Main

Aging is the progressive deterioration of cellular and organismal function. Age-dependent decline is linked in large part to the passage of time and therefore the chronological age of an individual. But such decline is not inexorable. At the same chronological age, some individuals have better organismal and tissue fitness (biological age) than others. Furthermore, aging trajectories can be slowed, and some aspects of aging can be reversed by specific interventions, including dietary restriction, exercise, reprogramming factors, senolytic compounds and young blood factors1,2,3,4,5,6. As aging is the primary risk factor for many diseases, particularly neurodegenerative diseases7,8, a better understanding of aging and ‘rejuvenation’ strategies could yield large benefits for a wide range of diseases.

Aging is complex and difficult to quantify. One quantification approach is to use machine learning to build age prediction models—‘aging clocks’—which can serve as integrative aging biomarkers. Such clocks should also accelerate our understanding of existing interventions and help identify new strategies to counter aging and age-related diseases. Machine learning models trained on high-dimensional datasets (for example, DNA methylation, transcriptomics and proteomics) can predict chronological age with remarkable accuracy. For example, regression-based aging clocks trained on DNA methylation profiles from multiple tissues (‘epigenetic aging clocks’)9,10,11,12,13 or blood plasma protein profiles14,15,16,17 have striking performance to predict chronological age in humans. Aging clocks directly optimized to predict biological age have also been developed on functional phenotypes12,13,18 or time remaining until death19,20. Interestingly, beneficial health interventions such as diet and exercise21,22,23 and genetic manipulations24,25,26 result in younger predictions from epigenetic aging clocks trained on chronological age. Thus, epigenetic aging clocks, despite being trained on chronological age, also capture dimensions of biological age.

So far, molecular aging clocks have largely relied on datasets built using bulk tissue input or purified cell populations9,10,11,12,13,27,28,29,30,31,32,33,34. Bulk tissue profiles (and even purified populations) average the molecular profiles from many cells, integrating tissue composition changes and cell-type-specific responses. Hence, the cell-type-specific contributions to aging and rejuvenation detected by these clocks remain unclear. While single-cell DNA methylation and transcriptomic data have started to be used to classify age35,36,37, cell-type-specific transcriptomic aging clocks have not yet been generated. Thus, it remains to be determined if aging clocks of different cell types ‘tick’ at different rates, which cell types predict age most accurately and how specific cell types respond to different interventions. The rapid advance of single-cell RNA-sequencing (RNA-seq) technologies provides a great opportunity to explore these unaddressed questions and identify new molecular aging clocks to study interventions to counter aging and age-related diseases.

Results

Cell-type-specific transcriptomic aging clocks

As a paradigm for tissue aging and functional decline in the brain, we focused on the neurogenic region located in the subventricular zone (SVZ) of the adult mammalian brain. The SVZ neurogenic region (or ‘niche’) contains neural stem cells (NSCs) that give rise to differentiated cells (neurons, astrocytes) that are important for olfactory discrimination and repair upon injury38,39,40,41,42,43,44,45. Importantly, this neurogenic region contains at least 11 different cell types and experiences age-related changes correlated with deterioration in tissue function42,46,47,48,11,21,22,23,24,31,32. Our results also highlight cell-type specificity for aging and possibly for rejuvenation interventions. This is unique to single-cell-based clocks and will allow a better understanding of cell heterogeneity in tissue aging and rejuvenation. Our data also reveal different potential for rejuvenation strategies, at least at the transcriptional level. These results raise the exciting possibility that aging clocks can serve to rapidly test the efficacy of rejuvenation interventions and to support combining specific interventions to counter aging and age-related diseases.

Methods

Our research complies with all relevant ethical regulations (AAALAC), under Institutional Animal Care and Use (IACUC) protocols 8661 and 16246 at Stanford University and VA Palo Alto Committee on Animal Research, ACORP LUO1736.

Animals

For aging cohorts and the exercise cohort, male C57BL/6 mice were obtained from the National Institute on Aging (NIA) Aged Rodent colony. For parabiosis cohort 1, old mice were male C57BL/6 mice from the NIA Aged Rodent colony and young mice were male B6.SJL-Ptprca Pepcb/BoyJ male (Pep boy) from the Jackson Laboratory. For parabiosis cohort 2, old mice were male C57BL/6J and young mice were male C57BL/6J or C57BL/6-Tg(UBC-GFP)30Scha/J from the Jackson Laboratory. Mice were housed in the Comparative Medicine Pavilion, ChemH/Neuroscience Vivarium or the SIM-1 Non-Barrier Rodent Facility at Stanford, or in the Veterinary Medical Unit at the Palo Alto VA. All these facilities provide equivalent standard conditions with a 12-h light–dark cycle, ad libitum food and water, ~21 °C temperature, and ~50% humidity. All mice were acclimated to their vivarium for at least 2 weeks before use in any experiment.

Tissue and cell collection for the subventricular zone neurogenic niche

For single-cell RNA-seq datasets, SVZ neurogenic niches were collected and processed as described in ref. 48. Briefly, mice were sedated with 1 ml of 2.5% vol/vol Avertin (Sigma-Aldrich, T48402-25G) and perfused with 15 ml of PBS (Corning, 21-040-CV) with heparin sodium salt (50 U ml−1; Sigma-Aldrich, H3149-50KU) to remove the blood, and brain collection was performed immediately. As previously described104, the SVZ from each hemisphere was microdissected and dissociated with enzymatic digestion with papain at a concentration of 14 U ml−1, rocking for 10 min at 37 °C. Note that the samples also contained some of the surrounding striatum, which contributed to the oligodendrocyte population in our study. The dissociated SVZ was triturated in a solution containing 0.7 mg ml−1 ovomucoid and 0.5 mg ml−1 DNase I (Sigma-Aldrich, DN25-100MG) in DMEM/F12 (Thermo Fisher, 11330032). The dissociated cells from the SVZ were centrifuged through 22% Percoll (Sigma-Aldrich, GE17-0891-01) in PBS to remove myelin debris. After centrifugation, cells were filtered through a 35-μm snap-cap filter (Corning, 352235), washed once with 1.5 ml of FACS buffer (HBSS (Thermo Fisher, 14175103), 1% BSA (Sigma, A7979) and 0.1% glucose (Sigma-Aldrich, G7021-1KG)) and spun down for 5 min at 300g. Cells were resuspended in 120 μl FACS buffer with live/dead staining performed using 1 μg ml−1 propidium iodide (BioLegend, 421301) and kept on ice until sorting. FACS sorting was performed on a BD FACS Aria II sorter, using a 100-μm nozzle at 13.1 PSI. Cells were sorted into low protein binding microcentrifuge tubes containing 750 μl of PBS with 1% BSA and 0.1% glucose. When not applying sample multiplexing (parabiosis cohort 1 and exercise cohort), cells were then centrifuged (300g for 5 min at 4 °C) and resuspended in 50 μl FACS buffer, counted and then immediately run on 10x Chromium to capture single-cell transcriptomes.

Cohorts of mice of different ages

To generate the single-cell RNA-seq dataset from mice of different ages and train aging clock models, we used four independent cohorts of aging mice. Each cohort had 4–8 male C57BL/6 mice from the NIA Aged Rodent colony, for a total of 28 mice. These 28 mice tiled 26 different ages (two pairs of mice had the same age), ranging from 3.3 months (young adult) to 29 months (geriatric adult).

Lipid-modified oligonucleotide multiplexing

Sample multiplexing was performed using LMOs, a method also known as MULTI-seq50. Lipid anchor and co-anchor reagents were kindly provided by the Gartner Laboratory at the University of California, San Francisco and custom oligonucleotides were ordered from Integrated DNA Technologies. We used MULTI-seq primer: 5′ CTTGGCACCCGAGAATTCC; and Universal.I5: 5′AATGATACGGCGACCACCGAGATCTACACTCTTTCCCTACACGACGCTCTTCCGATCT50.

We followed the exact protocol outlined by McGinnis et al.50 with the following modifications: (1) all labeling with LMOs was performed in a 4 °C cold room because, in our hands, the quality of labeling was very sensitive to temperature; (2) to avoid cell loss and cell clum**, cells were sorted into PBS with 2% BSA, and BSA was then removed using three PBS washes; (3) concentrations and volumes were adjusted to account for low cell numbers: 7.5 μl of 1 mM lipid anchor with oligonucleotide barcode mix was added to a 70 μl volume of resuspended cells followed by 7.5 μl of 1 mM lipid co-anchor; (4) labeling reactions were quenched with 2% BSA then samples were pooled before subsequent 1% BSA PBS washes to further reduce cell loss. The combined sample was resuspended at 50 μl for cell counting and single-cell RNA-seq.

Single-cell libraries and RNA sequencing

Single-cell RNA-seq was performed using a 10x Chromium machine and 10x Genomics V3.0 Transcriptomics kits (aging cohorts, parabiosis cohort 2 and exercise cohort) or a 10x Genomics V2 kit (parabiosis cohort 1). For sequencing, 10,000 cells per lane were targeted but typical yields were approximately 5,000 cells. Library preparation was done according to the manufacturer’s protocol (10x Genomics V3.0 or 10x Genomics V2 for parabiosis cohort 1). Sequencing was done to target a minimum of 25,000 reads per cell for transcriptome characterization and 5,000 reads per cell for LMO label recovery. The aging cohorts and the parabiosis cohort 2 samples were multiplexed with 4–8 samples per 10x Chromium lane. The parabiosis cohort 1 and the exercise samples were not multiplexed with LMO reagents. Sequencing was performed on either an Illumina HiSeq 4000 (aging cohorts and parabiosis cohort 1) or a NovoSeq using the 2 × 150-bp setting (parabiosis cohort 2 and exercise).

Analysis (quality control)

Cell Ranger (version 3.0.2) default settings were used to distinguish cells from background. Subsequent analysis was performed using R (version 3.6.3). Cells were filtered out in Seurat (version 3.2.3)105,106 if they contained fewer than 500 genes or greater than 10% mitochondrial reads. Small clusters of doublets that shared several marker genes from pure populations were identified and removed. LMO demultiplexing was performed using Seurat’s HTODemux function. A complete view of the data processing and quality-control parameters can be found at https://github.com/sunericd/svz_singlecell_aging_clocks.

Cell type annotation

Cell types in all datasets were manually annotated as described in ref. 48, and cross-referenced with annotations present in the single-cell database PanglaoDB107. Identification of major clusters was performed with the FindClusters() algorithm in the Seurat package, which uses a shared nearest-neighbor modularity optimization-based clustering algorithm106. Marker genes for each major cluster were found using the Seurat (version 4.1.1) function FindAllMarkers() using the Wilcoxon rank-sum test. Cell types were determined using marker genes identified from the literature and the marker genes were cross-referenced with annotations present in the single-cell database PanglaoDB107. This analysis identified ~11 clusters of cells (depending on the dataset), including astrocytes and qNSCs, aNSCs and NPCs, neuroblasts, neurons, oligodendrocyte progenitor cells, oligodendrocytes, endothelial cells, ‘mural’ cells (pericytes or smooth muscle) and microglia. The genes used for identification are included in Supplementary Table 2 and a clustering of a subset of these genes is presented in Extended Data Fig. 1c.

Consistent with our previous study48, we did not observe sufficient differences in transcriptomic signatures to separate astrocytes from qNSCs and aNSCs from NPCs. We have described these clusters as ‘astrocyte-qNSCs’ and ‘aNSC-NPCs’ throughout this study. Some cell types were not identified when using the LMO protocol (for example, T cells), probably because cells such as T cells are small and their membranes may not allow for efficient LMO labeling. We also identified only a few ependymal cells in several of our datasets, although these cells are known to be numerous in the SVZ neurogenic niche. This is probably because ependymal cells are too big to be efficiently uploaded in droplets and/or they are sheared in the 10x microfluidic device.

Cell cycle annotation and proliferative fraction

For cell cycle annotation (G1, S, G2/M) of cells in the SVZ neurogenic niche, we used Seurat’s CellCycleScoring function with default parameters. This annotation was used to calculate the ‘proliferative fraction’ in the SVZ neurogenic niche, that is, the percentage of cells predicted to be in S or G2/M phase. We used the proliferative fraction (ProliferativeFraction) as a functional metric of the SVZ neurogenic niche and used it to define ‘biological age’ in this study (‘Age prediction and validation strategy’).

To test the correlation between chronological age and proliferative fraction in the SVZ neurogenic niche, we used Pearson’s correlation. There was a negative correlation (Pearson R = −0.8) between chronological age and proliferative fraction in the SVZ.

Age prediction and validation strategy

Chronological or biological age (‘label’) was regressed onto all log-normalized gene expression values ln((gene transcripts / cell transcripts) × 10,000) (‘features’) in a particular cell type using the R package glmnet (version 4.0.2)51. To determine the most robust method to predict age from single-cell RNA-seq data, we tested various preprocessing approaches: SingleCell, Pseudobulk, BootstrapCell (‘BoostrapCell preprocessing’) and EnsembleCell (‘EnsembleCell preprocessing’). SingleCell uses bona fide single-cell transcriptomes with minimal processing as input to a lasso regression model to predict chronological or biological age. Pseudobulk involves naïve pseudobulking all cells from the same cell type and sample before using a lasso regression model to predict chronological or biological age. BoostrapCell uses lasso regression models and EnsembleCell uses elastic net models (described separately below)53. There was no manual filtering of genes. Both lasso regression and elastic net regression enforce sparsity in the model coefficients with tunable parameter such that only a subset of genes will have nonzero coefficients in the trained aging clock models.

Chronological age was defined as months since birth. Biological age was defined as 35 – (ProliferativeFraction × 100) where ProliferativeFraction was the number of cells predicted to be in S or G2/M phase divided by the total number of cells from that sample. The number 35 was selected to transform biological age into the same range as chronological age.

For validation, models were built on 3 of the 4 cohorts of mice, and validation was done on the remaining cohort (stringent ‘leave-one-cohort-out’ validation (cross-cohort validation)). For training of each model, hyperparameters were optimized with fivefold to tenfold cross validation. To quantify the performance of the models, the data were presented as a correlation between the actual chronological (or biological) age of the mouse from which the cell originated (x axis) and the median predicted chronological (or biological) age for that mouse (y axis). Density of cells is represented with graded colors and each mouse is represented as a dot. We fitted a linear model (black line) through the points as well as the 95% confidence interval (light gray) using geom_smooth (ggplot2). Pearson’s correlation (R) is indicated on the graph. In dot plots, both the R values and the MAE, that is, median absolute error across all the cells, are presented.

To test the correlation between chronological age and biological age, we used the Pearson correlation. There was a positive correlation (R = 0.84) between chronological age and biological age predictions.

BootstrapCell preprocessing

BootstrapCell uses a lasso model with the following characteristics: To generate a BootstrapCell, 15 single-cell transcriptomes were sampled without replacement from the pool of cells of a given cell type from a given animal (for example, oligodendrocytes from a single mouse). Gene counts were then summed. A BootstrapCell constructed from 15 cells was empirically found to balance the tradeoff between sample number and gene coverage per sample. This bootstrap** process was repeated 100 times for each cell type–animal combination. BootstrapCells were used as input into lasso regression models. This approach had the effect of normalizing the contribution of each animal rather than each single-cell transcriptome.

EnsembleCell preprocessing

We devised and evaluated a second preprocessing and age prediction technique to compare to our BootstrapCell approach and to test robustness to changes in preprocessing and model architecture. In the EnsembleCell approach, 20 elastic net models were trained for each cell type. For each model, gene expression data from cells were randomly partitioned into groups of 15 single-cell transcriptomes and the unique transcript counts for all cells in each group were summed to create ‘EnsembleCells’. To predict age from the gene expression profile of a cell, we used the weighted average of predictions across all 20 models, where weights were determined by the R2 (coefficient of determination) of the model on a held-out validation set (‘Age prediction and validation strategy’).

Use of aging clocks on independent mouse datasets

We determined if the single-cell-based models (‘aging clocks’) generated from our mouse SVZ neurogenic niche dataset could be applied to cells from an independent dataset and even to cells from another neurogenic region in the brain. To this end, we used a single-cell RNA-seq dataset of the SVZ neurogenic niche from young and old mice48 and a single-cell RNA-seq dataset of the dentate gyrus of the hippocampus from mice of three different ages60. These datasets were preprocessed as described above using the ‘BootstrapCell’ method. We examined the distribution of the predicted chronological or biological ages of each cell in these datasets, color coded by the age of the mouse of origin.

Use of aging clocks on human datasets

To determine if the single-cell-based aging clocks generated from the mouse SVZ neurogenic niche could apply to cells from other regions of the brain and in other species, we used a single-nucleus RNA-seq dataset of the middle temporal gyrus from humans of different ages61. The dataset was preprocessed using the ‘BootstrapCell’ method as described above. As oligodendrocytes and astrocytes were present both in the human dataset and our mouse SVZ neurogenic niche dataset, we applied our oligodendrocyte and astrocyte-qNSC chronological aging clocks to the corresponding cell types in the human dataset. We rescaled the raw predictions linearly to obtain rescaled predicted chronological ages for each human BootstrapCell (rescaled predicted age = m × raw predicted age + b, where m = 10 and b = 125.5 for oligodendrocytes; m = 5 and b = 32.75 for astrocytes). The linear rescaling did not change the reported correlation between predicted chronological age and actual chronological age. Correlation plots were generated as described in ‘Age prediction and validation strategy’.

Cell-type-specific aging clocks using Tabula Muris Senis

To determine whether the method we used to derive cell-type-specific aging clocks was generalizable to tissues other than neurogenic niches, we used the count matrices from the single-cell RNA-seq dataset of the multi-tissue aging atlas Tabula Muris Senis62. We chose three diverse cell types in different tissues: endothelial cells from limb muscle, mature natural killer T cells from spleen and podocytes from kidney. For each cell type, the data were preprocessed and aging clocks were trained using the BootstrapCell approach described above. The performance of these models was evaluated by iteratively training on all mice except for one mouse and obtaining predictions on the held-out mouse (‘leave-one-mouse-out’ cross validation (cross-mouse validation) instead of ‘leave-one-cohort-out’ cross validation (cross-cohort validation) because there were no distinct cohorts in this dataset).

Identification of genes that contribute to the aging clocks

Genes that contribute to each aging clock model were retrieved by selecting all genes from the clocks with nonzero coefficients (Supplementary Table 4). The weight of a gene on each clock model (that is, the level of contribution based on coefficient values) and the sign of the coefficient (positive, higher gene expression is associated with older age; negative, lower gene expression is associated with older age) are indicated using a donut plot, with sector size indicating the gene weight and color indicating coefficient sign. Genes with positive coefficient are mostly upregulated with age, and genes with negative coefficient are mostly downregulated with age. The regulation of each chronological and biological clock gene (compared to other genes) is presented using a volcano plot (Extended Data Fig. 4). Most genes selected by the clocks were differentially expressed during aging. Less than half of the genes selected by chronological and biological aging clocks in a particular cell type overlapped (Supplementary Table 4). To determine if chronological or biological clock genes were shared across cell types or specific to each cell types, we used UpSet plots. Most genes selected by chronological or biological clocks were cell-type specific. The ‘impact’ (sum of absolute values of coefficient) and ‘count’ (sum of gene number) of shared genes or specific genes are indicated as a stacked bar plot.

Properties of genes that contribute to the aging clocks

To determine if genes that contribute to the aging clocks have specific properties, we examined their variability by plotting the coefficient of variation as a function of mean expression. Genes used by the clocks were more highly expressed and, at a given level of expression, had a higher coefficient of variation (that is, were more variable) than genes not in the clock (Extended Data Fig. 3a).

We also verified that the increased variability of genes that contribute to the clocks was not merely due to sparsity in the single-cell RNA-seq dataset. On average, the majority of cells (for each cell type) express the genes that contribute to the clocks and this is higher than what was observed for genes that do not contribute to the clock (Extended Data Fig. 3b).

Gene-set enrichment analysis

GSEA was performed using Enrichr108 to query cell-type-specific clock genes for enrichment against GO biological process gene sets. Statistics were exported from the Enrichr web tool and processed and visualized in R with ggplot2 (version 3.3.3) package.

Parabiosis cohorts and single-cell RNA-seq dataset

Two independent cohorts of heterochronic parabiosis were generated (cohort 1 and cohort 2). Parabiosis cohort 1 involved six male mice across three pairings. We collected SVZ niches from one isochronic young mouse (5 months, control), one heterochronic young mouse (5 months, old blood), one heterochronic old mouse (26 months, young blood) and one isochronic old mouse (26 months, control), for a total of four SVZ niches (of six mice). Old parabionts were C57BL/6 male mice from the NIA Aged Rodent colony at Charles River. Young parabionts were B6.SJL-Ptprca Pepcb/BoyJ male (Pep boy) mice from The Jackson Laboratory and C57BL/6 male mice from the NIA. Of the young, only the Pep boy mice were used for transcriptomics. Congenic (rather than isogenic) pairings were performed to enable verification of blood chimerism by FACS with antibodies specific to CD45.1 (BioLegend, 110705; 1:100 dilution) or CD45.2 (BioLegend, 109814; 1:100 dilution) alleles. Mice were 4 and 25 months old at the start of the experiment, and parabiosis was conducted for 5 weeks until cell collection, when mice were 5 and 26 months old. Pairs were established as previously described69,75,80 by suturing the peritoneums of adjacent flanks and joining skin with surgical clips. Five weeks after the parabiosis surgery, mice were anaesthetized with 2.5% vol/vol avertin, euthanized by cardiac puncture and perfused with 15 ml PBS with heparin (50 U ml−1). SVZ dissection, digestion and FACS were performed as describe above. 10x Genomics single-cell transcriptome V2 libraries (one sample per 10x lane) were generated and sequenced on one Illumina HiSeq lane by the Stanford Function Genomics Facility. Animal care and parabiosis procedures were performed in accordance with Stanford University under IACUC protocols 8661 and 16246.

Parabiosis cohort 2 involved eighteen male mice across nine pairings. We collected SVZ niches from four isochronic young mice (5 months, control), four heterochronic young mice (5 months, old blood), four heterochronic old mice (21 months, young blood) and six isochronic old mice (21 months, control), for a total of eighteen SVZ niches (of eighteen mice). All mice in this cohort were sourced from the Jackson Laboratory and housed in the Veterinary Medical Unit at the Palo Alto VA77. Old mice were C57BL/6J and young were C57BL/6J or C57BL/6-Tg(UBC-GFP)30Scha/J. Mice were aged 4 and 19.5 months at the start of the experiment, and parabiosis proceeded for 5 weeks until cell collection, when mice were 5 and 21 months old. Surgeries were performed as described above. Five weeks after surgery, mice were anesthetized with 2.5% vol/vol avertin, euthanized by cardiac puncture and perfused with 15 ml PBS with heparin (50 U ml−1). SVZ dissection, digestion and FACS were performed as describe above. Tissue collection took place on three separate days and samples were multiplexed with LMOs. 10x Genomics single-cell transcriptome V3 libraries were generated in-house and sequenced by Novogene on an Illumina NovoSeq lane. Animal care and parabiosis procedures were approved by the VA Palo Alto Committee on Animal Research and listed on ACORP LUO1736.

Parabiosis cohort 1 and cohort 2 were generated in different animal facilities, by different surgeons, in different years, and they were analyzed with different versions of 10x Genomics single-cell transcriptomics kits. For visualization, data from the two independent cohorts were integrated on the cohort identity using the RunHarmony command from Harmony109. There were no statistically significant differences between young isochronic (control) predicted chronological ages across cohorts in all six cell-type-specific aging clocks (Wilcoxon rank-sum test for median predicted chronological ages), suggesting that there was not a major batch effect that could have influenced the age prediction.

Exercise cohort and single-cell RNA-seq dataset

C57BL/6 male mice from the NIA Aged Rodent colony at Charles River were housed in the Veterinary Medical Unit at the Palo Alto VA97. Young and old mice were aged 4.5 months and 21.5 months, respectively, at the start of the 5-week voluntary wheel running intervention, so they were 6 months and 23 months when tissues were collected. During the intervention period, mice (n = 4 for each age group) were singly housed in cages accommodating a running wheel. Control mice (n = 3–4 for each age group) had no access to a wheel. Running was verified by recording wheel revolutions. After 5 weeks, mice were anaesthetized with 2.5% v/v avertin, euthanized by cardiac puncture, perfused and cell suspensions from dissected SVZs generated as described in ‘Tissue and cell collection for the SVZ neurogenic niche’. Next, 10x Genomics V3.0 transcriptomics kits were used to generated libraries without upstream sample multiplexing. Tissue processing occurred across two separate mornings. SVZ libraries were pooled and sequenced on an Illumina NovoSeq.

Effect of rejuvenation interventions on the aging clocks

To measure the effect of heterochronic parabiosis and exercise on the aging clocks, we examined the distribution of predicted chronological or biological ages as described in ‘Use of aging clocks on independent mouse datasets’. We calculated the effect by the difference in median predicted chronological or biological age between intervention and control. In dot plots, these differences were represented as ‘effect’, using size and intensity of color, with blue indicating ‘rejuvenation’ and red indicating ‘aging’.

Comparison of heterochronic parabiosis and exercise effects

To compare the effect of heterochronic parabiosis and exercise, we calculated the mean of the difference between the median predicted chronological age for a mouse for each intervention (data from cohort 1 and cohort 2 for heterochronic parabiosis). Genes that were reversed by each intervention or by both, based on direction of average log fold change, were identified.

Differential expression analysis

To determine genes that were impacted by different interventions independently of the aging clocks, we used differential expression analysis, focusing on aNSC-NPCs (as this cell type is impacted by both interventions). MAST110 software was used to calculate differential expression statistics between three different conditions: age (young versus old), young blood (heterochronic parabiosis versus isochronic old control), exercise (exercise versus sedentary in old mice). To determine the DEGs between young and old, we defined ‘young’ as mice <7 months and ‘old’ as mice >20 months. Permissive cutoffs of 1.1-fold change and FDR < 0.1 were applied in each of the three different conditions. Overlap was presented as a Venn diagram.

Gene signature analysis

For specific gene signature analysis, we summed the expression of genes in one cell type from single-cell transcriptomic datasets within a specific gene signature defined by a specific GO term. Among the different signatures tested, we selected those that were significantly increased with age and reversed by at least one intervention. We focused on two signatures: the ‘interferon-γ response’ signature defined as the sum of all normalized expression values of genes in the interferon gene set defined by Dulken et al.48 and the ‘negative regulation of neurogenesis’ gene signature defined as the sum of all normalized expression values of genes in the GO term ‘negative regulation of neurogenesis’ gene set (v6.21)111. Data were presented as violin plots and statistical analyses were performed using the Wilcoxon rank-sum test at the cell level.

Intervention classification models

To evaluate the aging relevance of ‘rejuvenation’ interventions, we generated cell-type-specific models trained on the intervention rather than age as a label. We used classification models, based on logistic regression (cv.glmnet(type.measure = ‘mse’, family = ‘binomial’) using all log-normalized gene expression values ln((gene transcripts / cell transcripts) × 10,000) as features. These intervention classification models were trained on single-cell RNA-seq data from heterochronic parabiosis (young blood) versus isochronic parabiosis old (control) or from exercise versus sedentary old mice. The data were preprocessed using the same BootstrapCell approach as described above. For logistic regression, the label used corresponded to either the intervention (‘0’) or control (‘1’). Cross validation was performed on held-out cells (25% of the cells that were not used to build the models). After training and validating the intervention classification models, we applied these models to the single-cell RNA-seq dataset of the SVZ neurogenic niche from 28 mice, tiling 26 ages from young (3.3 months) to old (29 months). Data were plotted as described in ‘Age prediction and validation strategy’, with (log(p(control) / p(intervention))) as a function of the actual chronological age of aNSC-NPC BootstrapCell transcriptomes. Old mice were more likely to be classified as ‘isochronic old control’, whereas young mice were more likely to be classified as ‘heterochronic old’, indicating that the gene signature that distinguishes exposure to young and old blood is relevant to aging. R is the Pearson correlation. Higher correlation indicates that the main intervention signature overlaps with and reverses age-related changes. Correlations between intervention state prediction and chronological age across cell types and interventions were assessed, with a separate classifier built for each. The exercise classifiers were built to distinguish old sedentary from old exercised transcriptomes for each cell type. The lower correlation between intervention state predictions and age for the exercise samples implies that the signatures that distinguishes exercised and sedentary mice are less related to aging than those derived from parabiosis intervention classifiers.

Statistics and reproducibility

No statistical methods were used to predetermine sample sizes; we determined our sample sizes based on our previous analysis of similar types of datasets48. For study design, we used four independent cohorts of mice, each spanning different ages, to build the age prediction models. This design allows us to test the machine learning aging clock models with a robust cross-cohort validation (that is, ‘leave-one-cohort-out’ validation). Two independent experiments of heterochronic parabiosis were performed, involving 6 mice (4 collected, cohort 1) and 18 mice (cohort 2), with data collection spread across 4 d. One experiment of exercise (with controls lacking a running wheel) was performed, involving 15 mice processed across 2 d. Animals from group 3 from parabiosis cohort 2 were excluded because sample multiplexing failed and it was not possible to distinguish samples. The experiments were not randomized. Investigators were not blinded to allocation during experiments and outcome assessment, although the genomics analyses were performed in a systematic manner. To test correlations, we used Pearson’s correlation. To determine the statistical significance of the differences between intervention and control, we used the Wilcoxon rank-sum test (a non-parametric test).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.