AI Reads the 3D Architecture of Chromatin to Predict Stem Cell Aging

Deep inside every blood-forming stem cell, the genome is folded into an intricate three-dimensional architecture that changes as we grow older. A research team has now shown that a convolutional neural network can read those changes directly from ordinary fluorescence microscopy images and reliably distinguish young stem cells from aged ones. The new framework, called ChromAgeNet, was trained on high-resolution 3D confocal images of hematopoietic stem cell (HSC) nuclei stained with DAPI, a cheap and widely available DNA dye. Published in Aging Cell, the study demonstrates for the first time that deep learning can model age-associated chromatin remodeling from 3D images of single stem cell nuclei, opening a path toward image-based biological age estimation and, potentially, the discovery of drugs that rejuvenate aged stem cells.

Hematopoietic stem cells are the cornerstone of the blood and immune system, maintaining lifelong homeostasis of all blood lineages. As they age, their functional decline drives the deterioration of the entire hematopoietic system in the elderly, contributing to anemia, weakened immunity, and age-related blood disorders. Although biologists have catalogued twelve hallmarks of aging, including cellular senescence, stem cell exhaustion, and epigenetic alterations, the precise molecular processes that erode HSC function over time remain incompletely understood. Because HSCs are extraordinarily heterogeneous, with individual cells in the same aged animal showing very different degrees of decline, researchers have long needed tools that can assess age-associated chromatin states cell by cell rather than in bulk populations.

Existing biological age estimators rely on other molecular readouts. Epigenetic clocks based on DNA methylation, pioneered more than a decade ago, can predict mortality and disease risk and are used to evaluate rejuvenation therapies. Single-cell transcriptomic clocks, proteomic and metabolomic signatures, telomere length, and markers of genomic instability all capture complementary dimensions of aging. What has been missing is a way to exploit the spatial organization of chromatin inside the nucleus, which is known to deteriorate with age at every scale of the genome’s folding hierarchy, from chromosome territories and A/B compartments down to topologically associating domains and chromatin loops. During aging and senescence, cells lose facultative heterochromatin, chromatin redistributes from the nuclear periphery toward the interior, and lamin-associated domains erode, detaching peripheral heterochromatin from the nuclear lamina.

To capture these changes, the team collected 1,229 3D confocal microscopy images of HSC nuclei from young mice aged 10 to 16 weeks and aged mice older than 80 weeks. Because long-term HSCs are rarer in young animals, more young mice were needed to obtain comparable cell numbers. The researchers segmented each 3D stack, resized the images to a uniform isotropic resolution, and extracted 2D slices along the Z-axis, using interpolation to generate additional synthetic slices that preserved biologically meaningful intensity distributions. The result was a balanced training set of 81,643 2D slices, roughly 44 percent from 552 young HSCs and 56 percent from 551 aged HSCs, acquired by three independent operators on two different confocal microscopes.

ChromAgeNet itself is a compact convolutional network, a reduced adaptation of the Xception architecture with depthwise separable convolutions and residual connections, containing only about 350,000 trainable parameters. The network processes each 2D slice and outputs a probability that the chromatin pattern belongs to a young nucleus. Predictions from all slices of the same 3D nucleus are then aggregated through soft voting to produce a single nucleus-level score, a strategy that measurably improved performance. In fivefold cross-validation, with data split at the level of whole nuclei to prevent information leakage, the model achieved an average validation AUROC of 0.77 at the nucleus level, up from 0.72 for individual slices, with the best fold reaching 0.81. After post hoc beta calibration, the output scores aligned closely with true class frequencies, with an expected calibration error of just 0.02.

The researchers benchmarked this approach against classical machine learning pipelines trained on 70 handcrafted chromatin features, including intensity statistics, shape descriptors, and texture measures computed according to the Imaging Biomarker Standardization Initiative. A calibrated XGBoost classifier performed respectably, reaching an AUROC of 0.73, but fell short of the deep learning model. Importantly, the interpretable feature analysis revealed which chromatin properties changed with age: texture descriptors such as size zone nonuniformity and sum entropy, along with Wavelet median intensity and least axis length, consistently ranked highest. Features derived from Laplacian of Gaussian and Wavelet filtered images were especially informative, suggesting that fine-scale spatial variations in chromatin density carry much of the aging signal. A UMAP embedding built from these top features showed clear separation between young and aged cell populations.

Because convolutional networks are notoriously opaque, the team applied a battery of explainable AI techniques to peer inside ChromAgeNet’s decisions. Occlusion sensitivity, which measures how masking small image patches affects predictions, proved the most reliable attribution method in systematic benchmarking. The resulting attention maps told a coherent biological story. In young HSCs, model attention concentrated along the nuclear envelope, forming a continuous thin line of DAPI signal presumably corresponding to lamin-associated facultative heterochromatin, as well as at nuclear invaginations and regions near nucleoli. In aged HSCs, attention was more dispersed and entropy in the attribution maps was higher, highlighting large, irregular DAPI-dense blobs and a thickened, disrupted envelope signal, consistent with known age-related heterochromatin deterioration. Notably, the model’s scores showed no meaningful association with technical variables such as microscope operator, image intensity, noise, or Z-stack size, indicating that it captures biology rather than acquisition artifacts.

The most striking application came when the researchers treated aged HSCs with candidate rejuvenating compounds and asked whether ChromAgeNet could detect a shift toward a youthful chromatin state. They tested four drugs in two mechanistic classes: inhibitors of Rho GTPase signaling, including the Cdc42 inhibitor CASIN and the RhoA inhibitor Rhosin, and inhibitors of H3K9 methylation, including the G9a/GLP methyltransferase inhibitor UNC0646 and the broad-spectrum demethylase inhibitor IOX1. Baseline aged HSCs received a mean youthful score of 0.34. CASIN and Rhosin nudged scores modestly upward, to 0.40 and 0.48 respectively. Remarkably, UNC0646 and IOX1 raised scores to 0.56 and 0.55, essentially matching untreated young HSCs, which averaged 0.55. The finding suggests that pharmacological modulation of H3K9 methylation can restore a youthful nuclear architecture in aged stem cells, and it positions ChromAgeNet as a scalable phenotypic screening readout for rejuvenation drug discovery.

The authors are careful to note the limitations. Ground-truth labels came from the chronological age of the donor mice, a surrogate for biological age that introduces learning noise given the heterogeneity of aged HSCs. Batch effects, a known challenge for image-based profiling, were mitigated through standardized preprocessing, data augmentation, and cross-validation, but wider application will require validation on independent datasets from other institutions, microscopes, and imaging conditions. DAPI staining alone also cannot reliably separate true aging from cellular senescence, whose chromatin signatures overlap, and translation to human samples will demand model adaptation because chromatin organization differs between mouse and human cells. Future work could incorporate additional molecular layers, such as RNA polymerase II or specific histone modifications, and could adopt fully 3D convolutional architectures or visual transformers as larger datasets become available.

Even so, the study marks a conceptual advance: aging leaves a legible fingerprint in the three-dimensional texture of chromatin that a small, inexpensive neural network can decode from a single DNA stain. Because DAPI is affordable and compatible with standard imaging workflows, and because the model is light enough for high-throughput use, the approach could be extended to other stem cell compartments and cell types where chromatin remodeling is a conserved feature of aging. The team has released its curated 3D HSC microscopy dataset together with ChromAgeNet as an open resource, addressing the current scarcity of publicly available stem cell imaging data. If validated functionally and across laboratories, image-derived youthful scores could one day complement DNA methylation clocks, providing a rapid, single-cell view of biological age and accelerating the search for therapies that restore youthful architecture to aging tissues.

Subject of Research: Deep learning-based prediction of hematopoietic stem cell aging from 3D chromatin imaging

Article Title: Deep Learning Predicts Hematopoietic Stem Cell Aging From 3D Chromatin Images

Article References: Picazo, P. I., Mejía‐Ramírez, E., Di Bari, D., Vitali, E., Florian, M. C., & Petrone, P. (2026). Deep Learning Predicts Hematopoietic Stem Cell Aging From 3 D Chromatin Images. Aging Cell, 25(10), Article e70656. https://doi.org/10.1111/acel.70656

Image Credits: AI Generated

DOI: 10.1111/acel.70656

Keywords: hematopoietic stem cells, chromatin architecture, deep learning, convolutional neural networks, biological aging, 3D confocal microscopy, epigenetic clocks, nuclear organization, explainable AI, drug screening, rejuvenation, H3K9 methylation