AI Reads 3D Chromatin Images to Predict Stem Cell Aging and Screen Rejuvenation Drugs
A deep learning model trained on 3D chromatin images can distinguish young from aged blood stem cells — and flag when epigenetic drugs reverse that aging signature.
Summary
Blood stem cells decline with age, and scientists lack good ways to measure that decline in individual cells. Researchers trained an image-analysis AI, ChromAgeNet, on 3D microscope pictures of DNA-stained nuclei from young and old mouse blood stem cells. It told young from old cells with moderate accuracy (AUROC 0.77) and beat traditional machine learning methods. Explainability tools showed what it looked at. Young cells had a thin, continuous rim of packed DNA along the nuclear edge, while aged cells had larger, irregular DNA clumps. When aged cells were treated with four experimental drugs, the model scored two that target histone methylation (UNC0646 and IOX1) as nearly as youthful as young cells. This is early mouse work, but it points toward image-based screening for rejuvenation therapies.
Detailed Summary
The blood-forming system weakens with age, and hematopoietic stem cells (HSCs) sit at the root of that decline, contributing to immune dysfunction and reduced healthspan. Epigenetic clocks based on DNA methylation are well established, but they mostly read bulk tissue. Chromatin architecture within the nucleus is a recognized hallmark of aging, yet it has rarely been used as a quantitative, single-cell readout. This study asks whether 3D images of nuclei alone can reveal a stem cell's biological age.
The team collected 1,229 confocal images of DAPI-stained HSC nuclei from young (10–16 weeks) and aged (>80 weeks) C57Bl6 mice. After segmentation, the dataset contained 552 young and 551 aged nuclei, which were split into 81,643 2D slices. Their convolutional neural network, ChromAgeNet, scores each slice. A soft-voting step then combines slice scores into one calibrated nucleus-level "youthful score." Performance was assessed with fivefold cross-validation split at the nucleus level.
At the slice level, AUROC was 0.72 ± 0.02 (accuracy 0.66, F1 0.64). Aggregating across the Z-stack raised nucleus-level AUROC to 0.77 ± 0.03 (accuracy 0.68, F1 0.69), with the best fold reaching 0.81. Sensitivity was higher for young cells (78%) than aged cells (61%). Beta calibration produced reliable probabilities (expected calibration error 0.02, Brier score 0.22). A calibrated XGBoost model trained on 70 handcrafted texture, intensity, and shape features reached AUROC 0.73 ± 0.04, so the CNN outperformed it. Top features included size zone nonuniformity and sum entropy. Acquisition variables such as operator, noise, and intensity explained almost none of the predictions (R² ≤ 0.03).
Explainable AI methods showed what the network used. Occlusion sensitivity gave the most reliable attention maps. In young HSCs, attention concentrated along the nuclear envelope, where a thin, continuous DAPI line likely reflects lamina-associated heterochromatin. In aged HSCs, attention fell on large, rounded DAPI blobs and a thicker, irregular envelope signal. Aged-cell attention maps had higher Shannon entropy, meaning more dispersed signals. Attention peaked about 30 pixels (1.5 μm) from the nuclear border.
As a proof of concept, the authors treated aged HSCs with four compounds. Baseline aged cells scored 0.34 ± 0.14. The Rho GTPase inhibitors CASIN and Rhosin modestly raised scores to 0.40 ± 0.18 and 0.48 ± 0.19. The H3K9 methylation modulators UNC0646 and IOX1 reached 0.56 ± 0.13 and 0.55 ± 0.16, closely matching untreated young HSCs at 0.55. This suggests image-based scoring could support high-throughput rejuvenation screens.
Several caveats apply. All data are from mice, and an AUROC of 0.77 is useful but far from definitive for individual cells. The model misclassifies more aged cells than young ones. Chromatin changes in normal aging overlap with those in senescence, so the two are hard to separate from images alone. The score shows that chromatin looks younger, not that the cells function better. Whether the findings carry over to human HSCs remains untested in the portion of the paper available here.
Key Findings
- Nucleus-level AUROC of 0.77 ± 0.03 (accuracy 0.68, F1 0.69) for classifying young vs aged mouse HSCs; best fold reached 0.81
- Soft-voting aggregation across the 3D stack improved AUROC from 0.72 (slice level) to 0.77 (nucleus level)
- ChromAgeNet beat the best classical ML model on 70 handcrafted chromatin features (XGBoost AUROC 0.73 ± 0.04)
- Sensitivity was 78% for young HSCs vs 61% specificity for aged HSCs; calibration was good (ECE 0.02, Brier 0.22)
- Occlusion maps showed young HSCs rely on a thin nuclear-envelope heterochromatin rim, while aged HSCs show large, irregular DAPI blobs and higher attention entropy
- Aged HSCs scored 0.34 ± 0.14; UNC0646 (0.56 ± 0.13) and IOX1 (0.55 ± 0.16) shifted scores to the level of young HSCs (0.55)
- CASIN (0.40 ± 0.18) and Rhosin (0.48 ± 0.19) produced smaller increases in youthful score; technical confounders explained R² ≤ 0.03 of predictions
Methodology
The study used 1,229 3D confocal DAPI images from young (10–16 weeks) and aged (>80 weeks) C57Bl6 mice, yielding 552 young and 551 aged HSC nuclei and 81,643 2D slices. A CNN was trained on slices, with nucleus-level predictions made by soft voting and beta calibration. Fivefold cross-validation split at the nucleus level was used, with AUROC, accuracy, F1, ECE, and Brier score as metrics. Benchmarks included 32 combinations of classical classifiers and feature selection on 70 radiomic features, interpreted with SHAP, activation maximization, and occlusion sensitivity.
Study Limitations
The model's discrimination is moderate and less sensitive for aged cells, and the data come from mice imaged in a single lab with different operators contributing different cell counts. Chromatin changes in aging overlap with senescence, making them hard to disentangle from images alone. The drug experiment is only a proof of concept, and the available text reports no functional validation or conflict-of-interest statement.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: