ImAge quantitates aging and rejuvenation

Abstract

For efficient, cost-effective and personalized healthcare, biomarkers that capture aspects of functional, biological aging, thus predicting disease risk and lifespan more accurately and reliably than chronological age, are essential. We developed an imaging-based chromatin and epigenetic age (ImAge) that captures intrinsic age-related trajectories of the spatial organization of chromatin and epigenetic marks in single nuclei, in mice. We show that such trajectories readily emerge as principal changes in each individual dataset without regression on chronological age, and that ImAge can be computed using several epigenetic marks and DNA labeling. We find that interventions known to affect biological aging induce corresponding effects on ImAge, including increased ImAge upon chemotherapy treatment and decreased ImAge upon caloric restriction and partial reprogramming by transient OSKM expression in liver and skeletal muscle. Further, ImAge readouts from chronologically identical mice inversely correlated with their locomotor activity, suggesting that ImAge may capture elements of biological and functional age. In sum, we developed ImAge, an imaging-based biomarker of aging with single-cell resolution rooted in the analysis of spatial organization of epigenetic marks.

Extended Data Fig. 1 |. ImAge accuracy measurements in blood.

a, b, Peripheral blood mononuclear cells (PBMCs) were isolated from mice aged 2, 5, 9, 15, 21, 23, 32 months (n = 2 for all groups), and immunolabeled for CD3 and H3K4me1 and colabelled with DAPI. Lineplots: Accuracy measurements versus bin size split by train and test data. 95% Confidence intervals are shown. Measurements were made for all models constructed: a centroid-based b, support vector machine SVM data.

Extended Data Fig. 2 |. ImAge versus orthogonal distance plots.

a, b, Peripheral blood mononuclear cells (PBMCs) were isolated from mice aged 2, 5, 9, 15, 21, 23, 32 months (n = 2 for all groups), and immunolabeled for CD3 and H3K4me1 and colabelled with DAPI. Scatter-plots show ImAge (normalized 0–1) versus orthogonal distance to the ImAge axis. ImAge was calculated using either a, centroid- or b, support vector machine (SVM)-based method. c, Performance comparison of Euclidean multidimensional scaling (EMDS) and hyperbolic multidimensional scaling (HMDS) within a 12-dimensional space. HMDS demonstrates significantly reduced distortion and uncertainty of distances (R² = 0.99) following the embedding process in hyperbolic space as compared to EMDS (R² = 0.67). This outcome supports the notion that our data exhibits an inherent hyperbolic structure.

Extended Data Fig. 3 |. Age-related ImAge progression in multiple solid organs.

ImAge readouts of the brain, cardiac muscle (heart), kidney, liver, and skeletal muscle (quadriceps) collected from three differentially aged cohorts of mice: 2 months (n = 5) 15 months (n = 4) and 27 months (n = 4). Two plates were analyzed, both immunolabeled with H3K27ac+DAPI and then with either H3K27me3 or H3k4me1. Data for overlapping channels (DAPI & H3K27ac) were combined for computations. Boxplots of the test set of bootstrapped data are shown with the normalized ImAge readouts from maximum-minimum to 0–1. Differences of means were calculated via Tukey’s HSD. Significance values for all tests shown represent: * = p < 0.05, ** = p < 0.01, *** = p < 0.001, **** = p < 0.0001

Extended Data Fig. 4 |. ImAge accuracy measurements in multiple solid organs.

a–e, Nuclei isolated from solid organs of young (2 months, n = 4) and aged (27 months, n = 4) control mice were immunolabeled and imaged with two sets of antibodies: (H3K27ac & H3K4me1), (H3K27ac & H3K27me3), both colabelled with DAPI. Test accuracy versus bin size was shown with 95% Confidence intervals. Dashed or solid lines represent plate/antibody set of origin: (H3K27ac & H3K4me1), (H3K27ac & H3K27me3), respectively.

Extended Data Fig. 5 |. Separation between different cell types with ageing, including the Brain.

a, 2-dimensional Euclidean multidimensional scaling (EMDS) of young (2 months) and old (27 months) liver, kidney, quads, and heart and brain. The observed clustering pattern reveals that the brain tissue cluster maintains a distinct separation from the clusters representing other tissue types, such as Kidney, Liver, Skeletal Muscle, and Cardiac Muscle. b,c, Silhouette scores of 5 organs at indicated ages for individual marks (b) or their combination (c) using the information distance metric. The Silhouette scores do not indicate a decline in tissue differentiation with ageing in the presence of brain tissue, likely due to a slower progression of ageing in the brain.

Extended Data Fig. 6 |. ImAge accuracy metrics for the interventions in biological age.

a–c, Nuclei from young and aged control mice were immunolabeled for H3K27ac and either H3K9me3 (a and b), or H3K27me3 and H3k4me1 (c) colabelled with DAPI. Lineplots: Accuracy measurements versus bin size split by train and test data. 95% Confidence intervals are shown. a, Accuracy measurements in Caloric Restriction (CR) dataset for separating 3 month (n = 3) and 24 month (n = 3) old mice at various bin sizes. b, Accuracy measurements in Doxorubicin (DXR) dataset for separating 1 month (n = 4) and 27 month (n = 4) old mice at various bin sizes. c, Accuracy measurements in induced hepatocarcinoma (tumor) dataset for separating 2 month (n = 5) and 8 month (n = 6) old mice at various bin sizes. Dashed or solid lines represent plate/antibody set of origin: (H3K27ac & H3K4me1), (H3K27ac & H3K27me3).

Extended Data Fig. 7 |. Statistics of the behaviors and the ImAge.

a, Performance comparison of Euclidean Multidimensional Scaling (EMDS) and Hyperbolic Multidimensional Scaling (HMDS) within a 9-dimensional space. HMDS demonstrates significantly reduced distortion of distances (R² = 0.87) following the embedding process in hyperbolic space as compared to EMDS (R² = 0.35). b, The ImAge distribution along the ageing geodesic in hyperbolic space and the orthogonal distance distribution to the same ageing geodesic for skeletal muscle cells in chronologically identical mice (25 months old) utilized for behavior performance evaluation. The ImAge axis captures ~82% of the total variance (the orthogonal projections ~18% of the variance). c, d, Clustering of individual behaviors; Person correlation (left) and significance (right)., Orthogonal bases from individual behaviors were used to identify unique linear coefficients. The criterion for behaviors clustering together was a high (>0.7) and significant (p < 0.05) Pearson correlation between the behaviors. e, Correlation between ImAge and individual orthogonal clusters of behaviors. f, Histogram of the test statistics for the correlation between ImAge and the optimized linear combination of behaviors. g, Correlations between ImAge and the individual clusters of behaviors. h, Histogram of the test statistics for the correlation between ImAge and the linear combination of individual orthogonal clusters of behaviors (without optimization). i Scatter plot for the correlation between ImAge and the linear combination of individual orthogonal clusters of behaviors (without optimization).

Extended Data Fig. 8 |. ImAge detected partial reprogramming of multiple organs.

The chronological ages of young and aged mice are 3.2 and 13.8 months, respectively. a, A graphical representation of the experimental condition: aged mice were treated with doxycycline to overexpress OSKM factors to evaluate the degree of reprogramming on the liver, quadriceps and cardiac muscle (heart). b-d (left), Distribution of ImAge obtained from 100 iterations of the test procedure using the linear support vector machine. Round-shaped symbols on the right side of violin plots for each age group represent the mean ImAge of each mouse sample. b-d (right), Sample-wise comparison of ImAge. Each violin plot represents the distribution of the ImAge for each mouse sample with the numbers from one to five for young, Aged-OSKM and aged groups. In all panels, an asterisk indicates the significant decrease from the aged mouse showed the lowest ImAge in the aged group with respect to the median ImAge (Mann-Whitney U-test, significant threshold; p < 0.05).

Extended Data Fig. 9 |. The degree of partial reprogramming reported by ImAge varies depending on tissues and epigenetic marks.

The chronological ages of young and aged mice are 3.2 and 13.8 months, respectively. a, Each violin plot represents the distribution of the ImAge for each mouse sample with the numbers from one to five for young, Old-OSKM and aged groups. The Left and right columns are for the liver and muscle, respectively. In all panels, an asterisk indicates the significant decrease from the aged mouse showed the lowest ImAge in the aged group with respect to the median ImAge (Mann-Whitney U-test, p < 0.05). b, c, proportion of cells with young and aged signature ImAge in each age group in the liver (b) and muscle (c) (Mann-Whitney U-test, p < 0.05). Partial reprogramming after doxycycline-induced OSKM factors (i4F mice, old-OSKM) was observed in both HNF4a+ and HNF4- cell populations using indicated marks.

Extended Data Fig. 10 |. Overview of young/aged indicative cells revealed at the single-cell level in liver samples following partial reprogramming in vivo.

Uniform manifold approximation and projection (UMAP) of three-dimensional threshold adjacency statistic (TAS) features for single cells. The green and brown data points represent young and aged indicative cells, respectively. The gray points represent the intermediate single cells.

Fig. 1 |. Emergence of chromatin trajectories of aging.

a, Graphical representation of the microscopic imaging of epigenetic landscapes workflow: nuclear isolation, immunofluorescence imaging, image preprocessing, nuclear segmentation, texture feature extraction and downstream analysis. b,c, Multidimensional scaling plots of CD3⁺ and CD3⁻ subsets of PBMCs from C57BL/6NJ males aged from 1.7 to 32 months (1.7, 2.2, 5.3, 8.7, 15.1, 21, 22.3 and 32.2) (n = 2 per age group) based on 2D EMDS (b) and 3D HMDS (c) of 504 texture features. Each axis represents the score obtained from EMDS or HMDS indicated as EMDS1 and EMDS2 and HMDS1–HMDS3. Each data point was colored with its chronological age based on the color scale on the right side of the plot. d,e, Top, graphical representation of the method of ImAge axis construction; bottom, scatter-plots of ImAge versus chronological age for CD3⁺ subset (top), the CD3⁻ subset (middle) and the whole population (PBMC) (bottom). d, ImAge axis constructed using the geodesic connecting the centroids of the youngest and oldest groups in Euclidean or hyperbolic space. e, ImAge axis constructed using linear SVM fit to the youngest and oldest groups. Pearson’s correlation (r) between ImAge and chronological age and P values.

Fig. 2 |. Age-related ImAge progression in multiple solid organs.

a–e, ImAge readouts of the brain, cardiac muscle (heart), kidney, liver and skeletal muscle (quadriceps) collected from three differentially aged cohorts of mice: 2 months (n = 5), 15 months (n = 4) and 27 months (n = 4). Two plates were analyzed, both immunolabeled with H3K27ac + DAPI and then with either H3K27me3 or H3k4me1. Data for overlapping channels (DAPI and H3K27ac) were combined for computations. Boxplots of the test set of bootstrapped data are shown with the normalized ImAge readouts from maximum-minimum to 0–1. Differences of means were calculated via Tukey’s HSD. f, Statistically significant (P < 0.005) ImAge feature weights for H3K27ac were compared across all tissues (left) and pairwise Spearman correlations were calculated (right). Bracket and box annotations were added to highlight strong and significant correlations between heart, kidney and liver correlations. Note that for visual representation (left) H3K27ac-based ImAge features were ranked based on average correlation for all five tissues. g,h, Statistically significant (P < 0.005) ImAge feature weights for liver (g) and quads (h) were compared across all channels (left) and pairwise Spearman correlation matrices were calculated (right). Note that for visual representation (left), skeletal muscles (quads) and liver ImAge features for all epigenetic marks were ranked based on their correlation between H3K27ac and H3K4me1 marks (that is a different order compared to that in f). Bracket and box annotations were added to highlight strong and significant correlations between H3K27ac and H3K4me1. Spearman’s R and P values with Bonferroni correction for multiple comparisons. Significance values for all tests shown represent: *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001; NS, not significant.

Fig. 3 |. Age-related erosion of cellular states based on information distance metrics.

a, A 2D EMDS of young (1.7 months) and old (32.2 months) CD3⁺ and CD3⁻ subsets of PBMCs. b,c, Silhouette scores of CD3⁺ and CD3⁻ subsets at indicated ages for individual marks (b) or their combination (c) using the information distance metric (based on mutual information and Shannon entropy, see Methods). d, the KS distance analysis of CD3⁺ and CD3⁻ subsets across indicated ages performed on significant features. e, A 2D EMDS of young (2 months) and old (27 months) liver, kidney, quads and heart. f,g, Silhouette scores of five organs at indicated ages for individual marks (f) or their combination (g) using the information distance metric. h, the KS distance analysis of four organs across indicated ages performed on significant features. In d and h, The significant features were selected based on (1) statistically significant (P < 0.05, Pearson |r| > 0.85) KS distances between cell types and (2) a statistically significant (P < 0.05, Pearson |r| > 0.95) correlation between KS distance and age.

Fig. 4 |. Diet, chemotherapy and cancer affect ImAge.

a–c, ImAge calculations separated by epigenetic marks (left) and several marks combined (right) for indicated conditions for caloric restriction, doxorubicin (DXR) or induced hepatocarcinomas (tumor). Young (3 months, n = 4) and old (24 months, n = 4) control mice were used to construct an ImAge axis upon which CR (7 months, n = 4) and control (7 months, n = 4) ImAge values were measured (a). Young (1 month, n = 3) and old (27 months, n = 3) control mice treated with PBS were used to construct an ImAge axis upon which young DXR-treated mice (1 month, n = 4) ImAge values were measured (b). Liver tissue from old mice (8 months, n = 3) with induced tumors was separated by the presence or absence of tumors (c). Normal old tissue along with young control mice (2 months, n = 3) was used to construct an ImAge axis upon which old tumor ImAge values were measured. In c, two plates were analyzed, both immunolabeled with H3K27ac+DAPI and then with either H3K27me3 or H3k4me1. Data for overlapping channels (DAPI and H3K27ac) were combined for computations. Significance was calculated using Tukey’s HSD. All ages shown are in months. P value cutoffs are *0.01 < P ≤ 0.05; **0.001 < P ≤ 0.01; ****P ≤ 0.0001. ImAge axis correlates and explains approximately half of the changes induced by DXR and CR. d, heatmap of 192 statistically significant texture features shared by ImAge and DXR axes. e. A heatmap of 232 statistically significant texture features shared by ImAge and CR axes. The color bar indicates the direction (blue-negative, red-positive) and the degree (color intensity) of correlation. Statistically significant texture features are features with non-zero weight using a one-sample t-test against 0 mean value, P < 0.005. See text for the axes correlation statistics.

Fig. 5 |. ImAge inversely correlates with locomotor activity in chronologically identical mice.

a, A 3D representation of HMDS and its 2D projection (view from the top) of the young (2 months) and old (27 months) mouse skeletal muscle (quadriceps) samples utilized as references to obtain centroids for the ImAge axis. b, ImAge distributions of the reference samples (young and old) with chronologically identical (25 months) mice from the behavioral cohort. c, The nine orthogonal clusters of behavior with the coefficients for linear optimization correlating behavioral/functional readouts and ImAge. The direction of ImAge association with each cluster (older and younger) is proportional to the cluster’s α_i. d, Correlations between ImAge and a linear combination of all behavioral readouts (top) and locomotor activities only, clusters 1, 3, 4 and 7 (bottom).

Fig. 6 |. ImAge reveals heterogeneity of partial reprogramming in vivo.

The chronological ages of young and aged mice are 3.2 and 13.8 months, respectively. a, Evaluating the degree of reprogramming after doxycyclineinduced OSKM factors (i4F mice, aged-OSKM) in the liver and muscle. b,c, Distribution of ImAge (100 iterations of the test is shown) in the liver (b) and muscle (c). Dots on the right side of violin plots are the mean ImAge of individual mouse samples. d,e, Violin plots representing the distribution of the ImAge within individual samples in the liver (d) and muscle (e). Statistically significant differences were assessed between all aged-OSKM samples and the aged mice sample with the lowest ImAge (the ‘youngest’ aged mouse). f, Distribution of ImAge at a single-cell level. The mean accuracy of the young and old segregation was 0.620 ± 0.001 (100 iterations). The young and old indicative cells are defined by the threshold of ImAge readout at 5th (and lower) and 95th (and higher) percentile values of aged and young single-cell ImAge readout, respectively. g,h, The proportion of young/aged indicative cells in the liver (g) and muscle (h) for each animal. The dotted and dashed lines indicate the reference proportions of young and aged indicative cells, respectively, in the ‘youngest’ aged mouse. i, Uniform manifold approximation and projection (UMAP) of single-cell texture features obtained from the liver samples. The green and brown data points represent the single cells with young and ageds ImAge signatures, respectively. The intermediate single cells are in gray. P < 0.05, Mann–Whitney U-test.