Rapid emergence of transcriptional heterogeneity upon molecular stress predisposes cells to two distinct states of senescence

Abstract

Slowing aging can reduce the risk of chronic diseases. In particular, eliminating senescent cells is a promising approach to slow aging. Previous studies found that both cells from older animals and senescent cells have noisy gene expression. Here, we performed a large-scale single-cell RNA-sequencing time course to understand how transcriptional heterogeneity develops among senescent cells. We found that cells experiencing senescence-inducing oxidative stress rapidly adopt one of two major transcriptional states. One senescent cell state is associated with stress response, and the other is associated with tissue remodeling. We did not observe increased stochastic gene expression. This data is consistent with the idea that reproducible, limited, distinct, and coherent transcriptional states exist in senescent cell populations. These physiologically distinct senescent cell subtypes may each affect the aging process in unique ways and constitute a source of heterogeneity in aging.

Supplementary Information

The online version contains supplementary material available at 10.1007/s11357-022-00709-x.

Introduction

Aging incurs a tremendous socioeconomic cost. Life expectancy has significantly increased since the twentieth century. People spend a smaller portion of their life working, while concurrently requiring more resources for health care with increasing age. Age itself is the biggest risk factor for many diseases, including heart disease, cancer, and neurodegenerative diseases like Alzheimer’s disease [1]. Thus, understanding the aging process and how to slow it, is a means of reducing the incidence of chronic diseases and decreasing the socioeconomic burden of aging.

Recent advances in aging research have shown that excessive accumulation of senescent cells with age is a major factor that precipitates age-related physiological decline and mortality [2–4]. Senescence is a physiological state cells can enter in response to severe molecular damage; it is one of an array of natural cellular stress responses [5]. Depending on the particular stress, cells may initiate several different responses such as (1) repair the damage and resume normal function, (2) undergo apoptotic or necrotic cell death, or (3) enter senescence, a state of stable proliferative arrest, associated with secretion of inflammatory cytokines.

In the present study, we set out to understand the physiological heterogeneity among senescent cells by sequencing the transcriptomes of individual cells. Senescent cell populations are known to be heterogeneous. This is evidenced by senolytic compounds that work only on certain types of senescent cells [6], and by transcriptionally distinct senescent cells produced by different types of molecular stress [7]. The existence of different types or “flavors” of senescent cells necessitates better characterization and maybe even better definition of senescence itself. Understanding how senescence manifests, and what types of senescent cells are there is paramount for understanding how to more effectively target them to improve the aging process.

The process of becoming a senescent cell may increase cell-to-cell variation in gene expression. Studies found senescent mouse and human cells had increased cell-to-cell variation in gene expression compared to their non-senescent counterparts [8, 9]. Additionally, many studies have found that older age increases cell-to-cell variation in gene expression among cells isolated from tissues [10–13]. Alternatively, at least one other recent study found that cell-to-cell variation in gene expression can also decrease with age [14]. In any case, abnormal changes to heterogeneity occur during aging [15], and it is not entirely clear how they originate. Since aging is accompanied by, and maybe even caused by, molecular damage, the cell-to-cell variation in gene expression among senescent and nonsenescent aged cells may have similar origins. It could be due to preexisting epigenetic differences, or due to the random nature of molecular damage events. Understanding how and why cell-to-cell variation in gene expression changes with age may provide novel insights into the mechanisms of biological aging.

Here, we used large-scale single-cell RNA-sequencing (scRNA-seq) to understand the mechanisms of cell-to-cell variation in gene expression upon induction of senescence. We set out to discriminate between the hypotheses transcriptional heterogeneity among senescent cells arises from preexisting heterogeneity, or that senescent cell heterogeneity arises purely from stochastic noise in gene expression. In the first case, subgroups of cells would exhibit coordinated gene expression changes. In the latter case, individual genes would exhibit increased uncoordinated variance; there would be no coherent patterns. To distinguish between these possibilities, we performed a transcriptome-wide analysis of gene expression at the single-cell level using the SPLIT-seq approach [16]. We performed time course experiments during senescence induction to generate whole-transcriptome data for thousands of individual cells across many individual experiments, finding reproducible patterns of gene expression variation. In total performed six independent experiments to rule out the contribution of technical noise and focus on biological cell-to-cell variation. Our results indicate that cells after stress fall into distinct reproducible transcriptional clusters with different levels of metabolic signaling. This kind of reproducible, limited heterogeneity could be interpreted to be part of a functional bet-hedging mechanism in response to the same type of DNA damage or as a labor division between senescent cells, which may be required for complex processes involving senescent cells, such as wound healing.

Methods

Cell culture and treatments

Human fetal lung fibroblasts IMR-90 were received from Coriell Cell Bank. Cells were cultured in Dulbecco’s modified Eagle’s medium containing 10% fetal bovine serum and penicillin/streptomycin. Cells were maintained at 37 °C and 5% CO₂. In all the experiments, we used replicatively young cells with cumulative population doublings less than 30.

For treatments, cells were seeded at ~ 10⁴/cm² and all treatments started the next day. For induction of quiescence, cells were kept in media with 0.2% serum for 7 days. For induction of senescence/senescence-like arrest with nutlin-3a, cells were kept for 7 days in media supplemented with 2.5uM nutlin-3a. For induction of senescence with oxidative stress, cells were treated with 55uM hydrogen peroxide for 2 h. In one experiment (Fig. S1h–n), cells were treated with 75uM hydrogen peroxide twice, on Day 0 and Day 3. As in other experiments, cells were collected on Day 7 after the first treatment. In this experiment, quiescence was induced by low serum treatment for 3 days, from Day 4 to Day 7. In the time course experiments, cells were collected at different time points after stress treatment, which was performed on Day 0. For assaying senescence-associated ß-galactosidase activity, cells were fixed with formaldehyde/glutaraldehyde and stained with 5-bromo4-chloro-3-indolyl P3-D-galactoside (X-Gal) as described before (Dimri et al., 1995). Presented X-gal staining images were lightly processed to increase brightness.

Untreated cells were collected on the same day as treatments were started, that is Day 0 in time course experiments. During the experiments, we changed media twice on other samples (treated with low serum, hydrogen peroxide, or nutlin-3a) to ensure that cells do not experience nutrient stress. All the applied treatments suppress cell division; hence, cells did not need to be split during the experiments.

Preparation of cDNA library and sequencing

We followed the standard SPLIT-seq protocol as described before [16]. Briefly, cells were fixed in formaldehyde, then permeabilized with triton, filtered through a 40um strainer, and counted. Cells were subjected to 3 rounds of barcoding using 3′ targeting polyT oligos in the first round. After barcoding, cells were digested with proteinase K and isolated cDNA was further amplified, size-selected, and indexed. The libraries were subjected to 2 × 150 bp paired-end sequencing on Illumina platforms by Genewiz.

Determining transcriptional profiles of individual cells

Reads were aligned to the human genome assembly hg19 using STARsolo following developers’ guidelines. Read2 was used to decode cellular barcodes and unique molecular identifiers, while read 1 was actually mapped to the transcriptome. Further processing was done with scanpy using developers’ guidelines [17]. We filtered out cells with low read count (less than 1000 or 2000 depending on sequencing depth) and cells with more than five percent of reads originating from the mitochondrial genome. Genes expressed in less than 100 cells were excluded from the analysis. We selected ~ 2000 most variable genes in each experiment and used them for principal component analysis (PCA). After PCA, cells were embedded using UMAP [18] and community detection was performed with Louvain algorithm [19]. Markers of transcriptional groups were determined by Wilcoxon rank-sum test.

Gene set enrichment analysis (GSEA)

Gene set enrichment analysis was performed using its Python implementation, GSEAPY. Genes were ranked by Wilcoxon rank-sum-based comparison of different transcriptional groups. The ranked list of genes was used as input for GSEAPY. We used “Hallmarks” gene sets and performed the analysis with 1000 permutations of gene sets. Benjamini–Hochberg Correction was used for adjusted p values. For plotting, only gene sets with adjusted p values of 0.05 or less and False Discovery Rate of 0.25 or less were included. When comparing two groups, positive enrichment scores indicate that gene sets are enriched in the first group compared to the second group, while negative enrichment scores indicate the enrichment of gene sets in a second group compared to the first one.

Scoring gene expression signatures

To derive gene expression signatures of senescent cell fractions, we selected genes that reproducibly show statistically significant (adjusted p values 0.05 or less) differences in expression between two fractions of senescent cells in independent experimental trials: in two non-time course experiments and three time course experiments. The final time point, Day 7, was used in the time course experiments. Experiment with repetitive stimulation with hydrogen peroxide was not used to derive these gene expression signatures. Enrichment of the signatures in all cells of the time course experiments was performed using score_genes function of the scanpy Python library. To score the enrichment of glycolytic genes, we used genes of the KEGG_GLYCOLYSIS_GLUCONEOGENESIS set.

Results

Two distinct transcriptional states emerge after oxidative stress

To generate senescent cells, we exposed human IMR-90 fibroblasts to hydrogen peroxide for 2 h [20]. To compare senescence to other non-proliferative states, cells were cultured in low serum media (0.2% FBS) to induce quiescence or with nutlin-3a, an MDM-2 inhibitor (Fig. 1a). Nutlin-3a was previously shown to induce either senescence or senescence-like arrest depending on conditions [21, 22]. After 1 week, cells were stained with X-gal to analyze senescence-associated (SA) ß-galactosidase activity. We confirmed that both oxidative stress and nutlin-3a treatment induced robust SA ß-galactosidase staining (Fig. 1b, Fig. S1o, p). Cells after oxidative stress also exhibited enlarged morphology, another classical marker of senescence (Fig. S1p). Importantly, before starting the experiments with the treatments, we conducted pilot experiments to confirm that untreated samples contained a minimal number of SA ß-galactosidase-positive cells (Fig. S1o). In all the experiments, we used replicatively young cells with a cumulative population doubling of less than 30.

Fig. 1.

Two transcriptionally distinct types of senescent cell populations emerge after oxidative stress. a Experimental outline. Cells were treated with 2.5uM nutlin-3a to induce senescence/senescence-like arrest, or with 55uM H₂O₂ for 2 h to induce senescence, or they were maintained for 7 days in 0.2% FBS media to induce quiescence. “Untreated” cells were collected in the beginning of the experiment. b Fraction of cells in each condition that were positive for senescence-associated ß-galactosidase as measured by X-gal staining. * indicates statistical significance with p < 0.05. Statistical testing was performed using non-parametric ANOVA (Kruskal–Wallis test) with Conover’s post hoc test and Holm-Sidak adjustment. c Uniform Manifold Approximation (UMAP)-based representation of transcriptional profiles of individual cells. Oxidative stress induces two distinct senescent states with different transcriptional signatures. d Louvain community detection analysis of individual cells. e. Post-oxidative stress cells that were transcriptionally similar to nutlin-treated cells were labeled as OxStress-TR, and cells that were more transcriptionally distinct were labeled as OxStress-CP. f Cell cycle phase of “Untreated” cells were determined by scoring expression of cell cycle specific genes. g, h Markers of OxStress-CP and OxStress-TR fractions were determined by Wilcoxon rank-sum test. Shown are top 10 statistically significant markers for each fraction. Also see Supplementary Spreadsheet File S2. j Expression of select markers of each fraction. Two OxStress fractions exhibit antagonistic pattern of expression of the markers

We collected cells from parallel plates (not used for X-gal staining) and processed them for single-cell RNA-sequencing (scRNA-seq) with SPLIT-seq 3′-tag protocol [16]. To determine the gene expression changes of the stress-induced senescence, we first compared the average expression profiles of untreated cells and cells after oxidative stress. We found strong overlap with the previously published signature of senescent fibroblasts [7]: out of 1311 genes in the universal signature of senescent fibroblasts, 1072 were present in our dataset, and when adjusted for multiple comparisons, 205 of them had a statistically significant change of expression in the same direction as in the universal signature. That is these genes were upregulated or downregulated upon the oxidative stress-induced senescence compared to the untreated sample in our dataset and were respectively upregulated or downregulated in senescent cells in the previous large-scale analysis [7] (Supplementary Spreadsheet File S1).

To proceed to single-cell analysis, we performed dimensionality reduction and embedding with Uniform Manifold Approximation (UMAP) [18]. The results are shown in Fig. 1c. Upon examining the UMAP space, we found that cells induced into senescence with oxidative stress fall into two distinct transcriptional clusters (Fig. 1c, OxStress cells). This transcriptional heterogeneity was specific to the post-oxidative stress cells, as both quiescent cells and nutlin-treated cells formed singular clusters (Fig. 1c, “LowSerum” and “Nutlin” cells respectively). Louvain community detection analysis also indicated that two fractions of OxStress cells fall into distinct Louvain clusters, further underscoring their transcriptional heterogeneity (Fig. 1d, e). Interestingly, one of the OxStress clusters localized closer to the “Nutlin” cells in the UMAP space, indicating stronger transcriptional similarity to the nutlin-treated cells. The only other sample in our experiment that formed two clusters was the control sample with untreated cells which segregated based on their cell cycle phase (Fig. 1c, f). We examined markers that differentiate the two OxStress clusters from each other (Fig. 1g, h, Supplementary Spreadsheet File S2). Fraction of OxStress cells that were transcriptionally closer to the nutlin-treated cells and quiescent cells was marked by higher expression of genes associated with cell motility/adhesion/organization: UACA, FN1, TRIO, SERPINE1, ROBO2. Hence, we labeled this fraction as “OxStress-TR” (tissue remodeling) for further discussion (Fig. 1e). The fraction of OxStress cells that were transcriptionally more distinct from the quiescent and the nutlin-treated cells was marked by higher expression of the genes associated with cellular homeostasis (Fig. 1g): FTL and FTH1 are subunits of ferritin, which stores intracellular iron, SQSTM1 is a regulator of autophagy; thioredoxin reductase TXRND1 and cystine/glutamate transporter SLC7A11 both help maintain cellular redox balance. Hence, this fraction of OxStress cells was labeled as “OxStress-CP” (cytoprotective) (Fig. 1e). As shown in Fig. 1j, identified markers exhibit antagonistic expression pattern in “cytoprotective” and “tissue remodeling” fractions of senescent cells. Thus, upon oxidative stress, senescent fibroblasts exhibited two distinct transcriptional signatures: cytoprotective response and response associated with tissue organization/remodeling.

It was important for us to distinguish technical noise associated with scRNAseq from the genuine biological variation of gene expression [23]. Therefore, we performed independent biological repetitions of the experiment (Fig. S1a–g). Most of the cells after oxidative stress were again positive for SA ß-galactosidase activity (Fig. S1a). Consistent with the first experiment, we found that after oxidative stress, cells separated into two transcriptional fractions, while low serum- and nutlin-treated cells formed singular transcriptional communities (Fig. S1b). Analysis of the two OxStress fractions confirmed that they are distinguished by the same reproducible markers (Fig. S1e–g, Supplementary Spreadsheet File S3). Thus, we found reproducible transcriptional heterogeneity among senescent cells.

We considered a possibility, that one of the OxStress fractions corresponds to cells that resumed proliferation after stress. However, proliferating cells in S and G2/M phases were clearly distinct from other clusters in transcriptional space (see Figs. 1c, f) and only a small number of cells from OxStress, LowSerum, and Nutlin samples co-localized with untreated cells in S/G2/M phases in transcriptional space (see Figs. 1c, f). Consistently, we observed a large number of mitotic cells on untreated plates, but not on LowSerum-, OxStress-, or nutlin-treated plates (not shown). Hence, neither of the large transcriptional fractions of OxStress cells represent cells that resumed proliferation.

Next, we examined the possibility that either of the OxStress fractions may be representing cells that were somehow not strongly affected by oxidative stress. Two lines of evidence argue against this possibility. First, oxidative stress rendered a vast majority of cells SA ß-galactosidase-positive in the initial experiment, inconsistent with the possibility that a large fraction of OxStress cells were not strongly affected by the oxidative damage (see Fig. 1b). Second, we performed another independent repletion of the experiment with harsher oxidative stress treatment: cells were treated with a higher concentration of hydrogen peroxide and were treated with it twice (Fig. S1 h–n). Almost 100% of cells were SA ß-galactosidase-positive after the treatment (Fig. S1h), but the OxStress cells were still split into two transcriptional fractions (Fig. S1i–k) distinguished by a similar set of marker genes as seen earlier (Fig. S1i–n, Supplementary Spreadsheet File S4). Thus, in this repetition of the experiment with harsher oxidative stress and almost all cells being SA ß-galactosidase-positive, both fractions of OxStress cells were still present indicating that both fractions represent senescent cells. Two fractions were distinguished by the same marker genes as in previous independent repetitions (Fig. S1 l–n).

In summary, we performed 3 independent experiments with consistent outcomes. Our results show that, upon oxidative stress, most cells subsequently exhibit signs of senescence, yet they fall into two distinct transcriptional states. The observed transcriptional heterogeneity is in line with the recent studies showing changes in transcriptional heterogeneity in vivo and in vitro with aged cells [8–13]. Our results highlight that the term “senescent cells” encompasses a heterogeneous array of subtypes of cells that may require additional classification and consideration when attempting to understand or eliminate senescent cells.

Subgroups of oxidative stress-induced senescent cells exhibit distinct functional and metabolic signatures

To better characterize the observed fractions of senescent cells, we performed gene set enrichment analysis (GSEA) and compared these two fractions to each other and cells from other conditions. We chose “Hallmarks” gene sets that represent well-defined biological states or processes and reduce redundancy and noise [24]. First, we compared OxStress-CP and OxStress-TR to quiescent (LowSerum-treated) cells and we noted that both fractions exhibited signs of stress. Compared to quiescent cells, OxStress-CP cells expressed hallmarks “Unfolded Protein Response,” “Reactive Oxygen Species Pathway,” “p53 Pathway,” and “DNA Repair,” while OxStress-TR cells were enriched for a hallmark “Unfolded Protein Response,” but not others (Fig. 2a, b). In another independent repetition (same experiment as in Fig. S1a–g), both fractions exhibited hallmarks of stress when compared to quiescent cells, but the OxStress-CP group exhibited a stronger stress response (Figs. S2a, b). Another distinction between the two fractions was metabolism. OxStress-CP cells were again more distinct from quiescent cells and exhibited hallmarks of the MTORC1 pathway (“MTORC1 Signaling,” “PI3K/AKT/MTOR Signaling”), “E2F Targets,” and “Oxidative Phosphorylation.” Hence, between the two fractions, OxStress-CP mounted a broader stress response and seemed to be in a more active metabolic state compared to OxStress-TR cells.

Fig. 2.

Each OxStress fraction has a distinct functional specialization. a “Hallmarks” gene sets enriched in OxStress-CP compared to Quiescent cells (low serum-treated cells). b Hallmarks gene sets enriched in OxStress-TR compared to Quiescent cells. c Hallmarks gene sets enriched in OxStress-CP compared to OxStress-TR cells (positive values) or in OxStress-TR compared to OxStress-CP (negative values). d Fraction of mitochondrial transcripts in transcriptome of individual OxStress cells. e Fraction of mitochondrial transcripts in Untreated cells and OxStress-CP and OxStress-TR groups. f Enrichment of glycolysis-related genes in individual OxStress cells. g. Enrichment of glycolysis-related genes in Untreated, OxStress-CP, and OxStress-TR groups. * indicates statistical significance with p < 0.05. Statistical testing was performed using non-parametric ANOVA (Kruskal–Wallis test) with Conoverss post hoc test and Holm-Sidak adjustment

Next, we directly compared OxStress-CP and OxStress-TR fractions to each other (Fig. 2c). Consistent with the earlier analysis, OxStress-CP cells were more metabolically active (hallmarks “Myc Targets V1,” “Myc Targets V2,” “E2F Targets,” “mTORC1 Signaling,” “Oxidative Phosphorylation”) and exhibited stronger activation of stress-related pathways (“Reactive Oxygen Species,” “UV Response Up,” “p53 Pathway,” “Interferon Alpha Response,” “Unfolded Protein Response,” “DNA Repair”) than OxStress-TR cells. Also consistent with an earlier notion, OxStress-TR cells were more enriched for cell- and tissue-organization activities (“Apical Surface,” “Apical Junction,” “Epithelial Mesenchymal Transition” Hallmarks). Our analysis showed a similar difference between OxStress-CP and OxStress-TR fractions in a repetition experiment (Fig. S2c). Hence oxidative stress-induced senescence comprised two signaling states with distinct functional specializations.

We performed a more detailed analysis of the two OxStress fractions to understand how the transcriptional changes may relate to differences in cell physiology, focusing on mitochondrial and glycolytic metabolism. We found that the two fractions differ in the abundance of mitochondrial genes, with OxStress-CP cells being more enriched in mitochondrial transcripts (Figs. 2d–e, S2d-e). However, when we compared OxStress cells to the untreated control, we found that the difference was not due to elevated expression of mitochondrial genes in OxStress-CP cells compared to non-stressed cells, but rather due to strongly decreased expression of mitochondrial genes in OxStress-TR cells compared to both OxStress-CP and untreated cells (Fig. 2e, S2e). Importantly, the abundance of the mitochondrial transcripts was regressed from the gene expression datasets before the initial analysis, so the appearance of the two fractions of senescent cells is not a technical artifact due to a variable abundance of mitochondrial transcripts. We then examined the expression of glycolytic enzymes as another measure of metabolic activity. OxStress-CP cells were again more enriched in glycolytic transcripts, underscoring higher metabolic activity of this cell subset (Figs. 2f–g, S2f–g). Interestingly, in one experiment mitochondrial enrichment did not overlap with glycolytic enrichment within the OxStress-CP subset, indicating potential heterogeneity within the OxStress-CP fraction itself. The smaller size of the OxStress-CP fraction in a repetition experiment did not allow such a detailed examination (Fig. S2f). Thus, two fractions of post-stress cells differ in metabolic signaling, and both glycolytic and mitochondrial genes contribute to this difference.

The OxStress-TR fraction exhibited signatures of the TGF-beta Pathway and Notch Signaling (Figs. 2c, S2c). Remarkably, recent studies identified two senescence fates upon oncogene-induced senescence (OIS), one of which was characterized by inflammatory cytokines and the other one by Notch signaling [25, 26]. Hence, we uncovered that a Notch-associated senescent state emerges not only upon OIS but also in the context of oxidative stress-induced senescence.

In the OIS model, TGF-beta/Notch-associated senescence (“secondary senescence”) appeared to be a response to signals secreted by “primary” senescent cells, where Notch was mediating a switch from the “inflammatory” fate to the TGF-beta fate. We wanted to get a better understanding of the senescence fate determination in our system, and how it emerges. It is possible that all cells exhibit the same response to oxidative stress at early stages and the fate split happens later during the stress response. Alternatively, heterogeneity may establish immediately after the stress response. Discriminating between these possibilities will help better understand the mechanisms behind the distinct senescence states.

Transcriptional heterogeneity emerges early during the stress response

To determine if heterogeneity is preexisting before oxidative stress, or if it arises as cells become senescent, we performed a new set of experiments with time course analysis. We analyzed cells before stress (untreated control) as well as 4 h, 1 day, 2 days, 3 days, 4 days, and 7 days after stress (Fig. 3a). Consistent with earlier experiments, cells at the final time point clustered into two transcriptional groups (Fig. 3a). To better analyze the emergence of the distinct post-stress fates, we combined genes that reproducibly distinguish OxStress-CP and OxStress-TR fractions in all experiments into two gene sets (“CP” signature and “TR” signature) (Supplementary Spreadsheet Files S5, 6). We then scored all cells in the time course experiment to determine when “CP” and “TR” signatures emerge. Both signatures were the most pronounced at the final time point and were developing gradually over the course of time (Fig. 3b, c). We then examined each time point more closely. Figure 3d shows cells and enrichment of the two signatures at the different time points. By following the time course from the last time point to the first, we noticed that the clear separation of the two fractions was not apparent until Day 7 (Fig. 3d). At the earlier time points, cells were mostly in a single transcriptional cluster starting from Day 1 (Fig. 3d). Analysis was complicated at earlier time points because of the differences in cell cycle phases (Fig. 3a, d). Notably, at all the time points between Day 1 and Day 7, there was clear antagonism between “CP” and “TR” signatures where cells were leaning toward one or another fate. In one of the independent repetitions, cells separated into two transcriptional groups earlier during the time course (Figs. S3a–d). Yet another independent trial gave similar results to the first one (Figs. S3f–j). In all independent trials, bias toward a particular post-stress fate was observed early during the stress response. Hence, we propose that senescence fate is at least partially determined early after stress.

Fig. 3.

Transcriptional signatures of senescent cells emerge early. Cells were collected at different time points after oxidative stress. a UMAP representation of time course analysis. Cells are colored based on their sample of origin. b, c Enrichment of CP and TR signatures across the time course. CP and TR signatures are sets of genes that distinguish OxStress-CP and OxStress-TR cell fractions (see Supplementary Spreadsheet Files S5,6). d Enrichment of CP and TR signatures at individual time points. e. Fraction of mitochondrial transcripts and enrichment of glycolytic transcripts in CP and TR group cells at Day 1. At this time point, cells were assigned to CP or TR fraction depending on which respective signature was higher. f, g Genes from the universal signatures of senescent fibroblasts that changed expression in either of the OxStress fractions in the same direction (up- or downregulated) as in the universal signature were combined into the respective signatures. Intensity of each signature throughout the time course is shown in UMAP plots * indicates statistical significance with p < 0.05, determined by Mann–Whitney test

We then examined if the metabolic difference between cell fractions also establishes early during the stress response. We found that similar to the later time points, “TR”^high at Day 1 (cells in which “TR” signature scored higher than “CP” signature) exhibited a lower abundance of mitochondrial transcripts compared to “CP”^high cells (Fig. 3e, S3e,j), indicating less metabolically active state. Interestingly, the abundance of glycolytic transcripts did not show a consistent difference between “CP”^high and “TR”^high cells on Day 1 (Fig. 3e, S3e,j). Thus, the cell-to-cell difference in mitochondrial activity is among early signs of post-stress heterogeneity, while the intercellular difference in glycolytic activity develops later. Overall, our data do not support the model of a Notch-mediated switch between senescence states, and instead, these data suggest that the observed senescence states develop simultaneously and are at least partially determined early during the stress response.

In addition, we wanted to examine how fast cells progress toward each senescence state by focusing on senescence-specific genes. For that, we determined genes that consistently discriminate each of the OxStress fractions from untreated cells in all five experiments with single oxidative stress treatment (Supplementary Spreadsheet Files S7–10). We then cross-referenced those genes with the universal signature of senescent fibroblast, focusing on the genes that change expression in at least one OxStress fraction in the same direction as in the universal signature (Supplementary Spreadsheet File S11). We combined each resulting gene set into the respective signatures (e.g. genes upregulated both in the universal signature and in OxStress-CP compared to non-treated cells) and visualized the progression of each signature (Fig. 3f, g). It can be seen from the UMAP plots that signatures change early during the stress response, indicating again that responses leading to either senescence state start to manifest early.

Increased stochastic noise is not a universal feature of senescence

After establishing that oxidative stress-induced senescence is associated with increased variability of signaling states (signaling noise), we asked if senescence also increased stochastic noise, that is an uncoordinated cell-to-cell variation of gene expression. Observable stochastic noise of gene expression strongly depends on the expression level with lower expression being associated with increased noise due to both technical and biological factors [23, 27, 28]. Normalizing read counts across all cells to the same value, which is a typical step in scRNA-seq data analysis, may therefore give exaggerated noise estimates for samples with lower read counts. Hence, for this analysis, we used non-normalized data. We calculated coefficients of variation (CV) as well as the mean expression for each gene in OxStress and untreated cells. The scatter plot in Fig. 4a shows that expression variation for most genes is decreased in OxStress cells. The decrease is likely explained by the increased mean expression level (Fig. 4b). Indeed, we noticed that the CV of some genes increased in OxStress cells relative to the majority of other genes (shown in red in Fig. 4a), and that coincided with the decreased mean expression of these genes after oxidative stress (Fig. 4b).

Fig. 4.

Increased stochastic noise of gene expression is not a universal feature of senescence. a. Coefficient of variation (CV) of genes in untreated and OxStress cells. b Mean expression of genes in untreated and OxStress cells. c Mean expression of genes with similar expression in untreated and OxStress cells. d CV of genes with expression was similar in untreated and OxStress cells. e Coefficient of variation (CV) of genes in quiescent and OxStress cells. f Mean expression of genes in quiescent and OxStress cells. g Mean expression of genes with similar expression in quiescent and OxStress cells. h CV of genes with similar expression in quiescent and OxStress cells. All axes are in log scale

To further examine the changes in the transcriptional heterogeneity, we selected genes with similar average expression levels in both post-stress and untreated conditions (Fig. 4c). We found that the CV of those genes changed little between the two conditions (Fig. 4d). To exclude the effect of the cell cycle, we also compared OxStress cells to quiescent cells and we observed similar results (Fig. 4e–h). An independent repetition produced the same results (Fig. S4). Hence, we did not find evidence for the increased stochastic noise upon oxidative stress-induced senescence.

Expression of senescence markers in different fractions of senescent cells

Identification of reliable markers of senescent cells for research and clinical use is an area of active work. We used our single-cell data to examine the expression of the previously reported universal markers of senescent fibroblasts [7]. When we cross-referenced universal senescence markers of fibroblasts with our list of genes that distinguish either OxStress-CP and OxStress-TR fractions from untreated cells (Supplementary Spreadsheet Files S11), and with genes that discriminate each fraction from each other (Supplementary Spreadsheet Files S5,6). As a result, we obtained senescence markers that are specific to each fraction of senescent cells and changed expression in this fraction compared to untreated cells in the same direction (up- or downregulation) as in the universal signature of senescent fibroblasts. OxStress-CP fraction is characterized by the increased expression of FGF2, IFR2R, GBE1, KTN1, and AKR1B1, and much stronger upregulation of CLCN3 than in OxStress-TR cells. OxStress-CP is also characterized by the decreased expression of MEIS1, SEC24D, and MAGI3. OxStress-TR is characterized by increased expression of RAI14, SMURF2, and PLAT, and much stronger upregulation of RORA compared to OxStress-CP cells. OxStress-TR cells downregulated SPATS2L more strongly than the OxStress-CP fraction. Some of these changes are shown in Fig. 5.

Fig. 5.

Senescent cell biomarkers are differentially expressed between the two transcriptionally distinct senescent cell types. All the included genes show significant (p < 0.05) difference in expression between OxStress-CP and OxStress-TR cells in all experiments with single application of oxidative stress. All the included genes were previously shown to be part of the universal signature of senescent fibroblasts. a Genes with previously reported increased expression in senescent cells that were only upregulated in OxStress-CP, but not OxStress-TR cells compared to untreated cells. b Genes with previously reported increased expression in senescent cells that were only upregulated in OxStress-TR, but not OxStress-CP cells compared to untreated cells. c Genes with previously reported decreased expression in senescent cells with significantly different expression between OxStress-CP and OxStress-TR cells. MEIS1 and SEC24D were only downregulated in OxStress-CP, but not in OxStress-TR cells. SPATS2L was downregulated in both fractions of senescent cells, but retain significantly higher expression in OxStress-TR cells compared to OxStress-CP. Statistical significance was determined by Mann–Whitney test

Finally, we examined what components of senescence-associated secretory phenotype (SASP) characterize each fraction of senescent cells. We used the list of SASP components from the recent proteomic study of SASP induced by different stressors [29]. We again cross-referenced the lists of genes that discriminate each OxStress fraction from each other (Supplementary Spreadsheet Files S5,6), with the lists of genes that consistently discriminated each fraction from untreated cells with the list of SASP components. We thus derived the lists of SASP components that were specific to each fraction of senescent cells. As shown in Fig. S5, each fraction of senescent cells includes some unique SASP components.

Discussion

Recent studies highlighted increased cell-to-cell variation in gene expression with age, both in vivo and in vitro [8–13]. Here, in order to understand how cell-to-cell variation in gene expression originates in senescent cells, we used advances in genomic technologies which allowed scalable and multiplexed single-cell RNA-sequencing. We were able to perform multiple independent experiments which allowed us to extensively characterize a time course of cell-to-cell variation in gene expression as cells become senescent.

In a previous study, senescent cells exhibited increased cell-to-cell variation in gene expression variation for genes measured with single-cell qPCR [8]. We extended prior findings with whole transcriptome analysis. We also found that induction of senescence through molecular damage is associated with increased cellular heterogeneity, with cells grouping into two major transcriptional clusters. This was unlike cells induced into quiescence by low serum treatment or into senescence/senescence-like arrest with nutlin, which formed singular transcriptional clusters. Metabolic difference emerged as a prominent feature that differentiated two post-stress fractions and it was supported by the abundance of both mitochondrial and glycolytic transcripts. Given a central role of metabolic pathways in aging and stress adaptation [30], it is tempting to speculate that differences in metabolic signaling may be a driving force for two distinct cell fates. Further investigations will allow to test this hypothesis and better understand the relationship between metabolism and senescence and potentially different subtypes of senescence.

Our results are consistent with the recent reports of two distinct senescent fates in the oncogene-induced senescence (OIS) model [25, 26]. Similarly to those results, we also observed two senescence fates, one of which is characterized by TGF-beta and Notch signaling. Unlike the OIS model, we did not observe that the non-TGF-beta-associated fraction, namely OxStress-CP, was more pro-inflammatory compared to the TGF-beta-associated fraction, OxStress-TR. Rather, OxStress-CP was more enriched in cytoprotective and cellular stress-response signaling. We propose that this difference with the OIS model is due to the nature of molecular stress, namely oxidative stress that we used in our study.

Similar functional heterogeneity observed in this study and the OIS model suggests the in vivo relevance of this phenomenon. We are yet to learn why two branches of senescence response exist. Based on the functional annotations of signaling pathways characterizing each fraction of senescent cells, we speculate that such heterogeneity represents labor division between cells for complex tissues tissue reorganization by senescent cells, e.g. at the site of injury. This idea is supported by our finding, that SASP profiles of two fractions include some unique components. Distinct contributions of different types of senescent cells into tissue remodeling may afford a more flexible response to injury or other insults, such as oncogenic transformation [31, 32].

We extended prior studies of senescence fates by performing a detailed time course analysis, tracking the emergence of the two transcriptionally distinct senescent cell fates. Our data indicate that distinct cell fates start emerging early during the stress response, potentially reflecting either pre-existing heterogeneity among untreated cells or stochastic differences in molecular damage during treatment, or the interaction of these two factors. Prior reports labeled the observed fates as primary and secondary senescence with the idea that Notch signaling of primary senescent cells gives rise to secondary senescent cells. We are yet to see if a similar hierarchy is applicable in our system, however, the early emergence of fate bias suggests that other mechanisms also come into play, such as pre-existing stochastic differences in cell signaling [33]. It is interesting to notice that nutlin-3a treatment did not give rise to the inflammatory fate, but only to the fate that was transcriptionally close to OxStress-TR (see Fig. 1). This is consistent with the previous report that nutlin-3a attenuates the inflammatory phenotype of senescent cells [22]. It also suggests that intensity and duration or dynamics of p53 signaling may be another determining factor for the particular senescence fate [34–36].

In our experimental system, we did not find increased stochastic noise of gene expression. The cells that we measured were not more dissimilar from one another. These results are distinct from some previous reports, but not all [14]. We do not yet understand the cause of these differences. It seems likely that different cell types and different senescence conditions will have different effects on the amount of cell-to-cell variation in gene expression and other kinds of cellular heterogeneity. It is suggested by Anderson et al. that there will be both abnormal increases and decreases in cellular heterogeneity that will both have pathological consequences [15].

Finally, various types of molecular stress induce the formation of senescent cells with distinct transcriptional profiles. It will be important to analyze these different types of senescent cells at the single-cell level to better understand the source of the differences. We may find that different stressors induce completely distinct types of senescent cells. Alternatively, both shared and distinct subtypes of senescent cells may emerge after different types of molecular stress. Indeed, in our system, we observed that one of the post-stress fractions, OxStress-TR, is transcriptionally more similar to nutlin-treated cells, while OxStress-CP is more distinct. It will be important to expand the single-cell analysis to different types of senescent cells for a better understanding of the senescence and identification of more reliable markers of senescence for research and therapeutic applications.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contribution

NB and ARM conceived the study and designed the experiments. JO advised on cell culture practice and experimentation. NB performed the experiments and data analysis. NB drafted the manuscript. NB, JO, and ARM revised the draft and revisions for reviewers.

Funding

Nikolay Burnaevskiy was supported by the award K99AG061216 from NIA, Junko Oshima was supported by the award R01CA210916 from NCI, Alexander R. Mendenhall was supported by the award R01CA219460 from NCI. Dr. Mendenhall was also supported by the Nathan Shock Center for Excellence in the Basic Biology of Aging center grant from the National Institutes for Health National Institute on Aging, P30AG013280.

Data Availability

The high-throughput sequencing data will be made available through NCBI GEO.

Declarations

Conflict of interest

The authors declare no competing interests.