The bottom line: Researchers trained an artificial intelligence system called ChromAgeNet to distinguish young from aged mouse blood stem cells using three-dimensional images of chromatin inside the nucleus. The work shows that nuclear organization carries measurable information about cellular aging, but it is an experimental research tool, not a clinically validated biological-age test.
Evidence at a glance
- Study type: Laboratory imaging and machine-learning study in mouse hematopoietic stem cells.
- Data: 1,229 three-dimensional images of nuclei from young and aged cells, converted into more than 81,000 two-dimensional image slices for model development.
- Main result: At the whole-nucleus level, ChromAgeNet separated young from aged cells with an average area under the receiver operating characteristic curve of 0.77 and an accuracy of 68% during cross-validation.
- What the model saw: Age-associated patterns involving chromatin texture, compact DNA near the nuclear edge and dense chromatin regions around structures such as nucleoli.
- Evidence level: Early preclinical research. The model was trained on mouse cells and has not been validated as a test of aging in people.
A picture of how DNA is packed
Hematopoietic stem cells, or HSCs, reside mainly in bone marrow and continually supply the body with blood and immune cells. With age, the HSC population changes. Older cells can become less balanced in the types of blood cells they produce, less effective at regeneration and more prone to changes associated with blood disease.
The new study focuses on chromatin, the complex of DNA and proteins that packages the genome inside a cell nucleus. Chromatin is not a featureless spool. Some regions are compact and relatively inaccessible, while others are open to the machinery that reads genes. Its three-dimensional arrangement helps a cell maintain its identity and decide which genetic programs to use.
Aging can disturb that organization. Boundaries may become less orderly, compact chromatin can shift, and structures near the nuclear envelope or nucleolus can change. These differences are often subtle and heterogeneous, meaning that two stem cells from the same animal may not look equally old.
How ChromAgeNet was trained
The researchers collected high-resolution confocal microscopy images of HSC nuclei from young mice, up to 16 weeks old, and aged mice, more than 80 weeks old. They stained the nuclei with DAPI, a widely used fluorescent dye that binds DNA and makes chromatin visible.
Each nucleus was imaged as a stack of slices through its depth. ChromAgeNet, a convolutional neural network designed to recognize spatial patterns in images, analyzed those slices and produced a “youthful score.” Scores across the stack were then combined into one prediction for the complete nucleus.
The team used five-fold cross-validation, repeatedly training the model on most nuclei and testing it on nuclei held out from training. Splitting at the nucleus level was important because it kept slices from the same cell from appearing in both training and validation data.
At the whole-nucleus level, the model achieved an average AUROC of 0.77. AUROC measures how well a model ranks examples from two groups across possible decision thresholds, with 0.5 representing chance and 1.0 representing perfect separation. ChromAgeNet’s average validation accuracy was 68%. It was more sensitive to young cells than specific for aged cells, and many predictions clustered near the uncertain middle of the scale.
Why the AI added something
The researchers also built conventional machine-learning models using 70 features that scientists had defined in advance, including measurements of nuclear shape, intensity and texture. The best of those models reached an average AUROC of 0.73. ChromAgeNet’s 0.77 was a modest improvement, suggesting that the neural network found useful combinations of spatial detail beyond the handcrafted measurements.
The team then used explainable-AI methods to ask which parts of the images influenced the predictions. Important signals included chromatin entropy, a measure related to how irregular or unpredictable the image texture is; peripheral heterochromatin, the compact DNA often found near the nuclear envelope; and dense chromatin condensates around nuclear structures.
Young-cell predictions tended to draw attention to more structured features near the nuclear edge and nucleoli. Aged-cell attention was more dispersed and frequently involved large, dense or disrupted regions. These patterns fit existing ideas about age-related loss of nuclear organization, but the analysis identifies associations. It does not prove that any one visual feature causes stem-cell decline.
Different from an epigenetic clock
Many familiar biological-age measures are epigenetic clocks. They typically analyze DNA methylation, small chemical marks attached at selected sites in the genome, and use statistical relationships between those marks and age or health outcomes to calculate an age estimate.
ChromAgeNet measures something different. It uses microscopy to read the physical arrangement and texture of DNA-containing chromatin within individual nuclei. It does not measure methylation at specific DNA letters, sequence genes or return a person’s age in years. Its output is a probability that a mouse HSC image resembles the young or aged group used for training.
The two approaches could eventually be complementary. Methylation clocks summarize chemical patterns across many genomic sites, while nuclear imaging may capture cell-to-cell variation and spatial organization. That possibility still requires direct comparison and validation in independent samples.
A possible screening tool, not proof of rejuvenation
As a proof of concept, the team applied ChromAgeNet to aged HSCs exposed to four compounds that affect signaling or chromatin regulation. Two compounds that alter histone H3K9 methylation shifted the model’s scores toward those of untreated young cells. Other compounds produced smaller shifts.
That result shows the model can detect drug-associated changes in nuclear appearance. It does not show that the cells were functionally rejuvenated. The study did not establish that treated cells regenerated blood better, remained safe over time or improved health in an animal. A compound could make chromatin look more youthful to the model without restoring the full biology of a young stem cell.
The practical appeal is speed. DAPI staining is inexpensive and compatible with high-throughput microscopy. If the method proves robust, researchers might use it as one layer of a screening pipeline to flag interventions for deeper functional testing.
What remains to be tested
ChromAgeNet was developed on one type of stem cell from mice. It has not shown that the same visual signature works in human HSCs, other tissues, different laboratories or across a continuous range of ages. Imaging conditions can also create technical patterns that a model might learn, although the researchers used cross-validation, calibration and careful data splitting to reduce that risk.
The performance is useful for a research proof of concept but far from diagnostic certainty. The cells also showed substantial heterogeneity, which is biologically interesting but limits simple classification.
Before any clinical claim would be justified, the approach would need independent replication, standardized imaging, human validation and evidence that its scores predict meaningful functions or outcomes beyond chronological age. None of that has yet been demonstrated.
The Lifespan Brief assessment
This study makes a clear, limited point: the three-dimensional architecture of chromatin contains enough age-related information for an image model to distinguish groups of young and aged mouse blood stem cells better than chance.
The most promising near-term use is not consumer age testing. It is laboratory phenotyping, especially as a relatively inexpensive way to compare many individual cells and prioritize candidate interventions for more demanding experiments. Whether nuclear images can become a reliable measure of human biological aging remains an open question.
Primary sources
Deep learning predicts haematopoietic stem cell ageing from 3D chromatin images, Aging Cell, 2026.
Institutional research summary from IDIBELL, September 28, 2026.
This article is for general information and is not personal medical advice.