing would yield 0.5). The authors have already used it as a rapid screen for four candidate anti-aging compounds and obtained results that conflict with their own earlier work.
With age, the packaging of chromatin, the complex of DNA and proteins that determines which genes are active and which are silenced, deteriorates inside the cell nucleus. In young hematopoietic stem cells, a thin continuous band of tightly packed, inactive chromatin runs along the nuclear periphery; in old cells this band breaks apart and large, shapeless clumps appear inside the nucleus. The team trained the network to recognize this disorganization on 1,229 three-dimensional images of nuclei stained with the inexpensive dye DAPI: 551 from young mice (10 to 16 weeks) and 678 from old mice (80 weeks and older).
The model is compact, with roughly 350,000 parameters. Its score of 0.77 exceeded the 0.73 of a classical method based on manually selected chromatin features: the network learned to find the aging pattern on its own. Until now, the most common way to measure a cell's biological age has been epigenetic clocks, a method based on DNA methylation introduced over a decade ago. It requires sequencing, which means cost, time, and destruction of the cell itself. ChromAgeNet works from an ordinary image, leaves the cell intact, and is suited for screening hundreds of compounds where sequencing every sample would be prohibitively expensive. This measurement is independent of methylation: an epigenome map of blood from 120 donors showed that DNA methylation and chromatin accessibility in the same individual rarely overlap at the same genomic positions. The authors also released the full image dataset publicly, a rare resource in this narrow field.
The authors tested two distinct rejuvenation mechanisms on aged mouse cells. Without treatment, these cells had a mean youth score of 0.34, compared with 0.55 for young cells. Rhosin, which blocks RhoA (a protein that controls mechanical tension in the nuclear envelope), raised the score only to 0.48. The authors tested an independent hypothesis with UNC0646 and IOX1, compounds that remove from histones (the proteins that package DNA) a mark that "seals" chromatin regions: the score rose nearly to the young-cell level, reaching 0.55 to 0.56.
Rhosin targets the same nuclear-tension pathway this laboratory has studied since 2018, when it first showed that aging of these cells is linked to loss of lamin A/C from the nuclear envelope. In 2025, the same group demonstrated in Nature Aging that suppressing RhoA restores normal function to aged mouse hematopoietic stem cells. A compound with already confirmed functional benefit therefore scored lower on the new imaging measure than compounds whose effect has been verified only by chromatin appearance.
The gap between chromatin appearance in an image and actual cell function shows where tools like these can mislead. Inexpensive visual assays such as ChromAgeNet speed up compound screening by an order of magnitude, but a compound that merely changes the picture under the microscope, without a separate functional test, may turn out to be inert.