TYSHKOVSKII ALEKSANDR (US)
WO2020255095A1 | 2020-12-24 |
US20210388442A1 | 2021-12-16 | |||
US20220136037A1 | 2022-05-05 | |||
US20190106747A1 | 2019-04-11 |
WHAT IS CLAIMED IS: 1. A computer-implemented method for calculating expected total and remaining lifespan, chronological age, and lifespan-adjusted age of a biological test system, optionally a cell, tissue, organ, or organism, the method comprising: providing a biological test system, optionally a cell, tissue, organ, or organism; determining expression of two or more transcripts identified in Table A; accessing, from memory, a model for calculating lifespan, chronological age, or lifespan-adjusted age of a biological test system, wherein the model calculates lifespan, chronological age, or lifespan-adjusted age, i.e. chronological age of the organism normalized by its expected lifespan, based on expression of the two or more transcripts; and calculating, using the model, expected lifespan, chronological age, or lifespan- adjusted age for the system. 2. The method of claim 1, comprising determining expression of 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 125, 150, or more of the transcripts. 3. The method of claim 2, comprising determining expression of up to 100, 125, 150, 175, 200, 250, 300, 350, 400, or 500 of the transcripts. 4. The method of claims 1 to 3, further comprising applying an intervention to the system, determining expression of the one or more transcripts during and/or after an application of an intervention, and calculating chronological age, lifespan- adjusted age, or expected lifespan based on the expression of the one or more transcripts. 5. The method of claim 4, further comprising comparing the chronological age, lifespan-adjusted age, or expected lifespan to a reference chronological age, lifespan-adjusted age, or expected lifespan. 6. The method of claim 6, wherein the reference chronological age, lifespan-adjusted age, or expected lifespan is a baseline chronological age, lifespan-adjusted age, or expected lifespan obtained in the same test system before application of an intervention, or a chronological age, lifespan-adjusted age, or expected lifespan obtained earlier in time in the same test system, or a chronological age, lifespan- adjusted age, or expected lifespan in a reference system that represents the chronological age, lifespan-adjusted age, or expected lifespan in the absence of an intervention. 7. The method of claims 1-6, wherein chronological age, lifespan-adjusted age, or expected lifespan is calculated using a software-based modeling algorithm, optionally wherein the algorithm comprises a linear algorithm; principal component analysis (PCA); classification or decision trees; elastic net analysis; linear and polynomial support vector machines (SMV); shrunken centroids; random forest algorithms; support vector machines; gradient boosting; or neural networks. 8. The method of claim 7, comprising calculating a chronological age, lifespan- adjusted age, or expected lifespan using the determined expression and applying an algorithm to the levels. 9. The method of claim 8, wherein the algorithm comprises : tAge = intercept + b1 * T1 + b2 * T2 +...+ bn * Tn; Where b1 - bn are the pre-trained model coefficients for genes listed in Table A. and T1 - Tn are the normalized expression levels of given Transcripts. 10. The method of claim 9, wherein the model is a log-log model, and the method further comprises performing a transformation wherein: tAgem = 10 ** (-10 ** (-tAge)), wherein tAge is the estimate from the formula above, and tAgem is a transformed tAge that reflects the chronological or lifespan-adjusted biological age on a scale of 0-1, where 0 is a moment of birth, and 1 is a maximum achievable lifespan for the model. 11. The method of claims 1-10, further comprising identifying an intervention as having a positive effect when expected lifespan increases, and/or chronological age, or lifespan-adjusted age decreases, and/or identifying an intervention as having a damaging effect when expected lifespan decreases, and/or chronological age, or lifespan-adjusted age increases. 12. The method of claim 11, further comprising: selecting an intervention that has been identified as having a positive effect as a candidate intervention; applying the candidate intervention to an in vivo model of, optionally wherein the model is a non-human test animal or a human subject with a disorder or condition associated with aging; and determining whether the candidate intervention has a positive effect on the model, optionally on the disorder or condition related to aging. 13. A method of predicting an effect of an intervention on lifespan, the method comprising: providing a biological test system, optionally a cell, tissue, organ, or organism; determining expression of two or more transcripts identified in Table A; applying an intervention to the system, calculating chronological age, lifespan-adjusted age, or expected lifespan using the method of claims 1-12; comparing the predicted chronological age, lifespan-adjusted age, or expected lifespan to a reference chronological age, lifespan-adjusted age, or expected lifespan; and identifying an intervention as likely to have a protective effect on lifespan when chronological age or lifespan-adjusted age decreases and/or expected lifespan increases, and/or identifying an intervention as likely to have a damaging effect on lifespan when chronological age or lifespan-adjusted age increases and/or expected lifespan decreases. 14. The method of claim 13, comprising determining expression of 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 125, 150, or more of the transcripts. 15. The method of claim 14, comprising determining expression of up to 100, 125, 150, 175, 200, 250, 300, 350, 400, 500, or 1,000 transcripts, including determining expression of at least 5, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 125, 150, or all of the transcripts. 16. The method of claim 13, wherein the reference chronological age, lifespan- adjusted age or expected lifespan is a baseline chronological age, lifespan- adjusted age or expected lifespan obtained in the same test system before application of an intervention, or a chronological age, lifespan-adjusted age or expected lifespan obtained earlier in time in the same test system, or a chronological age, lifespan-adjusted age or expected lifespan in a reference system that represents the chronological age, lifespan-adjusted age or expected lifespan in the absence of an intervention. 17. The method of claims 13-16, comprising calculating chronological age, lifespan- adjusted age or expected lifespan using an algorithm, optionally wherein the algorithm comprises using a manual or software-based modeling algorithm, optionally wherein the algorithm comprises a linear algorithm; principal component analysis (PCA); classification or decision trees; elastic net analysis; linear and polynomial support vector machines (SMV); shrunken centroids; random forest algorithms; support vector machines; gradient boosting; or neural networks. 18. The method of claim 17, comprising calculating a chronological age, lifespan- adjusted age or expected lifespan using the determined expression and applying an algorithm to the levels. 19. The method of claim 18, wherein the algorithm comprises : tAge = intercept + b1 * T1 + b2 * T2 +...+ bn * Tn Where b1 - bn are the pre-trained model coefficients for every gene from Table A. and T1 - Tn are the normalized expression levels of given Transcripts. 20. The method of claim 20, wherein the model is a log-log model, and the method further comprises performing a transformation wherein: tAgem = 10 ** (-10 ** (-tAge)), wherein tAge is the estimate from the formula above, and tAgem is a transformed tAge that reflects the chronological or lifespan-adjusted biological age on a scale of 0-1, where 0 is a moment of birth, and 1 is a maximum achievable lifespan for the model. 21. The method of claim 20, further comprising: selecting an intervention that has been identified as having a protective effect as a candidate intervention; applying the candidate intervention to an in vivo model of a disorder or condition associated with aging, optionally wherein the model is a non-human test animal or a human subject in a clinical trial, optionally a model or a trial of a disorder or condition associated with aging; and determining whether the candidate intervention has a protective effect on the model. . |
EXAMPLES The invention is further described in the following examples, which do not limit the scope of the invention described in the claims. Materials and Methods The following materials and methods were used in the Examples below. Preprocessing of gene expression training data For ITP data, liver samples from compound-treated male and females mice along with age- and sex-matched corresponding controls were collected (Table S1). RNA was extracted from tissues with PureLink RNA Mini Kit as described in the protocol and passed to paired-end 150 bp sequencing on the Illumina NovaSeq 6000. Reads were aligned to Mus musculus genome (GRCm38) using STAR (version 2.5.2b) and counted with FeatureCounts (version 1.5). RNA-seq raw data was filtered using soft threshold: genes with less than 5 reads in more than 80% of samples were filtered out. RNA-seq data was then subjected to RLE normalization. RNA-seq data from public sources was preprocessed accordingly. Both microarray and RNA-seq data were log-transformed and subjected to scaling. Gene IDs from all used platforms were mapped to Entrez IDs, while gene IDs mapping to multiple Entrez IDs were removed. Human genes were mapped to the corresponding one-to-one mouse orthologs, according to bioMart annotation. Only genes detected in at least 10% of samples were maintained, resulting in the final subset of 19,310 genes. Besides classical scaling-based normalization, we also tested normalizations based on ranking and YuGene transformation 33 . For differential aging clock, random age group of healthy control animals was selected for each dataset. Median expression of every gene was calculated for this subset and subtracted from gene expression profiles for every sample. The resulting gene expression profiles were used as a training set for biological age prediction. Lifespan-adjusted age for every sample was calculated as a chronological age divided by maximum lifespan, defined as 90 th quantile lifespan of animals in the corresponding cohort. Survival curves for each animal model were extracted from public sources and used to calculate the maximum lifespan. Aging clock development To develop a model of biological age prediction, we utilized supervised machine learning. We randomly selected 90% of samples as a training set and used the remaining 10% of samples as a test set. To identify the most accurate clock, we trained 4 types of models, including Elastic Net linear model, random forest, support vector machines (SVM) and gradient boosting (XGBoost). Models were trained to predict chronological or lifespan-adjusted age in linear or log-log-transformed scale (as described in 9 ). Missed gene expression values were imputed based on median values for the corresponding genes calculated from training set. Hyperparameters were optimized with GroupKFold cross-validation on 5 folds, where source of each sample (GEO ID) was considered as a group. Mean Absolute Error (MAE) and R 2 were utilized as accuracy metrics during cross-validation and evaluation of final model. To obtain generalized estimate of model accuracy, we performed nested cross- validation, randomly dividing data into training and test sets 25 times. During each run, the model was trained on a training set, and its accuracy was estimated on a test set. This resulted in distribution of MAE and R 2 for each type of model and outcome variable transformation. Application of the clock model Preprocessing of the data prior to applying the clock included filtering of low- expressed genes using soft threshold (genes with less than 5 reads in more than 80% of samples were removed), mapping of gene IDs to the mouse Entrez IDs, RLE normalization (for RNA-seq data), log-transformation, scaling and ranking/YuGene transformation (optionally, for the corresponding clock models). Missing values corresponding to the clock genes not detected in the data were imputed with the average values precalculated during the training of the clock model. For single-cell RNA-seq datasets, prior to the described preprocessing pipeline, meta-cells were prepared by randomly pooling gene expression profiles across certain number of cells corresponding to the same cell type or tissues. Afterwards, the preprocessing algorithm was the same as for the bulk RNA-seq datasets. Preprocessed gene expression profiles were used as an input for the clock, and tAges were returned by the model as an output. Differences between tAges of control and treated groups were assessed using linear regression model (for OSKM reprogramming datasets) and independent t-tests (for other applications). When multiple comparisons were assessed, obtained p-values were adjusted with Benjamini- Hochberg method. Example 1. Rodent transcriptomic clocks of biological age. To construct transcriptomic clocks of chronological age in rodents, we aggregated gene expression data corresponding to 2,568 samples obtained from healthy mice and rats across 26 tissues from 84 independent datasets (Fig.1A). To separate detrimental biomarkers of aging from potential neutral or beneficial age- related changes and to build a clock of biological aging, we also introduced 588 tissue samples corresponding to various genetic, pharmacological and environmental interventions that either extend or shorten mouse lifespan. We further added lifespan- extending interventions to this dataset by sequencing 182 liver samples of male and female mice subjected to 20 different chemical compounds tested by the ITP program 1–4 , together with the age- and sex-matched controls. By normalizing the age of rodents by maximum lifespan of corresponding models obtained from survival studies, we calculated lifespan-adjusted age for each given sample, ranging from 0 (birth) to 1 (end of life). This resulted in 3,338 samples, for which both gene expression and lifespan-adjusted age were measured at once. The resulting data covered a substantial range of rodent lifespan, from 0.5 to 32 months, as well as various genetic backgrounds and models of lifespan-extending and -shortening interventions (Fig.1B-C). Using the collected data and response values of real and lifespan-adjusted age, we constructed clocks of chronological and biological age (Table A). For tissues highly represented in our dataset, such as liver, brain and skeletal muscle, we built tissue-specific clocks. We also constructed multi-tissue transcriptomic clocks using all the data across 26 tissues. To make our clocks robust across independent datasets, we utilized group K-fold approach, using independent datasets for training and cross- validation sets. We randomly assigned 15% of the sampling size to the test set and applied an elastic net algorithm, previously shown to be effective for the construction of epigenetic clocks 6–9 . Chronological liver, brain, muscle, and multi-tissue clocks were able to predict absolute real age of healthy mice with the Pearson’s correlation coefficients of 0.93, 0.85, 0.9 and 0.86, respectively (Fig.1D). Lifespan-adjusted clocks resulted in results of similar quality. To check if the obtained quality was robust across different subsets of training and test sets, we reproduced the algorithm using 25 independent random subsamplings. The median quality of the clocks estimated using Pearson’s correlation coefficient was found to be 0.81, 0.89, 0.95 and 0.84 for brain, liver, muscle, and multi-tissue clock, respectively (Fig.1G). To examine if this model can be expanded to human data, we also included 1,289 brain, skeletal muscle and skin samples from people of different ages as well as 10 brain samples from people with Hutchinson-Gilford progeria syndrome. The model trained on mouse, rat and human samples exhibited high quality on test set (Pearson’s correlation coefficient > 0.75), suggesting that a single transcriptomic clock can be applied to capture biological aging of tissues across multiple mammalian species, including human. Since gene expression data is often accompanied by substantial batch effects, we tested whether clock quality was higher when age deviation instead of absolute age was predicted. To do that, for every dataset we randomly selected one control group of a certain age and calculated ages and gene expression values of all samples within this dataset compared to the age and median gene expression profile of this random group. By applying the same machine learning algorithm, we obtained clocks with significantly higher quality (Fig.1E), resulting in median Pearson’s correlation coefficients on test sets equal to 0.88, 0.92, 0.97 and 0.93, respectively (Fig.1G). Linear model coefficients of top 500 genes with the highest absolute weights for trained models are stored in Table A. To check how robust the clocks are when applied to independent datasets, we estimated their quality on test sets assembled from datasets that were not included in the training set (Fig. 1F). The resulting median Pearson’s correlation coefficients of the clocks on independent datasets were equal to 0.88, 0.84, 0.89 and 0.8 for brain, liver, muscle, and multi-tissue clocks, respectively (Fig.1G). Therefore, despite a significant batch effect and noise associated with gene expression data, our models were able to predict both chronological and biological age of samples obtained from independent datasets with high accuracy, comparable to the clocks based on other types of omics data 6,10 . Interestingly, despite high accuracy of real age prediction, the chronological transcriptomic aging clock not adjusted for lifespan was unable to capture the difference between livers of control mice and mice subjected to Hutchinson-Gilford progeria syndrome (HGPS) (Fig.1H). Although mice with the progeria-associated mutations in Lmna gene have 10-fold reduced lifespan compared to wildtype mice, they are not predicted to be older by many chronological clocks, including epigenetic clocks 6 . However, our lifespan-adjusted clocks of biological age were able to capture the difference and predicted a significantly higher transcriptomic age (tAge) for HGPS mice (Fig.1H), suggesting that they are indeed more applicable for the assessment of the effect of interventions on healthspan than regular clocks trained on chronological aging. Example 2. Consistency between the predictions of transcriptomic and epigenetic clocks. To check if the predictions of our transcriptomic clocks were consistent with those estimated by the established epigenetic clocks, we utilized the data on mice subjected to isochronic and heterochronic parabiosis 11 . In this dataset, we carried out RNA-seq and used DNA methylation data from the same mice representing 4 groups: young (3-month-old) mice attached to young mice, old (21-month-old) mice attached to young mice, young mice attached to old mice, and old mice attached to young mice. After 3 months of attachment, mice were sacrificed, and their livers were collected. Besides, livers were collected from mice attached for 3 months, followed by detachment for 2 months. Previously, we observed a significant rejuvenation of old mice attached to the young ones compared to the isochronic group, that was preserved even after 2 months of detachment 11 . On the other hand, young mice attached to the old mice were older, but mostly recovered after detachment. Interestingly, the same pattern was observed when our transcriptomic liver clocks were applied (Fig.2A). Moreover, the Pearson’s correlation coefficient between age estimates obtained with liver epigenetic clocks and liver transcriptomic clocks was 0.89 (Fig.2C). Remarkably, the correlation between tAge and epigenetic age (eAge) remained the same even after chronological age was subtracted from the predictions (Fig.2B,D). Residuals of tAge and eAge (Δ Ages) were strongly correlated for chronological epigenetic clocks as well as liver transcriptomic clocks, both chronological and lifespan-adjusted (Fig.2B). Therefore, epigenetic and transcriptomic clocks provide consistent results even after adjustment for chronological age, pointing to shared molecular mechanisms of biological aging at the level of DNA methylation and gene expression. Example 3. Transcriptomic clocks capture age-related changes induced by diseases, sleep deprivation, cell reprogramming and embryogenesis. To determine if our clocks capture the effect of other established models that affect biological age, we applied them to the gene expression data from rodents subjected to various aging-related conditions and diseases as well as to control rodents of the same chronological age. We observed a consistent increase in biological age in mice subjected to ischemic stroke, both in brain and heart, as assessed using lifespan- adjusted brain and multi-tissue clocks, respectively (Fig.3A). Interestingly, we observed higher difference between control and diseased mice for samples collected from ipsilateral brain hemisphere than for samples from the contralateral hemisphere, although statistically significant difference was observed in both cases (Fig.3B). We also detected significantly higher biological tAge in brains of 5xFAD mice subjected to the rodent model of Alzheimer’s disease (AD) (Fig.3C). The same pattern of increased biological tAge was observed in kidneys of mice with chronic kidney disease (CKD) and type II diabetes (Fig.3D). Interestingly, contrary to the other age-related diseases, hepatocarcinoma resulted in a slight non-significant decrease of tAge in murine liver (Fig.3E). To investigate functional processes behind the observed difference in biological transcriptomic age, we utilized functional enrichment analyses (Fig.7A) looking for pathways enriched for genes, which expression is responsible for the observed increase of tAge (and decreased in case of carcinoma). We found significant upregulation of genes associated with apoptosis, p53 signaling, coagulation and NF- kB signaling related to inflammation, shared by the majority of diseases including ischemic stroke, AD, CKD and diabetes. On the other hand, carcinoma samples were characterized by the downregulation of aging-associated coagulation and NF-kB signaling genes. Therefore, the constructed clocks allow to investigate particular cellular mechanisms that lie behind observed age-related changes. To examine biological age of animals subjected to short-term detrimental interventions, we employed sleep deprivation (SD) model (Fig.2E). We subjected 2- month-old mice to 5 days of SD, 10 days of SD or 10 days of SD followed by 5 days of recovery, and compared their transcriptomic ages estimated in liver to those from age-matched control animals. Transcriptomic clocks revealed increase of biological age in response to 5 or 10 days of sleep deprivation, and it was restored to the original level after 5-day recovery (Fig.2F). Interestingly, epigenetic clocks applied to the same tissues also showed similar trend towards age acceleration driven by sleep deprivation (Fig.7B), but it was not significant. In contrast, transcriptomic clocks demonstrated a statistically significant difference between the groups (adjusted p- value < 0.05), suggesting that gene expression aging-related changes may be more sensitive to short-term interventions compared to changes in DNA methylation. Functional enrichment analysis revealed that the increase of biological age caused by SD was mainly associated with the upregulation of genes involved in inflammatory response, interleukin and TGF-β signaling, while these changes were reversed after 5 days of recovery (Fig.2G). Previously described epigenetic clocks demonstrated U-shaped eAge behavior during early murine embryogenesis, reaching the minimum eAge approximately between 6.5 and 10.5 days after fertilization 12 . This finding was consistent with the ground zero model stating that different molecular systems of the embryo are rejuvenated during early embryogenesis, followed by monotonous aging of the animal 13 . To test if this model is confirmed at the level of gene expression, we applied the multi-tissue lifespan-adjusted transcriptomic clock to the gene expression data ranging from the zygote to the 17 th day of embryogenesis 14 . Consistent with the epigenetic data, we observed U-shaped dynamics of biological tAge during early development with the minimum value approximately between 7 and 10 days after fertilization (Fig.3F) that agrees with the ground zero estimate measured based on DNA methylation data. Functional enrichment revealed that some of the pathways are both responsible for the decrease of tAge before ground zero and the increase of tAge after it, including TNFa signaling via NF-kB, p53 signaling, apoptosis, hypoxia and cholesterol homeostasis (Fig. 7C). To test if our multi-tissue transcriptomic clock trained on mouse and rat tissues can be applied to assess biological age of human samples, we evaluated transcriptomic age of prefrontal cortex and primary fibroblasts derived from people of various chronological ages (Fig.3G). Transcriptomic clocks correctly predicted age differences across these individuals and were especially accurate for prefrontal cortex (Pearson’s ρ = 0.95 and 0.69 for prefrontal cortex and primary fibroblasts, respectively; p-value < 2 . 10 -4 ). Interestingly, biological age of OSKM-induced pluripotent stem cells (iPSCs) derived from fibroblasts of these individuals was significantly lower than tAge of original cells and wasn’t correlated with the chronological age of donors (Pearson’s ρ = -0.029, p-value = 0.89). This is consistent with the previous findings, showing that epigenetic age of iPSCs is close to 0 regardless of the original chronological age of cells subjected to OSKM reprogramming. Using multi-tissue transcriptomic clock, we also showed that dynamics of rejuvenation associated with classical OSKM reprogramming is different between mouse and human cells 15 . In murine cells, stable minimum tAge was reached approximately after 10 days of OSKM induction. In addition, transcriptomic rejuvenation of human cells followed sigmoidal curve and took more time, reaching minimum approximately on the 28 th day of reprogramming. Interestingly, such rejuvenation dynamics is consistent with differences in the duration of reprogramming in mouse and human cells, lasting, on average, for 14 and 30 days in mouse and human cell lines, respectively 16 . Besides, our clock also captured rejuvenation in mouse and human cells subjected to chemical reprogramming with 7c cocktail 17,18 , demonstrating further consistency across predictions of epigenetic and transcriptomic clocks. Example 4. Transcriptomic clocks predict chronological aging of single cells. To examine if multi-tissue transcriptomic clocks built on bulk data can predict chronological age of individual cells, we applied them to murine scRNA-seq data collected by Tabula Muris Consortium 19 . This dataset covers single cell gene expression data from various tissues of mice of different age and sex. To increase coverage of single cell data, we utilized a meta-cell approach and pooled read counts from randomly selected 500 cells for every biological sample. Following this procedure, multi-tissue chronological clock was able to capture chronological age dynamics for most mouse organs (Fig. 4A) with median Pearson’s ρ = 0.87. By repeating the same procedure taking different number of cells used for meta-cell data aggregation, we found that 50 cells were sufficient to achieve median accuracy of Pearson’s ρ = 0.75 across organs (Fig.4B). This number of cells corresponds to the average coverage of approximately 430,000 reads per meta-cell (Fig.4C). To examine if biological age varies across different cell types, we aggregated meta-cells from randomly selected 50 cells for individual cell types from 1-month-old mice. We observed significant variation of tAge across cell types, both for chronological and lifespan-adjusted clocks (Fig.4D). Several types of white blood cells, including monocytes, granulocytes and macrophages, had higher biological age, in agreement with their role in inflammaging and development of aging-related diseases 20,21 . In contrast, mesenchymal stem cells exhibited the lowest transcriptomic age, presumably associated with the decreased proportion of stem cells with age and detrimental effect of stem cell exhaustion on organism’s healthspan 22,24 . We also assessed age-related change of transcriptomic age for each cell type presented in Tabula Muris dataset, utilizing meta-cell procedure, multi-tissue transcriptomic clock and linear regression model. Remarkably, for most cell types we observed significant positive tAge slope (Fig.4E), suggesting that aging is accompanied by systemic increase of biological age in the majority of cells. Interestingly, even mesenchymal stem cells exhibited significant increase of tAge with age. Therefore, these transcriptomic clocks were able to capture both age-related changes in cell composition of the organ (e.g., accumulation of macrophages and granulocytes) as well as aging-associated changes in gene expression within individual cells. Example 5. Transcriptomic clocks reveal aging-related dynamics of single cells during reprogramming. To check if the gene expression clock of biological age can capture trajectory of age-related dynamics during OSKM-induced cellular reprogramming, we utilized scRNA-seq data, covering 18 days of reprogramming of murine embryonic fibroblasts (MEFs) 23 . We applied lifespan-adjusted multi-tissue transcriptomic clock to each individual cell presented in the data and observed significant variability of tAge across different cell trajectories (Fig.5A). Using gene expression signatures of pluripotency, cell cycle, apoptosis and senescence-associated secretary phenotype (SASP), we found that some of them characterized cell types with higher average tAge (Fig.5B). Thus, stromal cells characterized by upregulation of SASP- and apoptosis-associated genes had also higher tAge, whereas mesenchymal-epithelial transition (MET) and induced pluripotent cells (IPS) showed increased expression of cell cycle genes and lower tAge. To focus on the dynamics of transcriptomic age throughout trajectory of different cell types generated during OSKM reprogramming, we applied an optimal transport method 23 . We observed a significant drop of tAge during first days of reprogramming, consistent with the mouse bulk reprogramming data (Fig.5C). Interestingly, after that biological age was significantly increased in cells moving towards stromal and epithelial trajectory, while some other cell types, including IPS, neural and MET cells, maintained lower average tAge. Remarkably, ancestors of stromal cells demonstrated a similar significant increase of SASP-associated gene expression during first days of reprogramming, while apoptosis-related genes were upregulated in stromal, epithelial and trophoblast cell types (Fig.5D), pointing to the impact of apoptosis and SASP pathways on the variation of tAge across different cell types. To test it, we estimated correlation between tAge and different gene expression signatures across all individual cells (Fig.5E). Indeed, we observed strong positive correlation between tAge and signatures of apoptosis and SASP, and negative correlation between tAge and signatures of cell cycle and pluripotency. This finding suggests that cell reprogramming doesn’t lead to systemic rejuvenation of all cells. Instead, it results in a significant variation of biological age across cell types, leading to a decrease of tAge in some of the cells (IPS, MET), but significant upregulation of aging-related biomarkers in the others (stromal, epithelial cells). Single-cell data and transcriptomic clocks allow deeper investigation of aging-related changes induced by different interventions and assessment of heterogeneity of biological age dynamics induced by these models. Example 6. Gene expression clocks of aging and lifespan predict new longevity interventions. To discover candidate interventions that slow down biological aging and extend mammalian lifespan and healthspan, we utilized ClockBase 25 and processed 1,349 publicly available RNAseq datasets corresponding to mouse tissues exposed to various pharmacological, dietary and genetic interventions. For every intervention, we used gene expression data to calculate tAges of subjected mice and compared them with tAges of age-matched control mice. This resulted in identification of a number of interventions with a statistically significant effect on tAge (Fig.6A). As expected, we observed an increase of biological age in transgenic models of Alzheimer’s disease and in mice subjected to LPS injection, interleukin 6 injection and choline-deficient, L-amino acid-defined, high-fat diet (CDAHFD) associated with development of nonalcoholic steatohepatitis. Interestingly, we also identified interventions leading to a decrease in biological age. In addition to the well-known lifespan-extending interventions, such as rapamycin, they included β3-adrenoceptor agonist CL 316,243 26,27 and the cardiac glycoside ouabain 28 . According to the transcriptomic clock, 3-month ouabain treatment resulted in an ~20% reduction of biological age in 24-month-old mice (Fig.6B). This compound was also shown to exhibit senolytic activity and improve physical function of aged rodents 28 . Our data suggests that ouabain may counteract age-related changes and produce lifespan-extending effect in mammals. To examine if longevity of rodents can be predicted based on gene expression data, we used 1,030 liver samples of mice and rats with known maximum lifespan to construct transcriptomic clock of expected lifespan. Using an elastic net model and group K-fold cross-validation, we built a clock that predicted expected maximum lifespan of the given animal with a median Pearson’s correlation coefficient of 0.76 and median MAE of 2.5 months, as assessed using 25 randomly subsampled test sets (Fig.6C). Remarkably, the introduction of chronological age to the feature space didn’t improve the model, suggesting that the transcriptome profile covers most of the aging-associated effect on longevity. Linear model coefficients of top 500 genes with the highest absolute weights for absolute lifespan and differential lifespan clocks are stored in Table A. We applied the constructed clock to the liver samples obtained from long- lived growth hormone (GH) deficient mutant mice of different ages together with the corresponding controls 29 . Predicted expected lifespans of both Ames dwarf mice (Prop1 df/df ) and Little mice (Ghrhr lit/lit ) were indeed significantly higher than those of control mice (Fig.6D). The gene expression clock captured the difference in lifespan between the models across presented ages, suggesting that certain longevity- associated molecular mechanisms may be preserved throughout the lifespan. To test if the constructed clock can be used to identify new longevity interventions, we applied it to the liver and kidney samples obtained from young UM- HET3 male mice subjected to 25 different compounds for 1 month together with age- matched controls. These compounds were chosen using a CMAP platform 30 and gene expression signatures of longevity identified in our previous studies 31 .23 of the 25 compounds induced differential expression of at least 1 gene (adjusted p-value < 0.05) and were selected for subsequent analysis. Transcriptomic clock revealed a significant positive effect for 10 of these compounds on expected lifespan of mice (adjusted p- value < 0.05 for data pooled across tissues) (Fig.6E). To validate the predictions obtained with transcriptomic longevity clock, we selected 6 of these compounds, including KU-0063794 (an mTOR inhibitor), selumetinib (a MEK1/2 inhibitor), valdecoxib (a nonsteroidal anti-inflammatory drug (NSAID)), AZD-8055 (an mTOR inhibitor), GDC-0941 (pictilisib, a PI3Kα/δ inhibitor), and ascorbyl-palmitate, and subjected 20-2425-month-old C57BL/6JNia mice to each of these compounds added to the animal diet. Remarkably, 3 out of 6 compounds predicted to be beneficial for mouse longevity according to the clock indeed significantly extended lifespan of old male mice (log-rank p-value < 0.05) (Figs.6F-H). These drugs included MEK1/2 inhibitor selumetinib and mTOR inhibitors KU-0063794 and AZD-8055, suggesting that the obtained model may indeed be applied to identify novel longevity interventions. We further tested the effect of KU-0063794 on mouse healthspan and found that this compound improved gait speed and decreased frailty index of 30-month-old mice 32 , suggesting that the methods described herein can be applied to identify novel compounds that extend both lifespan and healthspan of old animals. REFERENCES 1. Strong, R., Miller, R.A., Antebi, A., Astle, C.M., Bogue, M., Denzel, M.S., Fernandez, E., Flurkey, K., Hamilton, K.L., Lamming, D.W., et al. (2016). Longer lifespan in male mice treated with a weakly estrogenic agonist, an antioxidant, an α-glucosidase inhibitor or a Nrf2-inducer. Aging Cell 15, 872–884. 10.1111/acel.12496. 2. Miller, R.A., Harrison, D.E., Allison, D.B., Bogue, M., Debarba, L., Diaz, V., Fernandez, E., Galecki, A., Garvey, W.T., Jayarathne, H., et al. (2020). Canagliflozin extends life span in genetically heterogeneous male but not female mice. JCI Insight 5, e140019. 3. Strong, R., Miller, R.A., Bogue, M., Fernandez, E., Javors, M.A., Libert, S., Marinez, P.A., Murphy, M.P., Musi, N., Nelson, J.F., et al. (2020). Rapamycin-mediated mouse lifespan extension: Late-life dosage regimes with sex- specific effects. Aging Cell 19, 1–10.10.1111/acel.13269. 4. Harrison, D.E., Strong, R., Reifsnyder, P., Kumar, N., Fernandez, E., Flurkey, K., Javors, M.A., Lopez-Cruzan, M., Macchiarini, F., Nelson, J.F., et al. (2021).17-a-estradiol late in life extends lifespan in aging UM-HET3 male mice; nicotinamide riboside and three other drugs do not affect lifespan in either sex. Aging Cell 20, 1–10.10.1111/acel.13328. 5. Siler, W. (1983). Parameters of mortality in human populations with widely varying life spans. Stat. Med.2, 373–380. 6. Horvath, S. (2013). DNA methylation age of human tissues and cell types. Genome biol 14, R115.10.1186/gb-2013-14-10-r115. 7. Petkovich, D.A., Podolskiy, D.I., Lobanov, A. V, Lee, S., Miller, R.A., and Gladyshev, V.N. (2017). Using DNA Methylation Profiling to Evaluate Biological Age and Longevity Interventions. Cell Metab.25, 954-960.e6. 10.1016/j.cmet.2017.03.016. 8. Meer, M. V, Podolskiy, D.I., and Tyshkovskiy, A. (2018). A whole lifespan mouse multi-tissue DNA methylation clock. Elife 7, 1–16. 10.7554/eLife.40675. 9. Lu, A.T., Quach, A., Wilson, J.G., Reiner, A.P., Aviv, A., Raj, K., Hou, L., Baccarelli, A.A., Li, Y., Stewart, J.D., et al. (2019). DNA methylation GrimAge strongly predicts lifespan and healthspan. Aging (Albany. NY).11, 303– 327. 10. Lu, A.T., Fei, Z., Haghani, A., Robeck, T.R., Zoller, J.A., Li, C.Z., Lowe, R., Yan, Q., Zhang, J., Vu, H., et al. (2023). Universal DNA methylation age across mammalian tissues. Nat. Aging.10.1038/s43587-023-00462-6. 11. Zhang, B., Lee, D.E., Trapp, A., Tyshkovskiy, A., Lu, A.T., Bareja, A., Kerepesi, C., Mckay, L.K., Shindyapina, A. V, Dmitriev, S.E., et al. (2023). Multi- omic rejuvenation and life span extension on exposure to youthful circulation. Nat. Aging.10.1038/s43587-023-00451-9. 12. Kerepesi, C., Zhang, B., Lee, S.-G., Trapp, A., and Gladyshev, V.N. (2021). Epigenetic clocks reveal a rejuvenation event during embryogenesis followed by aging. Sci. Adv.7, eabg6082. 13. Gladyshev, V.N. (2021). The Ground Zero of Organismal Life and Aging. Trends Mol. Med.27, 11–19.10.1016/j.molmed.2020.08.012. 14. Xue, L., Cai, J.Y., Ma, J., Huang, Z., Guo, M.X., Fu, L.Z., Shi, Y.B., and Li, W.X. (2013). Global expression profiling reveals genetic programs underlying the developmental divergence between mouse and human embryogenesis. BMC Genomics 14.10.1186/1471-2164-14-568. 15. Kriukov, D., Khrameeva, E.E., Gladyshev, V.N., Dmitriev, S.E., and Tyshkovskiy, A. (2022). Longevity and rejuvenation effects of cell reprogramming are decoupled from loss of somatic identity. bioRxiv.10.1101/2022.12.12.520058. 16. Fu, K., Chronis, C., Soufi, A., Bonora, G., Edwards, M., Smale, S.T., Zaret, K.S., Plath, K., and Pellegrini, M. (2018). Comparison of reprogramming factor targets reveals both species-specific and conserved mechanisms in early iPSC reprogramming. BMC Genomics 19, 1–13.10.1186/s12864-018-5326-1. 17. Yang, J., Petty, C.A., Dixon-McDougall, T., Lopez, M.V., Tyshkovskiy, A., Maybury-lewis, S., Tian, X., Ibrahim, N., Chen, Z., Griffin, P.T., et al. (2023). Chemically induced reprogramming to reverse cellular aging. Aging (Albany. NY).15, 5966–5989. 18. Mitchell, W., Goeminne, L.J.E., Tyshkovskiy, A., Zhang, S., Paulo, J.A., Pierce, K.A., Choy, A.H., Clish, C.B., Gygi, S.P., and Gladyshev, V.N. (2023). Multi-omics characterization of partial chemical reprogramming reveals evidence of cell rejuvenation. bioRxiv. 19. Almanzar, N., Antony, J., Baghel, A.S., Bakerman, I., Bansal, I., Barres, B.A., Beachy, P.A., Berdnik, D., Bilen, B., Brownfield, D., et al. (2020). A single-cell transcriptomic atlas characterizes ageing tissues in the mouse. Nature 583, 590–595.10.1038/s41586-020-2496-1. 20. Kristof, V.A., Strecker, J.-K., Tulotta, C., Minnerup, J., Schulz, C., and Soehnlein, O. (2023). Neutrophils in aging and aging-related pathologies. Immunol. Rev.314, 357–375.10.1111/imr.13153. 21. Maeyer, R.P.H. De, and Chambers, E.S. (2020). The impact of ageing on monocytes and macrophages. Immunol. Lett.230, 1–10. 10.1016/j.imlet.2020.12.003. 22. Lopez-Otin, C., Blasco, M.A., Partridge, L., Serrano, M., and Kroemer, G. (2013). The hallmarks of aging. Cell 153, 1194–1217. 10.1016/j.cell.2013.05.039. 23. Schiebinger, G., Shu, J., Tabaka, M., Cleary, B., Subramanian, V., Solomon, A., Gould, J., Liu, S., Lin, S., Berube, P., et al. (2019). Optimal-Transport Analysis of Single-Cell Gene Expression Identifies Developmental Trajectories in Reprogramming. Cell 176, 928-943.e22.10.1016/j.cell.2019.01.006. 24. Lopez-Otin, C., Blasco, M.A., Partridge, L., Serrano, M., and Kroemer, G. (2023). Hallmarks of aging: An expanding universe. Cell 186, 243–278. 10.1016/j.cell.2022.11.001. 25. Ying, K., Tyshkovskiy, A., Trapp, A., Liu, H., Moqri, M., Kerepesi, C., and Gladyshev, V.N. (2023). ClockBase^: a comprehensive platform for biological age profiling in human and mouse. bioRxiv. 26. Himms-hagen, J., Taatjes, J., Lang, S.S., Claus, H., Taatjes, J., and Lang, S.S. (1994). Effect of CL-316,243, a thermogenic b3-agonist, on energy balance and brown and white adipose tissues in rats. Am. J. Physiol.266, R1371-82. 27. Wang, W., Ishibashi, J., Trefely, S., Shao, M., Cowan, A.J., Sakers, A., Lim, H., Connor, S.O., Doan, M.T., Cohen, P., et al. (2019). A PRDM16-driven metabolic signal from adipocytes regulates precursor cell fate. Cell Metab.30, 174– 189.10.1016/j.cmet.2019.05.005.A. 28. Guerrero, A., Herranz, N., Sun, B., Wagner, V., Gallage, S., Guiho, R., Wolter, K., Pombo, J., Irvine, E.E., Innes, A.J., et al. (2019). Cardiac glycosides are broad-spectrum senolytics. Nat. Metab.1, 1074–1088.10.1038/s42255-019-0122-z. 29. Amador-Noguez, D., Yagi, K., Venable, S., and Darlington, G. (2004). Gene expression profile of long-lived Ames dwarf mice and Little mice. Aging Cell 3, 423–441.10.1111/j.1474-9728.2004.00125.x. 30. Subramanian, A., Narayan, R., Corsello, S.M., Peck, D.D., Natoli, T.E., Lu, X.L., Gould, J., Doench, J.G., Bittker, J.A., Root, D.E., et al. (2017). A Next Generation Connectivity Map: L1000 platform and the first 1,000,000 profiles. Cell 171, 1437–1452.10.1016/j.cell.2017.10.049. 31. Tyshkovskiy, A., Bozaykut, P., Borodinova, A.A., Gerashchenko, M. V., Ables, G.P., Garratt, M., Khaitovich, P., Clish, C.B., Miller, R.A., and Gladyshev, V.N. (2019). Identification and Application of Gene Expression Signatures Associated with Lifespan Extension. Cell Metab.30, 573–593. 10.1016/j.cmet.2019.06.018. 3 2. Tyshkovskiy, A., Ma, S., Shindyapina, A. V, Tikhonov, S., Lee, S.-G., Bozaykut, P., Castro, J.P., Seluanov, A., Schork, N.J., Gorbunova, V., et al. (2023). Distinct longevity mechanisms across and within species and their association with aging. Cell 186, 1–21.10.1016/j.cell.2023.05.002. 33. Edgar, R. (2002). Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic Acids Res.30, 207–210. 10.1093/nar/30.1.207. OTHER EMBODIMENTS It is to be understood that while the invention has been described in conjunction with the detailed description thereof, the foregoing description is intended to illustrate and not limit the scope of the invention, which is defined by the scope of the appended claims. Other aspects, advantages, and modifications are within the scope of the following claims.