Single-cell morphology encodes functional subtypes of senescence in aging human dermal fibroblasts

Pratik Kamat; Nico Macaluso; Yukang Li; Anshika Agrawal; Aaron Winston; Lauren Pan; Teasia Stewart; Bartholomew Starich; Nicholas Milcik; Chanhong Min; Pei-Hsun Wu; Jeremy Walston; Jean Fan; Jude M Phillip

doi:10.1126/sciadv.ads1875

. 2025 Apr 25;11(17):eads1875. doi: 10.1126/sciadv.ads1875

Single-cell morphology encodes functional subtypes of senescence in aging human dermal fibroblasts

Pratik Kamat ^1,², Nico Macaluso ^1,², Yukang Li ², Anshika Agrawal ¹, Aaron Winston ¹, Lauren Pan ³, Teasia Stewart ¹, Bartholomew Starich ^1,², Nicholas Milcik ^2,³, Chanhong Min ^2,³, Pei-Hsun Wu ^1,², Jeremy Walston ⁴, Jean Fan ³, Jude M Phillip ^1,^2,^3,^5,^6,^*

PMCID: PMC12024660 PMID: 40279419

Abstract

Cellular senescence, a hallmark of aging, reveals context-dependent phenotypes across multiple biological length scales. Despite its mechanistic importance, identifying and characterizing senescence across cell populations is challenging. Using primary dermal fibroblasts, we combined single-cell imaging, machine learning, several induced senescence conditions, and multiple protein biomarkers to define functional senescence subtypes. Single-cell morphology analysis revealed 11 distinct morphology clusters. Among these, we identified three as bona fide senescence subtypes (C7, C10, and C11), with C10 exhibiting the strongest age dependence within an aging cohort. In addition, we observed that a donor’s senescence burden and subtype composition were indicative of susceptibility to doxorubicin-induced senescence. Functional analysis revealed subtype-dependent responses to senotherapies, with C7 being most responsive to the combination of dasatinib and quercetin. Our single-cell analysis framework, SenSCOUT, enables robust identification and classification of senescence subtypes, offering applications in next-generation senotherapy screens, with potential toward explaining heterogeneous senescence phenotypes based on the presence of senescence subtypes.

Single-cell morphologies reveal functional subtypes of senescence.

INTRODUCTION

Senescence represents a heterogeneous cellular phenotype characterized by stable proliferative arrest (1–3), up-regulation of cyclin-dependent kinase inhibitors such as p16 and p21 (2, 4, 5), increased secretions of pro-inflammatory molecules (6–8), and acute changes in cellular and nuclear morphologies (9–13). Although recognition of senescence-associated changes has been critical to shaping our current understanding of senescence across cell populations, it remains limited in precision. The limitations arise because of the following: (i) Although a handful of protein-based biomarkers and curated gene expression signatures (14, 15) is used as proxies for senescence (2, 5), now, there are no absolute senescence biomarkers. (ii) It is unclear whether the accumulation of senescent cells within aging tissues is primarily due to inefficiency in their clearance or whether certain cells are more prone to becoming senescent. In addition, (iii) cellular senescence is routinely defined as a binary phenotype, meaning that cells are either senescent or not (16, 17). Furthermore, the consideration of senescence as binary brings implied assumptions that all senescent cells are in some ways equivalent and may respond similarly to biological perturbations and senotherapies (18, 19). However, mounting evidence demonstrates that senescence is heterogeneous and context dependent, conferring both favorable and deleterious effects based on the biological environment (5, 20–24). Addressing these limitations is key to developing a deeper understanding of cellular senescence and the factors that determine functional phenotypes (25).

Advances in single-cell technologies and modern data-science approaches have led to innovative methods for profiling senescent cells (9–11, 26). Studies using next-generation sequencing technologies on bulk and single-cell populations have demonstrated reliable methods for profiling enormous quantities of senescent cells (14, 15). These approaches have enabled the simultaneous interrogation of multiple transcriptionally defined signatures and putative molecular profiles of senescence. For example, multiple subpopulations of fibroblasts expressing distinct gene patterns have been identified in human skin (21, 27), leading to questions regarding whether the subpopulations exhibit varied susceptibility to senescence induction and dissimilar roles in aging-related pathogenesis. A report using WI-38 fibroblast cells identified the emergence of distinct senescence clusters based on differential, single-cell gene expression patterns that were dependent on the mode of senescence induction (28). While it is intriguing that transcriptionally defined senescence subtypes exist and are based on how senescence was induced, performing single-cell sequencing brings the challenges of prohibitive cost, limited throughput, and low feasibility of profiling multiple samples and conditions. Collectively, these advances expose the need for robust, cost-effective, and high-throughput approaches to profiling senescent cells at a single-cell resolution.

Recently, single-cell strategies have used high-content microscopy coupled with machine learning to profile senescent cells cultured on flat substrates coated with physiologically relevant extracellular matrix proteins (9–11, 13). These studies used images of baseline and induced senescent cells (e.g., DNA damage induced and serial passaging) to train computational models and neural networks to identify senescent cells. Furthermore, they used downstream validation of senescence, performed according to assessed expression of protein-based senescence biomarkers or 5-ethynyl-2′-deoxyuridine (EdU) incorporation, to denote cell proliferation (9–11). Of note, one study demonstrated that training their model on the nuclear features of cultured senescent cells allowed direct application to predicting the senescence burden of intact tissue sections (10). While these recent studies provide a proof of concept for identifying senescent cells using imaging and machine learning, they relied on fundamental assumptions related to the following: (i) a binary phenotype (i.e. cells existing as either senescent or not), despite growing recognition of a diverse spectrum of senescence phenotypes; (ii) all treated cells were considered senescent, while all untreated cells were reported as nonsenescent (9–11). While these assumptions could be valid in certain biological contexts, accepting this generalization could bias the interpretation of results and measures of senescence phenotypes in other, naturally occurring situations. Collectively, image-based approaches coupled with machine learning can provide a powerful solution to the robust identification of senescence phenotypes. However, it is critical to go beyond binary classifications to investigate the presence and functions of putative senescence subtypes among the spectra of cell conditions.

Here, we present an integrated imaging and machine learning framework, termed SenSCOUT (senescence subtype classifier based on observable unique phenotypes). SenSCOUT is a coupled experimental-computational workflow framework for identifying and classifying functional subtypes of senescence based on specific properties of single cells including cellular and nuclear morphologies and biomarker expression. Using primary dermal fibroblasts from two donors, aged 23 and 85, we profiled the morphologies of cells at the baseline, after serial passaging (replicative senescence), and 8 days postinduction using four chemical inducers. By profiling single cells across each of these biological conditions, we identified 11 morphology subtypes, among which three proved to be bona fide senescence subtypes (C7, C10, and C11). Senescence subtypes were determined on the basis of exhibited morphology and the expression of multiple protein-based senescence biomarkers. As part of this workflow, we implemented and validated a label propagation scheme using imputation based on k-nearest neighbors. This imputation approach enabled us to quantify and validate the expression of multiple protein-based biomarkers for each cell (up to five biomarkers per cell) and to link expression profiles directly to distinct morphologies.

In addition to quantifying cellular and nuclear morphology features, we developed a machine learning scheme based on the Xception (29) architecture to analyze and classify cells and develop a senescence score per cell based on raw textured images of F-actin (cell) and DNA (nuclei). Using this senescence score, we showed that the senescence burden across cell populations was dependent on the mode of senescence induction, time postinduction, and the age of the donor. We also observed that the mode of senescence induction was critical to defining the magnitude and the rate of development of the senescence phenotypes. Notably, cells induced with doxorubicin (DOX) exhibited robust induction kinetics with a decrease in the abundance of cells within nonsenescent clusters and sustained enrichment of the bona fide senescence subtypes. However, cells induced with hydrogen peroxide (H₂O₂) exhibited biphasic induction kinetics. This means that although the population of nonsenescent clusters decreased at early time points (with a complementary increase in the fraction of senescence subtypes), there was a reemergence of nonsenescent clusters at later time points postinduction. Moreover, rather than representing a reversal of senescence, this reemergence of nonsenescent clusters was due to the proliferation of cells that did not convert toward senescence upon exposure to H₂O₂—so-called escaper cells.

To further establish the association between senescence, senescence subtypes, and age, we used primary dermal fibroblasts from an expanded cohort of 50 healthy donors ranging in age from 20 to 89 years. We observed that not only did the number of senescent cells increase with age but that senescence subtype C10 exhibited strong age dependence. Moreover, we discovered that the baseline senescence burden within a cell population was indicative of the senescence score after exposure to the senescence inducer DOX. Last, to determine whether the identified senescence subtypes encoded functional information, we profiled the responses of senescent cells to a small panel of senotherapies, with a focused analysis on responses to dasatinib + quercetin (D + Q). The results showed that senescent cells responded in four dominant patterns: short death, long death, stable viable, and unstable viable. Senescent cells treated with D + Q exhibited higher abundance in short death, long death, and unstable viable, with senescent cells in C7 being most sensitive to death after D + Q exposure. Our analyses provided a robust, single-cell framework using imaging, morphological profiling, and machine learning to identify and classify functional subtypes of senescence across multiple biological conditions.

RESULTS

Multiple induction conditions and biomarkers enable the robust profiling of cellular senescence

Cellular senescence can be induced across ex vivo cell populations using a variety of methods that take advantage of core mechanisms such as DNA damage responses, replication stress and telomere shortening, oncogenic activation (i.e., p53 and RAS), and epigenetic modifications (5, 7, 30–32). To recapitulate the range of physiologically relevant senescence phenotypes, we established a framework using primary dermal fibroblasts isolated from a 23 and 85 year old, referred to as young and old, respectively. To profile senescence, we exposed cells from both donors to optimized concentrations of four chemical inducers [i.e., bleomycin (BLEO), DOX, atazanavir (ATV), H₂O₂, and serially passaged cells (REP)] until proliferation was arrested (see the Materials and Methods) (7, 30, 33, 34). After induction, the senescence phenotype was allowed to develop over the course of 8 days, after which cells were plated at low density onto collagen I–coated glass-bottom dishes, fixed and stained for proteins/biomarkers of interest, and then imaged for downstream single-cell morphology analysis (Fig. 1A). For all experiments reported in this study, unless otherwise stated, cells were stained with Hoechst 33343 (H33342) to delineate nuclear boundaries; phalloidin, an F-actin stain to delineate cell boundaries; senescence biomarker p21; and either lamin B1 (LMNB1), high-mobility group box 1 (HMGB1), p16, or β-galactosidase (βGal) (Fig. 1B and fig. S1, A and B) as biomarkers of senescence (1, 5, 35, 36). To quantify single-cell morphologies, uninduced and induced cells across both donors were segmented using an optimized CellProfiler pipeline (37) and processed through an in-house pipeline for data curation (i.e., removing mis-segmented cells and outliers) and analysis (see the Materials and Methods). Comprehensively, we generated an initial dataset consisting of >50,000 single cells, each having information related to cell and nuclear morphology and the expression of multiple protein-based senescence biomarkers for both ages across five senescence-induction conditions.

Fig. 1. — (A) Graphical illustration of experimental workflow to induce and characterize senescent cells. (B) Representative images of the immunofluorescence staining of five senescence-associated protein biomarkers across untreated (DMSO vehicle) and DOX-induced primary dermal fibroblast samples (23-year-old male). (C to G) Z-score fluorescence quantification relative to control of β-galactosidase (βGal) (C), p16 (D), p21 (E), HMGB1 (F), and LMNB1 (G) for young and old dermal fibroblast samples [n > 400 single cells per condition, means ± 95% confidence interval (CI), one-way ANOVA relative to age-matched control expression, P ≤ 0.001 relative to control]. (H) Heatmap of standard scaled expression of pro-inflammatory secretions for induction conditions. GM-CSF, granulocyte-macrophage colony-stimulating factor; IFN-γ, interferon-γ; IP-10, interferon γ induced protein 10 (also referred to as CXCL10); TNF-α, tumor necrosis factor–α. (I) Principal components analysis scatterplot of secretory profiles across uninduced and senescence-induced populations. Similar clustered groups are highlighted in red and blue transparent ovals. (J) Truncated waterfall plot of cellular and nuclear morphological feature enrichment to either untreated or senescence-induced samples across both ages (P ≤ 0.001 for all enrichments).

High-content imaging enables the profiling of senescence across biological conditions

To provide in-depth profiling of cells across all conditions, we quantified the expression of protein-based senescence biomarkers, cellular secretions, and morphological features describing cells and their corresponding nuclei. To establish and validate the senescence phenotypes across induced cell populations, we confirmed proliferation arrest by measuring EdU incorporation, monitoring Ki67 expression, and conducting direct cell counts (fig. S1, C to G). Given the verified proliferation arrest observed on the basis of single-cell counts and reduction of EdU- and Ki67-expressing cells, we then quantified the expression of multiple senescence biomarkers in cells from both donors across the control and the four senescence induction conditions. As expected, we observed an increase in the average expression of β-galactosidase, p16, and p21 and a decrease in the expression of HMGB1 and LMNB1 for both donors relative to their baseline uninduced cells (Fig. 1, C to G, and fig. S1H). We also observed that the extent of increase or decrease in the biomarker expressions was dependent on the mode of senescence induction, with slight shifts as a function of donor age (fig. S1H). Together, these results suggest that the mode of senescence induction contributes to the differential abundances of senescent and nonsenescent cells observed across the tested biological conditions.

To further establish senescence phenotypes, we profiled cellular secretions at the baseline and 8 days following senescence induction. We performed secretion analyses using the Bruker CodePlex multianalyte secretory panel, which contained analytes (cytokines, chemokines, etc.) overlapping with those identified in the core senescence-associated secretory phenotype (SASP), such as interleukin-6 (IL-6), IL-7, IL-8, IL-15, and MIP-1a (macrophage inflammatory protein–1a) (8, 38). The abundance of secreted factors was quantified by culturing 50,000 cells for each condition in six-well dishes with 5 ml of supplemented media for 48 hours, harvesting the conditioned media, and profiling the undiluted conditioned media. From this analysis, we profiled 22 secreted factors across 10 conditions. The results showed that uninduced baseline cells for both ages clustered together and were defined by low expression of SASP-related and other pro-inflammatory proteins (Fig. 1, H and I). The senescence-induced conditions, on the other hand, clustered together with elevated expressions of SASP-related proteins. Consistent with the protein-based biomarkers, we also observed induction-dependent trends in the secretions. For example, IL-7 was secreted in abundance by cells from both the young and old donors treated with DOX but showed expression reduced by 45 and 43%, respectively, when young and old cells were exposed to H₂O₂. Collectively, the analyses of cellular secretions provide additional confirmation of bulk senescence phenotypes across induction conditions.

Last, in collecting microscopy images of cells and nuclei across all conditions, we quantified senescence-associated changes in cellular and nuclear morphology for all treatments and controls. Consistent with the published literature, we observed statistically significant increases in the size of cells and nuclei postinduction (Fig. 1B) (1, 2, 5, 12). For each cell within our dataset, we quantified over 250 morphological parameters, which were reduced to 87 key parameters using factor analysis for those that best described the size, shape, curvature, and roughness of cells and nuclei (see the Materials and Methods and data S1) (39). To assess the specificity of this developing senescence phenotype, beyond simple increases in cell and nuclear size, we investigated which other morphological features were most strongly associated with uninduced cells and senescence-induced cells, respectively. Performing differential-morphology analysis on the induced and uninduced cells, we observed that most of the highly differential morphological features, among both induced and uninduced cells, were associated with nuclear morphology (Fig. 1J). However, features describing nuclear shape (e.g., more circular nuclei) were more strongly associated with the uninduced cells, whereas features describing the nuclear size (e.g., larger nuclear area) were more strongly associated with senescence-induced cells (fig. S2A). These results agree with recent findings showing that nuclear features strongly encode senescence information (10, 11) and may be more correlated with the senescence phenotype than whole-cell morphology changes alone.

Comprehensively, the results show that all biological conditions used in this study identified consistent senescence phenotypes. This provides a solid substantiation of the single-cell senescence framework described here and its potential for critically evaluating the emergence of functional senescence subtypes.

Single-cell morphology encodes heterogeneous senescence phenotypes in dermal fibroblasts

Given the capacity for cell and nuclear morphologies to encode senescence phenotypes, we developed a single-cell framework to identify and classify morphological subtypes of senescent and nonsenescent dermal fibroblasts. To establish this framework, cells from both donors across uninduced and senescence-induced conditions were imaged using high-content microscopy (see the Materials and Methods). Following imaging, cell and nuclear boundaries were segmented using optimized pipelines in CellProfiler (37). We then computed parameters describing cellular and nuclear morphologies, as well as the expression of two protein-based senescence biomarkers for each cell (each cell was costained for p21 combined with one of the following: p16, β-galactosidase, HMGB1, or LMNB1; see the Materials and Methods) (37).

Using this methodology, we pooled data across 12 biological conditions, consisting of cells from the young and old donors at the baseline (uninduced) and postinduction with optimized concentrations and durations of ATV, DOX, BLEO, H2O2, and REP (see the Materials and Methods and Fig. 2A). Data visualization using two-dimensional uniform manifold approximation and projection (UMAP) revealed specific cell localization patterns per condition, with cells under the uninduced conditions localizing to the lower left of the manifold and cells from senescence-induced conditions localizing toward the right of the manifold (Fig. 2B). Although different inducers exhibited varying degrees of spread across the manifold, there were core overlap areas for all inducers. To gain a deeper understanding of these spatial patterns across the manifold, we performed k-means clustering analysis to identify morphologically distinct groups of cells. k-Means analysis revealed 11 distinct morphological clusters, each having characteristic cell and nuclear morphologies (Fig. 2C and fig. S2, B to D).

Fig. 2. — (A) Schematic of the morphological analysis pipeline. (B) UMAP plot constructed using 88 morphological parameters across all conditions and age groups (n = 50,000 cells). Dots indicate individual cells. (C) Overlay of the 11 k-means clusters with representative cellular and nuclear morphologies. Contour lines indicate Gaussian kernel density function for each cluster. (D) Select morphological features layered on the UMAP space. Navy and yellow colors delineate low and high, respectively. UMAP-1 (x axis) was correlated with nuclear and cellular sizes, and UMAP-2 (y axis) was correlated with cell shapes. (E) Heatmap of fractional abundance for cells within each cluster for each experimental group (n > 1000 cells for each condition, dendrograms based on averaged Euclidean distances). (F) Contour overlay for nonsenescent and senescent classifications for βGal, p16, p21, HMGB1, and LMNB1. (G) Heatmap of the average biomarker expression for each cluster. (H) UMAP plots for imputed and measured biomarker expressions across young and old samples. (I) Imputation error using morphological parameters of single cells to predict protein biomarker expression (n > 500 cells for each biomarker, means ± 95% CI, multiple comparison Tukey test, P ≤ 0.001 for HMGB1 and LMNB1 relative to other biomarkers). Accuracy is defined as one minus the percent deviation from the model-predicted value divided by the ground truth value. (J) Heatmap of average imputation model accuracy for each biomarker in each morphological cluster (n > 200 cells per cluster per biomarker).

Notably, visual inspection of cellular morphologies indicated a pattern of increasing size from left to right and increasing linearity of the shapes from top to bottom of the manifold. These results were confirmed by color coding the magnitude of key morphological parameters onto the UMAP, with increased cell and nuclear area from left to right and increasing cell aspect ratio (corresponding to a decrease in cell circularity) from top to bottom (Fig. 2D). Notably, cluster C9 contained a large fraction of mis-segmented cells. As such, we retained C9 as a quality control cluster (mis-segmented cluster) but omitted cells classified in C9 from formal analysis and reporting in the results.

For these morphological clusters, we assessed the presence and abundance of cells per morphological cluster across induction conditions (fig. S2, E and F). Unsupervised hierarchical clustering of induction conditions per cluster revealed the emergence of three groups, depicted by the dendrogram branches in fig. S2 (E and F). Uninduced conditions for young and old donors and young cells induced with ATV comprised group 1; both young and old cells induced with DOX, BLEO, and H₂O₂ comprised group 2; and young and old cells induced via serial passaging (REP) and old cells induced with ATV comprised group 3 (Fig. 2E). Group 2 containing young and old cells treated with DOX, BLEO, and H₂O₂ exhibited enrichments in clusters C6, C7, C8, C10, and C11, whereas group 1 was depleted in those clusters but enriched for C1, C2, C4, and C5. Group 3, on the other hand, contained a mixture of enriched morphological clusters, with noted overlaps with both group 1 and group 2, and general enrichments in C4, C5, and C8. These associations between induction condition and morphological clusters point to the putative strength of the senescence inducers, with group 2 conditions being stronger inducers of the senescence phenotypes. Furthermore, we noticed that all induction conditions exhibited higher heterogeneity based on the Shannon entropy relative to uninduced conditions for both ages, suggesting that induced cells exhibited a wider range of morphologies relative to uninduced cells (fig. S2G).

The results show a robust framework for morphologically defining senescence phenotypes across multiple biological conditions. These findings suggest that investigating morphology-defined clusters may reveal functionally distinct groups of senescent and nonsenescent cells.

Single-cell morphology robustly encodes senescence biomarker expression

Current approaches to identifying senescence across cell populations rely on profiling the presence and expression of protein-based biomarkers, including p16, p21, HMGB1, LMNB1, and β-galactosidase (16, 35, 40). For example, compared to nonsenescent cells, senescent cells express higher levels of p16, p21, and β-galactosidase and decreased levels of HMGB1 (translocation out of the nuclei) and LMNB1. While these expression patterns have been collectively established across multiple studies, many studies evaluating senescence have profiled only one to two biomarkers (16). To provide a more comprehensive assessment of expression patterns across various biological conditions, we profiled all five of the biomarkers listed above. Because our imaging workflow was limited to monitoring four fluorescence channels simultaneously, we devised a multifaceted staining strategy, pairing p21 with one of the four other biomarkers. This was iterated across all conditions, a process through which biological replicates were stained with different biomarker combinations until all five biomarkers were represented across each of the 12 biological conditions (see the Materials and Methods). With this workflow, we quantified the localization of cells expressing elevated or depleted levels of senescence biomarkers within our UMAP space.

Subsequently, we investigated how biomarker expression is influenced by the diverse morphological parameters documented in the UMAP findings. To determine a strict, quantitative threshold for each biomarker for nonsenescent and senescent cells, we plotted and overlaid the distributions of single-cell expressions under the uninduced and senescence-induced conditions (fig. S2, H to J). From these distributions, we identified appropriate thresholds based on the realization that senescent cells would be enriched under the induced conditions and display lower representation in the uninduced states. Pragmatically, we defined thresholds for each biomarker based on the magnitude of the shifts in the distribution of postinduction single-cell expression (see the Materials and Methods). We found that even under the induced conditions, not all cells exhibited elevated senescence biomarker expressions, suggesting that not all cells shifted toward a biomarker-defined senescence phenotype. The results showed a clear separation of cells classified as senescent (red contours) from nonsenescent (blue), with senescence emerging from left to right of the manifold (large cells with large nuclei were localized to the right), which confirmed qualitative observations presented earlier (Fig. 2F). Intriguingly, while the UMAP contours showed clear overlap areas among all senescence biomarkers, unique topologies were apparent for each biomarker, suggesting that some biomarkers may identify cells as senescent, while others may lack this capacity. For example, the contours for senescent cells identified by p16, β-galactosidase, and LMNB1 had slightly less spread across the manifold relative to p21 and HMGB1, indicating that the use of HMGB1 or p21 alone may identify more cells as senescent relative to using p16, LMNB1, or β-galactosidase. Whether p21 or HMGB1 is more, or less, accurate remains to be determined; however, a combination of multiple biomarkers offers greater effectiveness in capturing senescence phenotypes.

In the next stage, we quantified the average expression of each senescence biomarker per cluster (average across individual cells in each cluster). The results showed that clusters C7, C10, and C11 had high relative expression of p21, p16, and β-galactosidase and low expression of HMGB1 and LMNB1, suggesting an enriched senescence phenotype within these clusters. Clusters C1, C2, C4, and C5 had the opposite expression pattern, suggesting that they were enriched for nonsenescent cells. Last, clusters C3, C6, and C8 showed intermediate expression of senescence biomarkers, suggesting that these clusters harbored an intermediate state of cells either approaching senescence (i.e., senescence “poised”) or expressing incomplete senescence induction (Fig. 2G). In summary, these findings highlight a strong relationship between cellular and nuclear morphologies and the expression of senescence biomarkers across our morphology-defined clusters, underscoring the notion that morphology robustly encodes heterogeneous senescence phenotypes.

Imputing senescence biomarker expressions based on single-cell morphologies

Given the robust association between single-cell morphology and the expression of senescence biomarkers, we investigated the possibility of using individual cell morphology to predict biomarker expression for each of the five biomarkers. To study this, we developed and implemented an imputation-based label propagation scheme using a k-nearest neighbor algorithm (41, 42). Briefly, we constructed an imputation scheme by leveraging cell and nuclear morphologies to determine senescence biomarker expression based on a cell’s proximity to other cells with known expression of a specific biomarker (fig. S3A). For example, to impute the expression of p16 in a cell stained for p21 and LMNB1 (but not stained for p16), we identified the 20 nearest cellular neighbors that had measured expression of p16 and then weighted the expression values based on the Euclidean distance from each of the 20 neighbors (see the Materials and Methods). This was iterated for every cell across all senescence biomarkers until we had generated an expression profile for each of the five senescence biomarkers across all cells. This dataset contained two measured and three imputed biomarker expression profiles for each of >50,000 individual cells (Fig. 2H). Analysis of the results revealed that the imputation algorithm was robust, with average prediction accuracies greater than ~85% for each of the senescence biomarkers (see the Materials and Methods) (Fig. 2I).

Unexpectedly, we observed that some cells had a low imputation accuracy, particularly for HMGB1 and LMNB1 (β-galactosidase, p21, and p16 had fewer of those cells). We deduced that this may indicate that these biomarkers have a lower dynamic capacity for identifying senescent versus nonsenescent cells and also that the distances of cells from a neighbor with known expression could be larger. However, even with the subset of cells having lower accuracies, the pattern of localization for the imputed cells across the manifold was consistent across both ages and for cells with measured expression for each senescence biomarker (Fig. 2H). To leverage the morphologically defined clusters, we assessed whether imputation accuracies differed for each cluster (Fig. 2J). Computing the average imputation accuracy score for each biomarker across morphological clusters, we observed a stable gradient of accuracy across all clusters. Notably, clusters C10 and C11 were larger and exhibited slightly higher accuracy. This finding highlights the robust performance of our imputation model, regardless of individual cell morphology.

Using our comprehensive dataset, containing detailed information regarding the five senescence biomarkers for each of >50,000 cells, we investigated patterns in biomarker expression to determine whether specific biomarkers identified senescence more effectively than others. We analyzed all cells across all biological conditions and grouped the cells according to the number of senescence biomarkers that were positively expressed using the previously described threshold criteria. Within each comparison (i.e., two, three, four, or five senescence biomarkers), we quantified the fraction of cells in which each biomarker was expressed (fig. S3B). We found that among all cells in which two senescence biomarkers were positively expressed, ~50% were classified as senescent based on biomarkers p21 and HMGB1. Of the cells classified as senescent based on three biomarkers, more than 70% involved β-galactosidase, HMGB1, and LMNB1. Furthermore, in more than 80% of the cells considered senescent based on the expression of four biomarkers, we found that β-galactosidase, HMGB1, and LMNB1 were differentially expressed (e.g., HMGB1 and LMNB1 assessed as low while β-galactosidase assessed as high). In addition, in the same group of cells characterized by the expression of four biomarkers, the fourth biomarker was p16 with a positive expression in ~70% of the cells. Within this framework, we also mapped the localization of cells with a varying number of senescence biomarkers from zero (i.e., nonsenescent cells) to five (senescence based on all five biomarkers). Qualitative results recapitulated the shift of left to right from zero to five markers, with a tight spread corresponding to the overlap areas across all the biomarkers (fig. S3C). In a comprehensive view, we found that information on protein-based senescence biomarker expression is nested within senescence-associated cell morphologies, offering the capability for imputing biomarker expression based on cell and nuclear morphology alone.

Machine learning strategies reveal morphological senescence subtypes

While we were able to establish robust associations between cell and nuclear morphologies and senescence phenotypes at a single-cell resolution, these findings were generated from computing parameters describing measured cell and nuclear morphologies, which included size, shape, curvature, and boundary roughness (data S1). To extract information for individual cells that may not be fully understood by measuring their morphological features, we implemented a machine learning strategy using raw textured images of cells (stained for F-actin) and nuclei (stained for DNA). The goal was to generate a senescence score for individual cells. To accomplish this, we constructed a three-layer model for individual cells, in which the first layer is the nucleus, the second layer is the cell’s F-actin framework, and the final layer consists of a composite image including the F-actin and nuclear structures. Each of these individual cell models (composite images) was cast onto a blank canvas and downsampled to 256 by 256 pixels, with the geometric centroid of each object (nucleus, cell, and composite) mapping to the center of the canvas. We hypothesized that, in addition to capturing the ensemble of morphologies associated with senescence, each data model also offered unbiased insights into which cell properties are most relevant to the identification and classification of senescence. The outputs of each model were labeled as either “nonsenescent” or “senescent” based on the previously established thresholds for indicating the expression of two measured senescence biomarkers per cell (fig. S2, H to J).

We trained our machine learning algorithm using the Xception (29) convolutional neural network architecture with a SoftMax output (Fig. 3A) (10). Briefly, Xception leverages convolution and depth-wise separations to provide an accurate and robust approximation for image-based classifications. SoftMax is a normalized exponential function and the final layer of the Xception architecture that outputs a probability for classification. In this case, the SoftMax output ranged from “0” to “1” based on the probability of a cell being senescent or not according to the threshold biomarker expressions. We implemented a balanced cohort of 6000 data runs spanning all induction conditions, both ages, and all five senescence biomarkers, with the algorithm trained for 250 epochs (fig. S4, A and B). The trained algorithm had an average accuracy of 89% for the training, test, and validation datasets (Fig. 3B). To further demonstrate the robustness of the algorithm, we plotted the receiver operator curve (ROC) of the test set, which yielded an area under the curve (AUC) of 0.95 (Fig. 3C).

In an effort to gain a deeper understanding of what morphological features the model was using to define senescence scores (e.g., the probability of a cell being senescent or not), we constructed a correlation matrix of the senescence scores for each cell versus the quantified morphological parametric data used previously in Fig. 2. We found strong correlations between senescence scores and parameters describing cell and nuclear sizes, with the Pearson correlation coefficients for nuclear area and equivalent diameter being ~0.8 and the same parameters of the cells being ~0.76. While these types of correlation analyses offer insights to connect interpretable morphologies to the Xception model, the actual morphological features most weighted by the model may be poorly translated into measured morphological features.

Previous studies using morphology to identify senescent versus nonsenescent cells have often used either nuclear (10, 11) or cellular (9) measurements. While both have proven effective in capturing aspects of the senescence phenotype, there has been limited investigation using both cellular and nuclear morphologies. To address this in an unbiased manner, we converted training and test datasets to include exclusively either nuclear or cellular data. The resulting datasets were used to iteratively train new algorithms for 50 epochs, and we computed accuracies based on protein-based biomarker expression. Notably, both models resulted in statistically similar accuracies, indicating that cell and nuclear morphologies provide useful information on senescence phenotypes (fig. S4C). To further illustrate this, we computed the classification error between the senescence scores per cell and the senescence classification per biomarker. Results showed an error of less than 0.1 (or 10%) across all senescence biomarkers, indicating that the Xception algorithm accurately captured the senescence phenotype in an unbiased manner (Fig. 3E).

While our trained algorithm robustly quantifies senescence scores ranging from 0 to 1 for individual cells, we wanted to determine a rigid threshold above or below which we could confidently determine a cell as senescent or not (with the underlying notion that “senescence” encompasses a spectrum of senescence subtypes). Conventionally, binary SoftMax outputs are rounded to the nearest integer value for classification, but we wanted to quantify how specific scores correlate with the accuracy of the classification. Thus, we needed to determine, for example, how the error of classifying a cell with a senescence score of 0.6 would change if the score had been 0.7. To assess this, we binned the range of senescence scores between 0 and 1 with increments of 0.02 and then determined how many cells in each bin were correctly classified on the basis of their biomarker expressions. From this analysis, we selected a score above 0.88 as senescent and below 0.1 as nonsenescent, corresponding to a prediction error below 15% (Fig. 3F and fig. S4D). To visualize how the algorithm determines senescence scores relative to the morphology-defined clusters, we projected the senescence score per cell across all clusters. Consistent with the senescence phenotypes based on biomarker expressions per cluster (Fig. 2G), we found that clusters C1, C2, C4, and C5 exhibited low senescence scores (<0.3); clusters C7, C10, and C11 exhibited high senescence scores (>0.9); and C3, C6, C7, and C8 exhibited intermediate scores on average (0.6 to 0.9) (fig. S4E). We show that combining our Xception-trained model with morphological parameters and senescence biomarkers yields a robust classification of senescence, with senescence scores increasing from left to right of the manifold (fig. S4F) and clusters C7, C10, and C11 exhibiting scores above 0.9 with large cell and nuclear morphologies (Fig. 3, G and H).

Last, given the high senescence scores for cells in C7, C10, and C11, we reevaluated the concept that different induction conditions result in differential senescence phenotypes. Results indicated induction-dependent senescence phenotypes with varying abundance of cells in each of the three senescence clusters (Fig. 3I). For example, cells from the young donor induced with BLEO harbored 22 and 21% of cells in C7 and C10, respectively, whereas cells from the old donor induced with H₂O₂ harbored 23% of cells in C11, only 10% of cells in C7, and 12% in C10. Therefore, we show the robust identification and classification of senescence phenotypes based on morphology, with validation metrics based on multiple protein-based senescence biomarkers. In addition, the identified morphology clusters exhibit varying senescence scores, with C7, C10, and C11 having a robust senescence phenotype and each having distinct patterns of morphological parameters (Fig. 3, J to L). Hereafter, we refer to clusters C7, C10, and C11 as senescence subtypes. Although clusters with intermediate senescence scores such as C6 and C8 may exhibit features of senescence, their classification at this point is uncertain, as they could represent a presenescent or poised senescence phenotype.

Emergent senescence phenotypes depend on inducer-specific kinetics

While our previous analysis provides a robust framework to describe morphological patterns of cellular senescence, it was developed on snapshots at a single time point (day 8). To establish robust temporal patterns across the progression of senescence, we performed senescence induction on cells from the young donor using DOX and H₂O₂, with snapshots of morphologies collected at multiple time points, specifically at 0, 2, 5, 8, 12, and 15 days postinduction. At each of these time points, we repeated the experimental and imaging workflows described previously (see the Materials and Methods) to collect information on cellular and nuclear morphologies, as well as the single-cell expressions of the five senescence biomarkers (Fig. 4A). From this analysis, we observed differences in the senescence progression based on whether cells were treated with DOX or H₂O₂ (Fig. 4, B and C). For the DOX-induced condition, cells exhibited a steady progression toward a senescence phenotype, with a decrease of cells in C1, C2, C4, and C5 as a function of time postinduction and a complimentary increase of cells in C3, C6, C7, C10, and C11 until day 15 (Fig. 4D and fig. S5, A to C). Conversely, H₂O₂-induced cells exhibited a biphasic progression of senescence, with cells in C1, C2, C4, and C5 decreasing to day 5 and then steadily increasing again. Simultaneously, cells in C6, C7, C10, and C11 followed the opposite pattern, increasing to day 5 and decreasing thereafter until day 15 (Fig. 4E and fig. S5, B and D). The depletion of cells within clusters is not due to the cells dying but rather to a fractional redistribution of cells within clusters as they progress toward senescence. Furthermore, the apparent recovery of senescence over time for cells treated with H₂O₂ is not a reversal of the senescence phenotype but represents a subset of cells (escapers) not converting toward senescence upon exposure to the inducer and proliferating to critical mass.

To quantify trends in the senescence subtypes, we profiled the senescence enrichments across the cell populations in both DOX-induced (Fig. 4F) and H₂O₂-induced (Fig. 4H) cells. For DOX-induced cells, there were a steady increase in senescence enrichment and a biphasic pattern for H₂O₂. Quantifying the temporal patterns for enrichment in each of the three senescence subtypes, we found that, for DOX, the fraction of cells classified as C7 decreased over time with a simultaneous increase in C10 and C11 and convergence of the fractional abundance of all senescence subtypes to ~30% (Fig. 4G). This trend was not as clear for H₂O₂-induced cells (Fig. 4I). To help explain these results, we also quantified the temporal patterns of expression in each of the protein-based senescence biomarkers for both DOX and H₂O₂ conditions. Consistently, early time points for both inducers showed deviations from baseline uninduced cells. However, at later time points, H₂O₂ regained biomarker levels like uninduced cells (Fig. 4, J to N). Collectively, these results suggest that the temporal patterns of senescence are inducer dependent, with cell populations fractionally redistributing among senescence subtypes during the progression. Furthermore, H₂O₂ may be a weaker inducer relative to DOX, with a higher fraction of “escapers” or cells not entering senescence, thereby resulting in a recovery and a subsequent increase in the proliferative cell population. This phenomenon is corroborated by enrichments of control-like morphological clusters at long time points (i.e., C1 and C2), as well as a corresponding convergence of senescence-associated protein biomarker expression to levels seen in controls.

Baseline senescence burden predicts susceptibility to senescence induction with DOX

Multiple studies have reported an increased abundance of senescent cells with increasing chronological age (2, 43–45). To quantify the relationship between chronological age and the abundance of senescent cells within the defined subtypes, we procured a gender-balanced cohort of low-passage, primary dermal fibroblasts from 50 individuals with an age range of 20 to 89 years (Fig. 5A). To evaluate patterns in the occurrence of senescent cells with age, we profiled the morphology of baseline uninduced cells ex vivo collected from each of the 50 donors and plotted the relationship between their chronological age and the average senescence score of the sampled cells. The results showed a positive correlation between the senescence score and the age of the donor with a Pearson correlation coefficient of 0.33 and P < 0.05 (Fig. 5B). This trend was conserved for cells from men and women (fig. S6, A and B) with Pearson correlation coefficients of 0.32 and 0.34, respectively. On the basis of this result, we investigated whether the abundance of cells classified as specific senescence subtypes was more strongly associated with chronological age. Plotting the fraction of senescent cells within each senescence subtype as a function of chronological age, we found that C10 had the strongest age association, with a Pearson correlation coefficient r = 0.45 and P = 0.002, followed by C11 (r = 0.30 and P = 0.041) and C7 (r = −0.16 and P = 0.286) (Fig. 5, C to E). Furthermore, other morphology clusters also exhibited age associations, such as C8 (r = 0.42, P = 0.003) (fig. S6, C to L). Although subtype C10 showed the strongest age association, it comprised a minority of the total senescence population, with fractions ranging from 2.5 to 17.5% across all ages assessed. C7, on the other hand, comprised most of the senescent population with a range of 20 to 90%, and C11 ranged between 0.5 and 3.5%. The initial induction experiments showed that C7 and C11 were the most dominant senescence subtypes, suggesting that they may be strongly associated with damage responses rather than chronological age. Collectively, these results indicate that a morphology-based classification of senescence recapitulates the overall age-related accumulation of senescence, while senescence subtypes exhibit varying degrees of age dependence.

Fig. 5. — (A) Graphical depiction of hypothesized divergent senescence susceptibility responses as a function of age. (B) Average senescence score of primary dermal fibroblast samples as a function of chronological age fit to a univariate linear regression model (n > 300 cells per sample, means ± SEM). (C to E) Fraction of senescent cells within each of the three morphological subtypes as a function of chronological age fit to a univariate linear regression model (from top to bottom: clusters 7, 10, and 11, respectively). (F) Double violin plot for age-dispersed senescence score distributions at the baseline (gray) and after DOX induction (red) (n > 100 cells per condition). YO, years old. (G) Linear regression model of the average DOX-induced senescence score against the chronological age of the sample. (H) Linear regression model of the average DOX-induced senescence score against the average baseline senescence score of the sample. (I) Senescence susceptibility postulate where the induction response is governed by baseline morphological profiles.

Given the relationship between age and senescence score, we wondered whether cells from older donors would be show greater susceptibility to senescence induction when exposed to appropriate inducers. To assess this, we selected DOX, given its clinical relevance to the treatment of cancers. From the expanded cohort of aging samples, we selected 20 randomly, spanning the entire age range from 20 to 89 and including the cohort’s range of baseline senescence scores. For each of the aging samples, we induced senescence using the previously described optimized induction protocols and profiled the senescence score as a function of chronological age (Fig. 5F and fig. S6, M and N). Plotting the average DOX-induced senescence score versus age, we did not observe a statistically significant correlation, with a Pearson correlation coefficient of 0.19 (Fig. 5G). This result is expected given the large donor-to-donor variability observed in senescence scores, even among donors of similar chronological age (Fig. 5B).

To gain deeper insight into factors associated with senescence susceptibility, we reanalyzed the pre– and post–DOX induction distribution of senescence scores of cells from each donor (Fig. 5H). Plotting the relationships between senescence scores at the baseline and postinduction, we observed a positive correlation with a Pearson correlation coefficient r = 0.54 and P = 0.016, suggesting that the baseline senescence score was indicative of postinduction score, likely representing a more sensitive predictor of senescence response than just chronological age. This suggests that senescence susceptibility is greater in cell populations having a higher senescence burden. The baseline senescence score was not indicative of the senescence subtype enriched after DOX induction (fig. S6, O to Q).

These results support the hypothesis that, while senescence burden across cell populations accumulates with increasing age, the relationship between chronological age and senescence susceptibility is more complicated. Furthermore, cell populations with higher senescence scores (i.e., indicator of senescence burden) at the baseline are more susceptible to senescence induction, as noted by the higher senescence scores postinduction (Fig. 5I).

Senescence subtypes exhibit differential responses to D + Q

With the emergence of senotherapies as a clinically viable strategy for mitigating senescence-associated dysfunctions, we conducted a preliminary inquiry to assess whether senescence subtypes exhibit differential responses to a small panel of senotherapies (18, 46–48). This exploration was stimulated in considering that, if cells of various senescence subtypes responded differently to various senotherapies, it could indicate functional differences among the senescence subtypes. To assess this, we profiled the responses of senescent cells to optimized concentrations of five clinically relevant senotherapies, comprising both senolytics [D + Q, ARV-825, and navitoclax (ABT263)] and senomorphics (fisetin and metformin) over the course of 3 days (see the Materials and Methods) (19). For this experiment, we induced cells from the young donor (GT22) with DOX as previously described and then cultured cells for 8 days, allowing senescence to progress. On day 7, the cells were trypsinized and seeded at low density onto collagen I–coated, 96-well, glass-bottom dishes, and on day 8, the cells were stained with live-cell markers (Spirochrome 650-FastAct and Spirochrome 555-DNA) to delineate cell (F-actin) and nuclear boundaries, respectively, and plates were scanned using high-content live-cell imaging. Baseline images of live cells were recorded, and the cells were subsequently treated with the panel of senotherapies.

Over the course of 72 hours, images of each condition (five senotherapies and untreated controls) were acquired hourly for a total of 60 hours and then analyzed to profile single-cell morphologies (see the Materials and Methods). Cells within each of the conditions were segmented at each time point, and each cell was assigned to a respective morphology cluster and senescence subtype. We chose to perform live-cell imaging on the senescent cells exposed to senotherapies specifically to map the dynamic response patterns at a single-cell resolution over time. Across all experiments, we recorded specific observations regarding cell death, whether cells exhibited plasticity among morphology clusters (i.e., transitioned among senescence subtypes or the other morphology clusters), and whether they were stable within their initial morphological cluster (Fig. 6A).

Fig. 6. — (A) Visualization of hypothesized differential therapeutic response profiles of senescence morphological subtypes. (B) Quasi-population level counts of senescent cell morphological subtypes for untreated, D + Q, metformin, navitoclax, fisetin, and ARV-825 conditions as a function of treatment duration time. The cell number was normalized relative to the count at the baseline for each condition (t = 0 hours, n > 150 per subtype per treatment condition). (C and D) Single-cell trajectory viability analysis of morphological subtypes as a function of treatment duration time for untreated (C) and D + Q treatment (D) conditions. Cells from the induced population were differentiated into the various senescence morphological subtypes based on morphology at the baseline (t = 0, n > 100 cells per cluster, Mann-Whitney P < 0.05 for all group comparisons between clusters); cell trajectories were considered nonviable after entering k-means clusters 1, 2, or 3 and subsequently losing the tracking ability in subsequent frames. The gray “Sen. avg.” line represents the weighted average viability trends. (E) Heatmap displaying representative morphological evolution of single cells for untreated and D + Q treatment conditions starting in one of the three senescence subtypes. Colors coordinate with k-means cluster colors, with black indicative of a cell trajectory that has lost tracking (n = 480 cell samples, ward cluster method, Euclidean distance, threshold of 190). Dendrogram clustering was used to determine four overarching response modalities. (F) Untreated versus DOX enrichment in each of the four response modalities identified in the trajectory heatmap. (G) Morphological subtype enrichment in each of the four response modalities identified in the trajectory heatmap, normalized to cell count. DOX-induced senescent cells in the sample from the 23-year-old male were used for this analysis.

With these time-lapse videos, we profiled bulk viability across the six tested conditions, which indicated a decreased viability for cells treated with D + Q, navitoclax, and ARV-825, while fisetin and metformin elicited viability responses like the untreated controls (fig. S7, A and B). In addition, while most of the tested conditions had an average senescence population of ~60% postinduction and treatment, the D + Q condition had a lower fraction of ~40%, showing greater killing (fig. S7, C and D), and was the only treatment condition showing a bulk decrease in cell count for all three of the senescence subtypes (Fig. 6B). Given the relevance of this result to ongoing clinical trials using D + Q, we decided to perform deeper analysis on D + Q responses. We quantified the fraction of senescent cells in the entire population and then within each of the three senescence subtypes as a function of exposure time for the untreated and D + Q–treated senescence-induced cells. The results indicated a decrease of ~10% among untreated senescent cells over the 60-hour experiment (Fig. 6C). C10 and C11 exhibited a stable trend over the 60 hours, while C7 exhibited a decrease to ~80%, which could be explained by spontaneous cell death (Fig. 6C). Conversely, D + Q–treated senescent cells exhibited >30% decrease among the overall fraction of senescent cells, with a complimentary decrease in the abundance of all three senescence subtypes (Fig. 6D). Cells in C10 had higher fractional viability after D + Q treatment relative to C11 and C7, while C7 exhibited the largest decrease in the fraction of viable cells over the 60-hour treatment.

To quantify these responses further, we constructed a single-cell response matrix for all senescent cells classified as C7, C10, or C11 at initiation of the experiment for both untreated and D + Q conditions. We performed unsupervised hierarchical clustering along the time series to identify response patterns (fig. S7, E to I). Qualitatively, we observed unique and shared response patterns across both untreated and D + Q–treated conditions. To enable a direct comparison among responses for untreated and D + Q senescent cells, we pooled the responses of both conditions and then reclustered the time series data using unsupervised hierarchical clustering. The results revealed four response patterns that were defined according to cell viability. We named these responses long death, short death, stable viable, and unstable viable. Short death referred primarily to cells cleared within the first 30 hours, long death were cells cleared at later time points (>30 hours), stable viable were cells that remained viable but exhibited little to no transition to other morphology clusters or other senescence subtypes, and unstable viable were cells that remained viable but exhibited high temporal plasticity and transitions to other morphology clusters (Fig. 6E).

Next, we quantified the abundance of cells within each of the response patterns from the untreated and D + Q conditions. As expected, the results showed that senescent cells treated with D + Q comprised a majority of the long and short death classes at 60 and 79%, respectively. Of the cells remaining viable over the 60 hours, the stable viable cells were under the untreated condition (84%), with 47% of the unstable viable cells coming from the D + Q treatment. These results suggest that D + Q treatments drove a strong cell death response (Fig. 6F). Further investigating the influence of the starting senescence subtypes, we observed that C7 was dominant in both the short and long death outcomes, comprising 47 and 43% of the cells, respectively. C10 was most dominant in the stable viable comprising 53% of cells, with C7 being dominant (43%) in unstable viable and C10 and C11 exhibiting a similar result (29%) (Fig. 6G and fig. S7J). We found also that cells classified as stable viable starting in C7 exhibited transitions primarily to C10 or C11, on average comprising ~28.7 and 27.9% of those clusters, respectively. Cells starting in C10 remained primarily in C10 (63.4%) and C11 (23.9%), and cells starting in C11 either remained in C11 (46.2%) or transitioned to C10 (39.9%) (fig. S7K). Of the unstable viable cells, senescent cells starting in C7 spent most of their time in either C7 (53.7%), C6 (21.2%), or C4 (7.9%). Senescent cells starting in C10 were morphologically very heterogeneous, spending most of the time in either C7 (30.5%), C10 (23.6%), C6 (17%), or C11 (9.3%). Last, senescent cells starting in C11 spent most of their time in C6 (37.1%), C11 (20.1%), C7 (19.5%), or C10 (5.1%) (fig. S7L).

Collectively, these results indicate that senescent cells within each of the three senescence subtypes responded differentially to D + Q, suggesting that they are functional subtypes of senescence. Notably, senescent cells in C7 were most responsive to D + Q, with cells either dying or becoming morphologically unstable during treatment. Alternatively, C10 senescent cells were most resilient to D + Q treatment. These findings also demonstrate that ex vivo morphological profiling of senescence subtypes can be a powerful tool for screening and mapping response patterns to senotherapy treatments.

DISCUSSION

Cellular senescence is a heterogeneous phenotype (23, 49). In summary, these investigations demonstrated the development of a robust single-cell framework, SenSCOUT, for classifying functional subtypes of senescence. Leveraging high-content microscopy coupled with machine learning in the development of single-cell profiles, we classified a panel of heterogeneous aging dermal fibroblasts across multiple induction conditions into 11 morphology clusters, among which we identified three bona fide senescence subtypes. This collaborative study demonstrates the following biological findings: (i) Expressed senescence phenotypes are dependent on the age of the donor, the mode of senescence induction, and the time interval since senescence induction. (ii) Single-cell morphology encodes senescence phenotypes and can be used to predict protein-based biomarker expression using imputation via k-nearest neighbor approaches. (iii) It is not appropriate to assume that all cells exposed to senescence inducers are senescent, because not all cells transition toward senescence when exposed to senescence inducers. Furthermore, a cell population exposed to a senescence inducer may harbor subpopulations of escaper cells that, subsequently, can proliferate to recover the population, as we observed with H₂O₂ induction. (iv) Various senescence subtypes do not exhibit the same degree of age dependence (e.g., in these studies, C10 was most strongly associated with the chronological age of donors and most resilient to D + Q treatments). (v) Chronological age by itself does not determine a cell’s susceptibility to senescence induction. (vi) Senescence subtypes exhibit differential responses to senotherapy (D + Q). We provide a robust and promising approach to ex vivo senescence profiling with potential applications in precision medicine.

Advances in single-cell technologies and machine learning have enabled deep profiling of cellular phenotypes. These and other recent studies applying similar techniques to understanding senescence have expanded the foundation for identifying senescence and gauging responses to prosenescence (inducers) and antisenescence (senotherapies) perturbations (9–11). However, many previous studies have considered senescence as a binary cellular state (cells are either senescence or not) rather than a heterogeneous collection of dynamic phenotypes. We demonstrated that subtypes of senescence do exist and that their existence could help explain the heterogeneity and context dependence of senescence phenotypes. Our approach provides insights and practical applications of ex vivo senescence phenotypes and cellular morphologies. Furthermore, because our framework is high throughput and cost effective, with the capacity to interrogate both snapshot and dynamic responses, it has direct applications for next-generation screening across a wide range of pro- and antisenescence perturbations. In addition, this framework could be integrated with other image-based or next-generation sequencing-based spatial, multiomic technologies to identify subtype-dependent molecular vulnerabilities and potentially allow target-specific senescence phenotypes.

While this single-cell, image-based approach holds much promise, it was developed ex vivo and has not yet been validated within the context of complex tissue structures. While cells do retain critical memory, for example, of their age and disease status, even when taken outside of their physiological context (50), it will be critical in future studies to investigate how the presence and abundance of senescence subtypes exist in vivo. To accomplish this, we will need to map senescence subtypes using features that translate well from ex vivo to in vivo, and vice versa. Because cell morphology does not translate directly from cell culture to tissues, we will need either nucleus-based signatures or robust molecular signatures (e.g., transcriptomic, epigenomic, and proteomic), which can be established and validated ex vivo for interrogation within tissues (intact or deconstructed tissues) using high-content imaging or spatial technologies. Furthermore, it will be interesting in future studies to investigate how the presence and abundance of specific senescence subsets alter or contribute to the senescence niche, for example, within aging skin and other aging tissues, or how targeting a particular cell subtype could disrupt the senescence niche or induce beneficial effects.

Since a previous study by Heckenbach et al. (10) demonstrated that an ex vivo trained model can be used to predict senescence burden within intact tissues, nuclear features may provide a means to identify senescence subtypes but only if we can accurately classify senescence subtypes with nuclear features alone. To assess this, we trained a model to classify senescence subtypes using either nuclear morphologies alone or cell morphologies alone. Our results indicate that a combination of cellular and nuclear morphologies appears most accurate currently. With these two models, we observed a high accuracy of senescence subtype classification using only nuclear morphologies with an average accuracy of 81.6%. Using cell morphologies yielded an average accuracy of only 69.1% (fig. S8, A and B). Together, these are exciting results, indicating that nuclear features can be used to classify cells into senescence subtypes. However, it remains to be determined how these models will work for cellular components within intact tissues.

Recent studies have also demonstrated the notion of molecular subtypes of senescence (26, 28). Considering these studies and others defining molecular signatures of senescence (14, 15), it would be interesting to learn the extent to which these molecular profiles associate with and map onto our morphologically defined senescence subtypes. Furthermore, we chose to define senescence subtypes as morphology clusters that have high average senescence scores (>0.9). Although nonsenescent clusters (C1, C2, C4, and C5) did harbor low senescence scores (<0.3), there were a few clusters, namely C3, C6, and C8, with intermediate senescence scores ranging between 0.6 and 0.82. Additional research is needed to determine whether these cells are in fact “poised” toward senescence, exhibiting a higher tendency toward senescence conversion, and what molecular features best describe these phenotypes. Having a better understanding of this will provide opportunities to target precursor senescence phenotypes, which may prove beneficial in certain biological contexts such as aging, for example. Furthermore, several studies have demonstrated the multifaceted roles of senescence (i.e., beneficial, deleterious, and ambiguous). By placing a greater emphasis on defining subtypes of senescence, we can better understand the context-dependent nature of cellular senescence. This is crucial, as targeting and eliminating senescent cells in beneficial contexts could result in physiological dysfunction. In summary, we present an innovative and robust approach to profiling senescence and senescence heterogeneity using a robust, cost-effective, and versatile single-cell approach.

MATERIALS AND METHODS

Cell culture

Low-passage human dermal fibroblasts were procured from the Baltimore Longitudinal Study on Aging to serve as the experimental basis for this work. We used 50 primary cell samples from a gender-balanced cohort with ages ranging from 23 to 89 years. Primary cells were cultured under standard conditions on cell culture–treated culture dishes (Corning, 353136) or on collagen I–coated glass-bottom dishes (Cellvis, P24-1.5H-N) with Gibson High-Glucose Dulbecco’s modified Eagle’s medium (Thermo Fisher Scientific, 11995073) supplemented with 15% fetal bovine serum (Thermo Fisher Scientific, A5256801) and 1% penicillin-streptomycin (Thermo Fisher Scientific, 15070063). Cells were grown at 37°C in a humidified incubator with 5% CO₂. Cells were passaged on average every 3 to 4 days. Once cultures reached ~70% confluency, they were plated onto 10-cm cell culture–treated petri dishes (Corning, 430167) at a density of 100,000 cells per dish and allowed to adhere overnight before senescence induction. Low-passage cells (maximum of four passages from stock vial) were used for all experiments, except when stated otherwise (i.e., for replicative senescence experiments).

Induction of senescence

All drugs used for senescence induction were diluted from their respective stock solutes in dimethyl sulfoxide (DMSO; <0.2%, v/v) according to the manufacturer’s guidelines and subsequently diluted in media to reach the final working concentration for senescence induction (see table S1 for respective concentrations). Briefly, cell culture media were removed from the petri dishes and replenished with 8 ml of induction media (culture media supplemented with respective concentrations of senescence inducers) and incubated for the specified time durations (table S1). Following induction, cells were washed twice with 1× phosphate-buffered saline (PBS; Corning, 21-031) before addition of 8 ml of fresh culture media. The senescence phenotype was allowed to develop, and cellular profiling was performed at various time points depending on the experiment (either 2, 5, 8, 12, or 15 days postinduction). For each of the specified induction time points, cells were subsequently plated onto collagen I (50 μg/ml; Corning, 354249)–coated 24-well glass-bottom plates (Cellvis, P24-1.5H-N) at a density of 1250 cells per well and a volume of 1 ml per well to allow for a sparse density and identification of single-cell morphologies. The end point for replicative senescence assessment for each sample (GT22 or GT125) was determined on the basis of a sustained plateau of proliferation rate for 1 week.

Profiling cellular secretions

To profile the secreted factors from biological conditions of interest, senescent and nonsenescent cells were plated at a density of 20,000 cells per 10-cm petri dish and allowed to adhere overnight. The following day, the media were replaced with fresh media and incubated for 48 hours to allow the accumulation of cellular secretions. After 48 hours, the conditioned media were collected, and cells were counted per condition (i.e., per 10-cm dish). Conditioned media were then profiled using Bruker CodePlex Human Adaptive Immune Panel (Bruker Cellular Analysis, CODEPLEX-2L01-4) to quantify the expression of 22 pro-inflammatory markers that overlap with the documented secretions of senescent cells (SASP) (5, 8). The raw readouts were calibrated to instrument controls and normalized using the individual cell counts to enable direct comparisons of secreted factors across biological conditions.

Immunofluorescence staining

Before immunofluorescence staining and quantifications, cells for each condition were seeded at low density (1250 cells per well of a 24-well plate) and allowed to completely adhere overnight. Once cells had adhered, they were subsequently washed with 1× PBS and fixed in freshly prepared 4% (w/v) paraformaldehyde (Electron Microscopy Services) for 12 min. Cells were then rinsed three times with PBS before being permeabilized by incubation in 0.1% Triton X in PBS solution for 10 min. Cells were again rinsed three times with PBS to remove any residual Triton X solution and then blocked with 2% bovine serum albumin (BSA; weight/volume in 1× PBS) for at least 30 min at room temperature. Following this blocking step, cells were incubated with various primary antibodies, including p16 (Abcam, ab108349, 1:250), LMNB1 (Abcam, ab229025, 1:500), HMGB1 (Abcam, ab18256, 1:500), or γ-H2AX (Abcam, ab81299, 1:250) in 1% BSA overnight at 4°C. The next day, cells were washed three times with PBS and incubated in p21 (Santa Cruz, SC-817, 1:200) in 1% BSA for 1 hour and 15 minutes at room temperature. Cells were washed three times with 1× PBS to remove any residual primary antibodies, and then they were incubated in a cocktail of secondary antibodies and small-molecule fluorophores, including Hoechst (Thermo Fisher Scientific, H3570, 1:250), Alexa Fluor 488 phalloidin (Thermo Fisher Scientific, A12379, 1:200), goat anti-rabbit IgG-Alexa Fluor 568 (Thermo Fisher Scientific, A-11036, 1:500), and goat anti-mouse IgG-Alexa Fluor 647 (Thermo Fisher Scientific, A-21240, 1:500) in 1% BSA. Last, cells were washed three times with 1× PBS and imaged using a Leica Stellaris 5 Confocal Microscope. All samples were either imaged immediately after staining or within 24 hours of secondary antibody incubation and intermediate storage at 4°C.

Senescence-associated β-galactosidase fluorescence staining was performed using a similar workflow as previously described. On the basis of the manufacturer’s protocol (Thermo Fisher Scientific, C10851), we incubated cells with fluorescence SA-gal before the addition of the p21 primary antibody. Following the incubation with the p21 primary antibody, the experimental workflow was consistent with that described above.

High-content microscopy and quantification of morphological features

Fluorescence images of all conditions were acquired using a Leica Stellaris 5 Confocal Microscope at 20× magnification using four laser lines (405 Diode, 488 Diode, 568 Diode, and 647 Diode). Images were taken at a resolution of 1024 by 1024 pixels and 0.568 μm per pixel. Individual nuclei and cell boundaries were segmented from raw TIFF files using CellProfiler in combination with in-house curation pipelines to ensure well-segmented single cells (51). Briefly, an immunofluorescence-focused segmentation algorithm used the DAPI (4′,6-diamidino-2-phenylindole) stain to delineate the nucleus boundaries and the phalloidin stain to delineate the general cell shape (39). We then used an in-house curation pipeline to remove cells in silico from high-density regions (more than three touching cells) to ensure the quantification of single cells. Touching cells made up a very small portion of the total cells because of optimized density and plating. Using the generated masks, we computed morphological features primarily describing each cell’s size, shape, boundary roughness, and boundary curvatures (data S1). We leveraged these masks to quantify the cellular expression of protein-based biomarkers of senescence for each cell. We profiled >250,000 single cells from photo images spanning all the observed biological conditions presented within this manuscript (not including single-cell tracking data).

With the masks generated as outputs from the CellProfiler pipelines, we used an in-house feature quantification pipeline to record ~200 morphological parameters for each cell (39). The morphological parameters ranged from basic geometric features such as area, perimeter, and curvature to more complex interpretations of morphology such as skeletal features. To identify key parameters driving variance within the cell sample set, we performed primary factor analysis (52) across all cells and all morphological parameters. Primary factor analysis was used to identify the most pertinent variables in describing the reduced latent space and, by extension, the most important variables in the full vector dataset. Parameters with a communality value below 0.2 were excluded from subsequent analysis. Note that a higher communality value indicates that a parameter captures more variance within a given cell population and is more relevant to the latent space (fig. S2B). This factor analysis resulted in a reduction from the total of 218 morphological features to 88 relevant features that were used for all subsequent assessments. Enrichment analysis of morphological features was conducted by comparing the average Z-score expression between noninduced and induced populations across the distribution of relevant morphological features. For interpretation, a large positive Z-value would indicate strong enrichment of a particular parameter in the induced population, whereas a large negative value would indicate a strong enrichment of that parameter in the uninduced population (note that zero signifies no enrichment in either the uninduced or induced population).

Analysis of protein-based biomarkers

Fluorescence-labeled protein-based biomarkers were analyzed at a single-cell resolution by quantifying the pixel intensity within the cells and nuclei. Specifically, HMGB1 and LMNB1 were quantified using the mean nuclear intensity (integrated intensity for each object) (i.e., cell or nucleus) divided by the total number of pixels (i.e., object size), p21 was quantified by total nuclear expression (i.e., summation of the pixel intensity of each nucleus), and p16 and βGal were quantified by total cellular expression (i.e., summation of the pixel intensity of each cell, including the nuclear region). To correct for variations between batches, we used reComBat software (https://github.com/BorgwardtLab/reComBat) for all datasets across biological replicates (53). Protein biomarker expressions for all cells across induction conditions were log transformed before Z-score normalization. The control distributions for each cell were used to compute the Z-score values for cells based on their individual protein expression (performed for each age sample).

To determine the relative cutoff/threshold for senescence (senescent versus nonsenescent cells), we plotted the biomarker distribution of baseline cells and induced cells and then identified the intersection point that defined the crossover for each of the distributions (i.e., cells with low values in the induced distributions but were high under the baseline conditions). Because the biomarker data were log normalized, the resulting normal distributions yield a clean Z-score to differentiate senescent and nonsenescent populations. This analysis was originally conducted independently for each age (because of our hypothesis that the older patient would be more enriched in senescent cells at the baseline) and each biomarker across BLEO, DOX, and H₂O₂ induction conditions (recognizing that BLEO and DOX are DNA damage response agents and that H₂O₂ is a reactive oxygen species inducer). This enabled a determination of the average intersection of the control distribution and the distribution of induced senescence conditions for the five biomarkers (fig. S2, H to J).

The average Z-scores of intersection across both ages, inducers, and biomarkers were ~1.3 for young and 1.0 for old. To account for variability and ensure that all biomarker intersections were included in the stricter cutoffs, we adjusted these thresholds by ~5% of the population metrics. The final Z-score cutoffs were determined to be 1.6 for young (GT22) and 1.3 for old (GT125), with the lower cutoff in older samples reflecting a higher baseline expression of the senescence phenotype. These values correspond to ~5% of healthy GT22 cells and 10% of GT125 cells being classified as senescent. Notably, these cutoffs are more stringent than those of any single biomarker intersection value for each age group, allowing for some flexibility while refining a more selective senescent pool for downstream analysis.

UMAP and k-means clustering

We constructed the reduced dimensionality framework using UMAP across all biological conditions and replicates (n = 3) for GT22 (23-year-old) and GT125 (85-year-old). All morphological parameters were independently log normalized and standard scaled (Z-score). This “preprocessed” morphological parameter dataset was subsequently used to construct a 2D-UMAP space. UMAP is a nonlinear dimensionality reduction algorithm that captures and projects the structure of highly dimensional data in a lower-dimensionality space (for this work, the 88-vector space was reduced to two with Euclidean convergence) (54). Each point in the UMAP space represents an individual cell. Considering that UMAP is a nonlinear reduction across multiple morphological parameters, there is no “equivalent” for each of the constituent vectors. However, UMAP-1 correlated well with cell and nuclear size, while UMAP-2 correlates with cell and nuclear shape. As a complement to the UMAP morphological analysis, we performed k-means clustering to differentiate and identify distinct morphologically defined groups within the dataset. We performed k-means clustering analysis using 11 clusters, which was determined on the basis of the inertia and silhouette values. The goal was to identify the number of clusters that would minimize the inertia (typically at the elbow) and maximize the silhouette values (i.e., stability of the respective clusters). Across the 11 clusters, we identified bona fide senescent and nonsenescent cell populations; however, cluster 9 harbored a large fraction of mis-segmented cells. For initial analysis, cluster 9 was isolated for cell classification but was removed from the analysis when reporting findings. To quantify morphological heterogeneity, the Shannon entropy was calculated using the following equation for the 10 validated k-means clusters (i.e., 11 clusters with the removal of cluster 9)

S = - \sum_{i = 1}^{10} p_{i} \cdot log (p_{i})

Here, S is the Shannon entropy (greater magnitude signifies a more heterogeneous population), and p_i is the fraction of cells classified within morphological cluster i (here, i range from 1 to 10).

Protein biomarker imputation

Considering the interconnectedness of morphology and senescence-associated protein biomarker expression, we developed and implemented a computational pipeline to determine biomarker expression based solely on cell and nuclear morphology parameters. This was particularly useful because our imaging workflow allowed staining a maximum of two biomarkers simultaneously. This biomarker imputation workflow enabled us to determine the expression of five biomarkers for a single cell across each of the biological conditions of interest. Principal components analysis was performed on the 88 log-normalized, standard-scaled morphological parameters that were used to construct the 2D-UMAP to identify the five principal components that capture 95% of the sample variance. For a given cell with an unknown biomarker, the 20 nearest-neighbor cells that were stained for that biomarker of interest were identified using the weighted Euclidean distance of the five principal components. For computational streamlining, nearest neighbors were restricted to the same cell line (GT22 or GT125) within the same k-means clusters. After the nearest neighbors were identified, we computed the expression using a weighted average of the biomarker. The weights were determined by Euclidean distance to the imputed cell, with cells in closer proximity given greater weight. These steps were iterated for each of the biomarkers.

Kinetics of senescence experiments

Human dermal fibroblast samples were induced with either H₂O₂ or DOX, and the senescence phenotypes were allowed to develop over 2, 5, 8, 12, or 15 days in separate dishes. At these time points, cells were seeded at low density on collagen I–coated 24-well glass-bottom plates. Cells were fixed, immunofluorescence stained for senescent-associated protein biomarkers, and imaged as described above. These data were analyzed for biomarker expression behavior and morphological cluster enrichment as a function of senescence induction time.

Machine learning approach to quantifying senescence subtypes using the Xception algorithm

Xception is a 48-layer convolutional neural network architecture for image classification, which we used to identify nonparameterized trends in morphological signatures associated with cellular senescence (29). This architecture was based on the original namesake using two key adjustments and 1,000,000 total trainable parameters: (i) Weights were initialized using ImageNet, and (ii) additional pooling, dropout, and SoftMax layers were appended to the architecture, allowing a senescence probability as the model’s final output. To construct the model input samples, a three-layer dataset consisting of 1536- by 1536-pixel images was created for individual cells in which the first layer consisted of the textured nucleus image (grayscale H33342 image), the second layer contained the textured actin architecture (grayscale phalloidin image), and the third layer was a composite overlay of both the nucleus and actin (combined H33342 and phalloidin images). Each image was geometrically centered and had pixel intensities normalized between 0 and 1. This dataset was constructed from the same cells used to create the UMAP manifold (GT22 and GT125 dermal fibroblast samples across the different induction conditions). Each layer of the input sample set was overlaid on a neutral background (pixel value of zero). Each sample of the training dataset was tagged as either senescent or nonsenescent based on protein biomarker expression Z-score cutoffs (described above). Specifically, cells expressing multiple senescence biomarkers (at least two measured biomarkers) were tagged as senescent. For computational efficiency, the data instances were downsampled into 256- by 256-pixel images.

Because there was a surplus of senescent-tagged cells compared to controls (because of the five senescence induction conditions and one control), we subsampled the training set from the complete dataset to balance the overall population of senescent and healthy cells. The training set also was balanced for age, and the senescent-tagged cells were balanced for induction conditions to minimize bias. Nonsenescent-tagged cells were biased toward control samples (>80% of samples were control conditions in either age group). A generative model was applied to augment the dataset by implementing nondistortive transformations (i.e., flips and rotations) to maintain the integrity of size and shape effects of the training set. Training of the model was conducted in two parts, with an 80-20 train-test split (fig. S4A). The first 200 epochs were trained with ImageNet-initialized weights, frozen backbone (indicating that all weights, except those in the final layers, retained locked values), Adam’s learning rate of 1 × 10⁻⁴, and accuracy-based optimization. After a plateau in frozen backbone accuracy, the backbone was unfrozen (indicating that all weights encompassing the model were open to tuning) and trained for an additional 50 epochs with an Adam’s learning rate of 1 × 10⁻³, which allowed for a finer convergence of algorithm accuracy (note that model training was stopped once it began to overfit on the training set). The rationale for the frozen-unfrozen training scheme was to finely tune the ImageNet weights to our senescence task without overfitting or nonconvergence. The accuracy of the model along with the AUC of the ROC was used to evaluate model performance

Accuracy = \frac{TP + TN}{TP + TN + FP + FN}

(1)

False positive rate = 1 - \frac{TN}{TN + FP}

(2)

True positive rate = \frac{TP}{TP + FN}

(3)

ROC - Graph of false positive rate vs true positive rate

(4)

TP is true positive, TN is true negative, FP is false positive, and FN is false negative. The final output of the trained model, produced by the SoftMax layer, would provide a senescence score between 0 and 1 and with a higher score indicative of the senescence phenotype. To determine a cutoff on what scores gave a high confidence of senescence, cells with known ground truth were binned by their respective senescence score (0.02 score intervals). The accuracy of classification based on the rounded value of the cells in a bin (i.e., a score of 0.7 rounds to 1, a senescence classification) was calculated. A score of 0.88 was the cutoff for cells to be classified correctly as senescent with at least 90% confidence. This score served as a cutoff for senescent morphological clusters.

Evaluating the effects of aging on senescence experiments

Forty-eight additional primary dermal fibroblast samples of known gender and chronological age were cultured and fixed at low densities using experimental workflows described above. The morphologies of each patient at the baseline were analyzed, and we determined their senescence score and their enrichment within morphologically defined senescence subtypes. Using a univariate linear regression model, we quantified how the abundance of senescent cells and the senescence score correlated with donor age. We selected 18 samples to assess the chronological age–associated response to senescence induction. Samples were either treated with DMSO or DOX, with an 8-day induction period, followed by fixing for morphological analysis. The morphology of cells under both conditions was analyzed by the trained Xception model for score distributions and analyzed by the k-means algorithm to determine senescence subtype distribution. Using this two-pronged orthogonal approach, we quantified changes in the abundance of senescence within each of the three bona fide senescence subtypes with increasing donor age.

Senotherapy response and live-cell imaging

GT22 human dermal fibroblast samples were induced using an 8-day DOX induction protocol and plated at a density of 300 cells per well in a 96-well collagen I–coated glass-bottom plate (Cellvis, P96-1.5H-N). After allowing the cells to adhere overnight, wells were replenished with fresh media containing Spirochrome 555-DNA (Spirochrome, SC201, 1:1000) and Spirochrome 650-FastAct (Spirochrome, SC505, 1:1000) and incubated for 4 hours under physiological conditions in a humidified cell culture incubator. A baseline series of cell images was taken before treatment with serotherapeutic drugs. Following the acquisition of the baseline images, as a one-time bolus treatment while plates were on the microscope, wells were supplemented with either dasatinib (SelleckChem, S1021, 1 μM) and quercetin (SelleckChem, S2391, 10 μM), metformin (SelleckChem, S1950, 5 μM), navitoclax (SelleckChem, S1001, 5 μM), fisetin (SelleckChem, S2298, 25 μM), or ARV-825 (SelleckChem, S8297, 10 nM). Images were recorded once per hour for a span of 60 hours to capture the response dynamics for each drug. Similar to the fixed images, cells in each frame were segmented using an optimized CellProfiler workflow, and the trackpy Python package was used to ascribe cell identities across frames of the same imaged well (https://soft-matter.github.io/trackpy/v0.6.2/). Segmented images were analyzed using the same UMAP and k-means analysis described above to determine the morphological clusters of individual cells as a function of time.

D + Q single-cell responses and viability

D + Q treatment was used as the primary focus for single-cell viability as it was the most effective senotherapy based on cell killing from our preliminary senotherapy screen. To identify dead or dying cells based on the time-lapse videos, two criteria were monitored over a cell’s morphological progression: (i) the point at which a cell has lost tracking; (ii) in the frame before the cell lost tracking, it must have been in either cluster 1, 2, or 3 (indicative of a senescent/dying phenotype). Although clusters 1, 2, and 3 were identified as nonsenescent clusters at the baseline, cells belonging to these clusters after senescence induction and senotherapy treatment exhibited enhanced expression of DNA damage as denoted by γ-H2AX (fig. S7D). Cells were classified by their initial morphological subtype using the pretreatment images, and we analyzed them as a function of senotherapy exposure time to assess subtype-specific responses. Given variations in the number of cells in each senescent morphological cluster, we normalized cell counts for each subtype. To identify how cell morphology transitions as a function of treatment, a hierarchal clustering algorithm was fed a dataset of all senescent morphological subtype trajectories [i.e., a cell with a baseline starting in one of the senescent morphological subtypes (C7, C10, C11) for both untreated and D + Q treatment conditions]. Hierarchical clustering used the ward algorithm to differentiate four response patterns: short death, long death, stable viable, and unstable viable.

Statistics and reproducibility

To establish statistical comparisons between multiple groups, we performed one-way analysis of variance (ANOVA) and post hoc pairwise Tukey tests. Correlations of linear regression fits were evaluated using R and P values. Data were log transformed, where applicable, to normalize distributions for Gaussian-based models. No data were excluded from this work. The experiments were not randomized, and the authors were not blinded to the study. All experiments were performed with at least three biological replicates with at least two in-plate technical replicates.

Software

All single-cell segmentations were performed using optimized workflows in CellProfiler. All dynamic tracking data were obtained using a custom algorithm combining CellProfiler outputs with the TrackPy algorithm. All analyses were performed in Python with the following software specifications: python 3.8.12, tensorflow-gpu 2.5.0, scipy 1.7.1, scikit-learn 1.1.1, pandas 1.4.3, matplotlib 3.4.2, seaborn 0.12.0, plotly 5.10.0, trackpy 0.5.0, umap-learn 0.5.3, and numpy 1.21.2. The source code for SenSCOUT can be found at https://datadryad.org/dataset/doi:10.5061/dryad.rbnzs7hp8 and on the Phillip Lab GitHub page (https://github.com/PhillipLabJHU/SenSCOUT).

Acknowledgments

Funding: We acknowledge the financial support for this work from American Federation for Aging Research (AFAR)—Glenn Foundation Junior Faculty Award (to J.M.P.); The Johns Hopkins University Older Americans Independence Center (OAIC) under the National Institute on Aging award number P30-AG021334 (to J.W. and J.M.P.); Johns Hopkins University Catalyst Award (to J.M.P.); start-up funds from the Biomedical Engineering Department and the Whiting School of Engineering at Johns Hopkins University (to J.M.P.); The National Institute of Health T32 Training Grant 5T32-CA153952-12 (to P.K.); The National Institutes of Health 1UG3-CA275681-01 (to P.-H.W. and J.M.P.); National Institute of General Medical Sciences under Award Numbers R35-GM142889 (to J.F.) and R35-GM157099 (to J.M.P.); and the 2024 Salisbury Family and Center for Innovative Medicine (CIM) Human Aging Project (HAP) award (to J.M.P.).

Author contributions: Conceptualization: P.K., N.M., and J.M.P.; investigation: P.K., N.M., A.A., Y.L., N.Mi., A.W., L.P., T.S., and J.M.P.; methodology: P.K., N.M., Y.L., T.S., P.-H.W., and J.M.P.; resources: J.W., P.-H.W., and J.M.P.; validation: P.K., N.M., C.M., N.Mi., A.W., and J.M.P.; visualization: P.K., N.M., B.S., and J.M.P.; data curation: P.K., N.M., and J.M.P.; formal analysis: P.K., N.M., N.Mi., A.W., and P.-H.W.; software: P.K., N.M., J.F., P.-H.W., C.M., and J.M.P.; supervision: J.F., J.W., and J.M.P.; project administration: J.M.P.; writing—original draft: P.K., N.M., and J.M.P.; writing—review and editing: P.K., N.M., A.A., Y.L., B.S., L.P., J.F., and J.M.P.; funding acquisition: J.M.P.

Competing interests: P.K., N.M., and J.M.P. are co-inventors on a provisional patent application submitted to 10 October 2024 held by Johns Hopkins University related to this work. The other authors declare that they have no competing interests.

Data and materials availability: All data needed to evaluate the conclusions in the paper are present in the paper and/or the Supplementary Materials. The source code for SenSCOUT can be found at https://datadryad.org/dataset/doi:10.5061/dryad.rbnzs7hp8 and on the Phillip Lab GitHub page (https://github.com/PhillipLabJHU/SenSCOUT).

Supplementary Materials

The PDF file includes:

Figs. S1 to S8

Table S1

Legend for data S1

sciadv.ads1875_sm.pdf^{(12.1MB, pdf)}

Other Supplementary Material for this manuscript includes the following:

Data S1

sciadv.ads1875_data_s1.zip^{(20.2KB, zip)}

REFERENCES AND NOTES

1.Gorgoulis V., Adams P. D., Alimonti A., Bennett D. C., Bischof O., Bishop C., Campisi J., Collado M., Evangelou K., Ferbeyre G., Gil J., Hara E., Krizhanovsky V., Jurk D., Maier A. B., Narita M., Niedernhofer L., Passos J. F., Robbins P. D., Schmitt C. A., Sedivy J., Vougas K., von Zglinicki T., Zhou D., Serrano M., Demaria M., Cellular senescence: Defining a path forward. Cell 179, 813–827 (2019). [DOI] [PubMed] [Google Scholar]
2.Hernandez-Segura A., Nehme J., Demaria M., Hallmarks of cellular senescence. Trends Cell Biol. 28, 436–453 (2018). [DOI] [PubMed] [Google Scholar]
3.Wiley C. D., Campisi J., The metabolic roots of senescence: Mechanisms and opportunities for intervention. Nat. Metab. 3, 1290–1301 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
4.Baker D. J., Wijshake T., Tchkonia T., Lebrasseur N. K., Childs B. G., Van De Sluis B., Kirkland J. L., Van Deursen J. M., Clearance of p16 Ink4a-positive senescent cells delays ageing-associated disorders. Nature 479, 232–236 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
5.Di Micco R., Krizhanovsky V., Baker D., d’Adda di Fagagna F., Cellular senescence in ageing: From mechanisms to therapeutic opportunities. Nat. Rev. Mol. Cell Biol. 22, 75–95 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
6.Chu X., Wen J., Raju R. P., Rapid senescence-like response after acute injury. Aging Cell 19, e13201 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
7.Lozano-Torres B., Estepa-Fernández A., Rovira M., Orzáez M., Serrano M., Martínez-Máñez R., Sancenón F., The chemistry of senescence. Nat. Rev. Chem. 3, 426–441 (2019). [Google Scholar]
8.Basisty N., Kale A., Jeon O., Kuehnemann C., Payne T., Rao C., Holtz A., Shah S., Ferrucci L., Campisi J., Schilling B., A proteomic atlas of senescence-associated secretomes for aging biomarker development. SSRN Electron. J. 18, e3000599 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
9.Kusumoto D., Seki T., Sawada H., Kunitomi A., Katsuki T., Kimura M., Ito S., Komuro J., Hashimoto H., Fukuda K., Yuasa S., Anti-senescent drug screening by deep learning-based morphology senescence scoring. Nat. Commun. 12, 257 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
10.Heckenbach I., Mkrtchyan G. V., Ben Ezra M., Bakula D., Madsen J. S., Nielsen M. H., Oró D., Osborne B., Covarrubias A. J., Idda M. L., Gorospe M., Mortensen L., Verdin E., Westendorp R., Scheibye-Knudsen M., Nuclear morphology is a deep learning biomarker of cellular senescence. Nat. Aging 2, 742–755 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
11.Duran I., Pombo J., Sun B., Gallage S., Kudo H., Mchugh D., Bousset L., Efren J., Avila B., Forlano R., Manousou P., Heikenwalder M., Withers D. J., Vernia S., Goldin R. D., Gil J., Detection of senescence using machine learning algorithms based on nuclear features. Nat. Commun. 15, 1041 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
12.Neurohr G. E., Terry R. L., Lengefeld J., Bonney M., Brittingham G. P., Moretto F., Miettinen T. P., Vaites L. P., Soares L. M., Paulo J. A., Harper J. W., Buratowski S., Manalis S., van Werven F. J., Holt L. J., Amon A., Excessive cell growth causes cytoplasm dilution and contributes to senescence. Cell 176, 1083–1097.e18 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
13.Welter E. M., Benavides S., Archer T. K., Kosyk O., Zannas A. S., Machine learning-based morphological quantification of replicative senescence in human fibroblasts. Geroscience 46, 2425–2439 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
14.Saul D., Kosinsky R. L., Atkinson E. J., Doolittle M. L., Zhang X., LeBrasseur N. K., Pignolo R. J., Robbins P. D., Niedernhofer L. J., Ikeno Y., Jurk D., Passos J. F., Hickson L. T. J., Xue A., Monroe D. G., Tchkonia T., Kirkland J. L., Farr J. N., Khosla S., A new gene set identifies senescent cells and predicts senescence-associated pathways across tissues. Nat. Commun. 13, 4827 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
15.Cherry C., Andorko J. I., Krishnan K., Mejías J. C., Nguyen H. H., Stivers K. B., Gray-Gaillard E. F., Ruta A., Han J., Hamada N., Hamada M., Sturmlechner I., Trewartha S., Michel J. H., Davenport Huyer L., Wolf M. T., Tam A. J., Peña A. N., Keerthivasan S., Le Saux C. J., Fertig E. J., Baker D. J., Housseau F., van Deursen J. M., Pardoll D. M., Elisseeff J. H., Transfer learning in a biomaterial fibrosis model identifies in vivo senescence heterogeneity and contributions to vascularization and matrix production across species and diverse pathologies. Geroscience 45, 2559–2587 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
16.Roy A. L., Sierra F., Howcroft K., Singer D. S., Sharpless N., Hodes R. J., Wilder E. L., Anderson J. M., A blueprint for characterizing senescence. Cell 183, 1143–1146 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
17.Kowald A., Passos J. F., Kirkwood T. B. L., On the evolution of cellular senescence. Aging Cell 19, e13270 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
18.Power H., Valtchev P., Dehghani F., Schindeler A., Strategies for senolytic drug discovery. Aging Cell 22, e13948 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
19.Rad A. N., Grillari J., Current senolytics: Mode of action, efficacy and limitations, and their future. Mech. Ageing Dev. 217, 111888 (2024). [DOI] [PubMed] [Google Scholar]
20.Huang W., Hickson L. J., Eirin A., Kirkland J. L., Lerman L. O., Cellular senescence: The good, the bad and the unknown. Nat. Rev. Nephrol. 18, 611–627 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
21.Taherian Fard A., Leeson H. C., Aguado J., Pietrogrande G., Power D., Gómez-Inclán C., Zheng H., Nelson C. B., Soheilmoghaddam F., Glass N., Dharmaratne M., Watson E. R., Lu J., Martin S., Pickett H. A., Cooper-White J., Wolvetang E. J., Mar J. C., Deconstructing heterogeneity of replicative senescence in human mesenchymal stem cells at single cell resolution. Geroscience 46, 999–1015 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
22.Cardoso T., Jose B., Vargas E., Oliveira A., Fraga L. R., Filippi-Chiela E., Subcellular structure, heterogeneity, and plasticity of senescent cells. Aging Cell 23, e14154 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
23.Kirschner K., Rattanavirotkul N., Quince M. F., Chandra T., Functional heterogeneity in senescence. Biochem. Soc. Trans. 48, 765–773 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
24.Wang B., Kohli J., Demaria M., Senescent cells in cancer therapy: Friends or foes? Trends Cancer 6, 838–857 (2020). [DOI] [PubMed] [Google Scholar]
25.de Magalhães J. P., Cellular senescence in normal physiology. Science 384, 1300–1301 (2024). [DOI] [PubMed] [Google Scholar]
26.Tao W., Yu Z., Han J. D. J., Single-cell senescence identification reveals senescence heterogeneity, trajectory, and modulators. Cell Metab. 36, 1126–1143.e5 (2024). [DOI] [PubMed] [Google Scholar]
27.Solé-Boldo L., Raddatz G., Schütz S., Mallm J. P., Rippe K., Lonsdorf A. S., Rodríguez-Paredes M., Lyko F., Single-cell transcriptomes of the human skin reveal age-related loss of fibroblast priming. Commun. Biol. 3, 188 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
28.Wechter N., Rossi M., Anerillas C., Tsitsipatis D., Piao Y., Fan J., Martindale J. L., De S., Mazan-Mamczarz K., Gorospe M., Single-cell transcriptomic analysis uncovers diverse and dynamic senescent cell populations. Aging 15, 2824–2851 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
29.F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2017), pp. 1800–1807. [Google Scholar]
30.Hernandez-Segura A., Brandenburg S., Demaria M., Induction and validation of cellular senescence in primary human cells. J. Vis. Exp. 2018, 57782 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
31.Palikyras S., Sofiadis K., Stavropoulou A., Danieli-mackay A., Hörl D., Nasiscionyte S., Zhu Y., Josipovic N., Rapid and synchronous chemical induction of replicative-like senescence via a small molecule inhibitor. Aging Cell 23, e14083 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
32.Nacarelli T., Liu P., Zhang R., Epigenetic basis of cellular senescence and its implications in aging. Genes (Basel) 8, 343 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
33.Brandl A., Meyer M., Bechmann V., Nerlich M., Angele P., Oxidative stress induces senescence in human mesenchymal stem cells. Exp. Cell Res. 317, 1541–1547 (2011). [DOI] [PubMed] [Google Scholar]
34.Xu M., Su X., Xiao X., Yu H., Li X., Keating A., Wang S., Zhao R. C., Hydrogen peroxide-induced senescence reduces the wound healing-promoting effects of mesenchymal stem cell-derived exosomes partially via miR-146a. Aging Dis. 12, 102–115 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
35.Davalos A. R., Kawahara M., Malhotra G. K., Schaum N., Huang J., Ved U., Beausejour C. M., Coppe J. P., Rodoer F., Campisi J., p53-dependent release of Alarmin HMGB1 is a central mediator of senescent phenotypes. J. Cell Biol. 201, 613–629 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
36.Freund A., Laberge R. M., Demaria M., Campisi J., Lamin B1 loss is a senescence-associated biomarker. Mol. Biol. Cell 23, 2066–2075 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]
37.Carpenter A. E., Jones T. R., Lamprecht M. R., Clarke C., Kang I. H., Friman O., Guertin D. A., Chang J. H., Lindquist R. A., Moffat J., Golland P., Sabatini D. M., CellProfiler: Image analysis software for identifying and quantifying cell phenotypes. Genome Biol. 7, R100 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]
38.Victorelli S., Salmonowicz H., Chapman J., Martini H., Vizioli M. G., Riley J. S., Cloix C., Hall-Younger E., Machado Espindola-Netto J., Jurk D., Lagnado A. B., Sales Gomez L., Farr J. N., Saul D., Reed R., Kelly G., Eppard M., Greaves L. C., Dou Z., Pirius N., Szczepanowska K., Porritt R. A., Huang H., Huang T. Y., Mann D. A., Masuda C. A., Khosla S., Dai H., Kaufmann S. H., Zacharioudakis E., Gavathiotis E., LeBrasseur N. K., Lei X., Sainz A. G., Korolchuk V. I., Adams P. D., Shadel G. S., Tait S. W. G., Passos J. F., Apoptotic stress causes mtDNA release during senescence and drives the SASP. Nature 622, 627–636 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
39.Wu P. H., Phillip J. M., Khatau S. B., Chen W. C., Stirman J., Rosseel S., Tschudi K., Van Patten J., Wong M., Gupta S., Baras A. S., Leek J. T., Maitra A., Wirtz D., Evolution of cellular morpho-phenotypes in cancer metastasis. Sci. Rep. 5, 18437 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
40.Noren Hooten N., Evans M. K., Techniques to induce and quantify cellular senescence. J. Vis. Exp. 2017, 55533 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
41.Beretta L., Santaniello A., Nearest neighbor imputation algorithms: A critical evaluation. BMC Med. Inform. Decis. Mak. 16, 74 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
42.Cubillos M., Wøhlk S., Wulff J. N., A bi-objective k-nearest-neighbors-based imputation method for multilevel data. Expert Syst. Appl. 204, 117298 (2022). [Google Scholar]
43.López-Otín C., Blasco M. A., Partridge L., Serrano M., Kroemer G., Hallmarks of aging: An expanding universe. Cell 186, 243–278 (2023). [DOI] [PubMed] [Google Scholar]
44.Phillip J. M., Aifuwa I., Walston J., Wirtz D., The mechanobiology of aging. Annu. Rev. Biomed. Eng. 17, 113–141 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]
45.Van Deursen J. M., The role of senescent cells in ageing. Nature 509, 439–446 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
46.Gasek N. S., Kuchel G. A., Kirkland J. L., Xu M., Strategies for targeting senescent cells in human disease. Nat. Aging 1, 870–879 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
47.Heinemann T., Kornauth C., Severin Y., Vladimer G. I., Pemovska T., Hadzijusufovic E., Agis H., Krauth M. T., Sperr W. R., Valent P., Jäger U., Simonitsch-Klupp I., Superti-Furga G., Staber P. B., Snijder B., Deep morphology learning enhances ex vivo drug profiling-based precision medicine. Blood Cancer Discov. 3, 502–515 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
48.Xu M., Pirtskhalava T., Farr J. N., Weigand B. M., Palmer A. K., Weivoda M. M., Inman C. L., Ogrodnik M. B., Hachfeld C. M., Fraser D. G., Onken J. L., Johnson K. O., Verzosa G. C., Langhi L. G. P., Weigl M., Giorgadze N., LeBrasseur N. K., Miller J. D., Jurk D., Singh R. J., Allison D. B., Ejima K., Hubbard G. B., Ikeno Y., Cubro H., Garovic V. D., Hou X., Weroha S. J., Robbins P. D., Niedernhofer L. J., Khosla S., Tchkonia T., Kirkland J. L., Senolytics improve physical function and increase lifespan in old age. Nat. Med. 24, 1246–1256 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
49.Cohn R. L., Gasek N. S., Kuchel G. A., Xu M., Xu M., The heterogeneity of cellular senescence: Insights at the single-cell level. Trends Cell Biol. 33, 9–17 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
50.Phillip J. M., Wu P. H., Gilkes D. M., Williams W., McGovern S., Daya J., Chen J., Aifuwa I., Lee J. S. H., Fan R., Walston J., Wirtz D., Biophysical and biomolecular determination of cellular age in humans. Nat. Biomed. Eng. 1, 0093 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
51.Bray M. A., Singh S., Han H., Davis C. T., Borgeson B., Hartland C., Kost-Alimova M., Gustafsdottir S. M., Gibson C. C., Carpenter A. E., Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nat. Protoc. 11, 1757–1774 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]
52.Tavakol M., Wetzel A., Factor analysis: A means for theory and instrument development in support of construct validity. Int. J. Med. Educ. 11, 245–247 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
53.Adamer M. F., Brü Ningk S. C., Tejada-Arranz A., Estermann F., Basler M., Borgwardt K., reComBat: Batch-effect removal in large-scale multi-source gene-expression data integration. Bioinform. Adv. 2, vbac071 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
54.L. McInnes, J. Healy, J. Melville, UMAP: Uniform manifold approximation and projection for dimension reduction. arXiv:1802.03426 [stat.ML] (2018).

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Figs. S1 to S8

Table S1

Legend for data S1

sciadv.ads1875_sm.pdf^{(12.1MB, pdf)}

Data S1

sciadv.ads1875_data_s1.zip^{(20.2KB, zip)}

[R1] 1.Gorgoulis V., Adams P. D., Alimonti A., Bennett D. C., Bischof O., Bishop C., Campisi J., Collado M., Evangelou K., Ferbeyre G., Gil J., Hara E., Krizhanovsky V., Jurk D., Maier A. B., Narita M., Niedernhofer L., Passos J. F., Robbins P. D., Schmitt C. A., Sedivy J., Vougas K., von Zglinicki T., Zhou D., Serrano M., Demaria M., Cellular senescence: Defining a path forward. Cell 179, 813–827 (2019). [DOI] [PubMed] [Google Scholar]

[R2] 2.Hernandez-Segura A., Nehme J., Demaria M., Hallmarks of cellular senescence. Trends Cell Biol. 28, 436–453 (2018). [DOI] [PubMed] [Google Scholar]

[R3] 3.Wiley C. D., Campisi J., The metabolic roots of senescence: Mechanisms and opportunities for intervention. Nat. Metab. 3, 1290–1301 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R4] 4.Baker D. J., Wijshake T., Tchkonia T., Lebrasseur N. K., Childs B. G., Van De Sluis B., Kirkland J. L., Van Deursen J. M., Clearance of p16 Ink4a-positive senescent cells delays ageing-associated disorders. Nature 479, 232–236 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R5] 5.Di Micco R., Krizhanovsky V., Baker D., d’Adda di Fagagna F., Cellular senescence in ageing: From mechanisms to therapeutic opportunities. Nat. Rev. Mol. Cell Biol. 22, 75–95 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R6] 6.Chu X., Wen J., Raju R. P., Rapid senescence-like response after acute injury. Aging Cell 19, e13201 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R7] 7.Lozano-Torres B., Estepa-Fernández A., Rovira M., Orzáez M., Serrano M., Martínez-Máñez R., Sancenón F., The chemistry of senescence. Nat. Rev. Chem. 3, 426–441 (2019). [Google Scholar]

[R8] 8.Basisty N., Kale A., Jeon O., Kuehnemann C., Payne T., Rao C., Holtz A., Shah S., Ferrucci L., Campisi J., Schilling B., A proteomic atlas of senescence-associated secretomes for aging biomarker development. SSRN Electron. J. 18, e3000599 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R9] 9.Kusumoto D., Seki T., Sawada H., Kunitomi A., Katsuki T., Kimura M., Ito S., Komuro J., Hashimoto H., Fukuda K., Yuasa S., Anti-senescent drug screening by deep learning-based morphology senescence scoring. Nat. Commun. 12, 257 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R10] 10.Heckenbach I., Mkrtchyan G. V., Ben Ezra M., Bakula D., Madsen J. S., Nielsen M. H., Oró D., Osborne B., Covarrubias A. J., Idda M. L., Gorospe M., Mortensen L., Verdin E., Westendorp R., Scheibye-Knudsen M., Nuclear morphology is a deep learning biomarker of cellular senescence. Nat. Aging 2, 742–755 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R11] 11.Duran I., Pombo J., Sun B., Gallage S., Kudo H., Mchugh D., Bousset L., Efren J., Avila B., Forlano R., Manousou P., Heikenwalder M., Withers D. J., Vernia S., Goldin R. D., Gil J., Detection of senescence using machine learning algorithms based on nuclear features. Nat. Commun. 15, 1041 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R12] 12.Neurohr G. E., Terry R. L., Lengefeld J., Bonney M., Brittingham G. P., Moretto F., Miettinen T. P., Vaites L. P., Soares L. M., Paulo J. A., Harper J. W., Buratowski S., Manalis S., van Werven F. J., Holt L. J., Amon A., Excessive cell growth causes cytoplasm dilution and contributes to senescence. Cell 176, 1083–1097.e18 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R13] 13.Welter E. M., Benavides S., Archer T. K., Kosyk O., Zannas A. S., Machine learning-based morphological quantification of replicative senescence in human fibroblasts. Geroscience 46, 2425–2439 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R14] 14.Saul D., Kosinsky R. L., Atkinson E. J., Doolittle M. L., Zhang X., LeBrasseur N. K., Pignolo R. J., Robbins P. D., Niedernhofer L. J., Ikeno Y., Jurk D., Passos J. F., Hickson L. T. J., Xue A., Monroe D. G., Tchkonia T., Kirkland J. L., Farr J. N., Khosla S., A new gene set identifies senescent cells and predicts senescence-associated pathways across tissues. Nat. Commun. 13, 4827 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R15] 15.Cherry C., Andorko J. I., Krishnan K., Mejías J. C., Nguyen H. H., Stivers K. B., Gray-Gaillard E. F., Ruta A., Han J., Hamada N., Hamada M., Sturmlechner I., Trewartha S., Michel J. H., Davenport Huyer L., Wolf M. T., Tam A. J., Peña A. N., Keerthivasan S., Le Saux C. J., Fertig E. J., Baker D. J., Housseau F., van Deursen J. M., Pardoll D. M., Elisseeff J. H., Transfer learning in a biomaterial fibrosis model identifies in vivo senescence heterogeneity and contributions to vascularization and matrix production across species and diverse pathologies. Geroscience 45, 2559–2587 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R16] 16.Roy A. L., Sierra F., Howcroft K., Singer D. S., Sharpless N., Hodes R. J., Wilder E. L., Anderson J. M., A blueprint for characterizing senescence. Cell 183, 1143–1146 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R17] 17.Kowald A., Passos J. F., Kirkwood T. B. L., On the evolution of cellular senescence. Aging Cell 19, e13270 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R18] 18.Power H., Valtchev P., Dehghani F., Schindeler A., Strategies for senolytic drug discovery. Aging Cell 22, e13948 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R19] 19.Rad A. N., Grillari J., Current senolytics: Mode of action, efficacy and limitations, and their future. Mech. Ageing Dev. 217, 111888 (2024). [DOI] [PubMed] [Google Scholar]

[R20] 20.Huang W., Hickson L. J., Eirin A., Kirkland J. L., Lerman L. O., Cellular senescence: The good, the bad and the unknown. Nat. Rev. Nephrol. 18, 611–627 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R21] 21.Taherian Fard A., Leeson H. C., Aguado J., Pietrogrande G., Power D., Gómez-Inclán C., Zheng H., Nelson C. B., Soheilmoghaddam F., Glass N., Dharmaratne M., Watson E. R., Lu J., Martin S., Pickett H. A., Cooper-White J., Wolvetang E. J., Mar J. C., Deconstructing heterogeneity of replicative senescence in human mesenchymal stem cells at single cell resolution. Geroscience 46, 999–1015 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R22] 22.Cardoso T., Jose B., Vargas E., Oliveira A., Fraga L. R., Filippi-Chiela E., Subcellular structure, heterogeneity, and plasticity of senescent cells. Aging Cell 23, e14154 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R23] 23.Kirschner K., Rattanavirotkul N., Quince M. F., Chandra T., Functional heterogeneity in senescence. Biochem. Soc. Trans. 48, 765–773 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R24] 24.Wang B., Kohli J., Demaria M., Senescent cells in cancer therapy: Friends or foes? Trends Cancer 6, 838–857 (2020). [DOI] [PubMed] [Google Scholar]

[R25] 25.de Magalhães J. P., Cellular senescence in normal physiology. Science 384, 1300–1301 (2024). [DOI] [PubMed] [Google Scholar]

[R26] 26.Tao W., Yu Z., Han J. D. J., Single-cell senescence identification reveals senescence heterogeneity, trajectory, and modulators. Cell Metab. 36, 1126–1143.e5 (2024). [DOI] [PubMed] [Google Scholar]

[R27] 27.Solé-Boldo L., Raddatz G., Schütz S., Mallm J. P., Rippe K., Lonsdorf A. S., Rodríguez-Paredes M., Lyko F., Single-cell transcriptomes of the human skin reveal age-related loss of fibroblast priming. Commun. Biol. 3, 188 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R28] 28.Wechter N., Rossi M., Anerillas C., Tsitsipatis D., Piao Y., Fan J., Martindale J. L., De S., Mazan-Mamczarz K., Gorospe M., Single-cell transcriptomic analysis uncovers diverse and dynamic senescent cell populations. Aging 15, 2824–2851 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R29] 29.F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (IEEE, 2017), pp. 1800–1807. [Google Scholar]

[R30] 30.Hernandez-Segura A., Brandenburg S., Demaria M., Induction and validation of cellular senescence in primary human cells. J. Vis. Exp. 2018, 57782 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R31] 31.Palikyras S., Sofiadis K., Stavropoulou A., Danieli-mackay A., Hörl D., Nasiscionyte S., Zhu Y., Josipovic N., Rapid and synchronous chemical induction of replicative-like senescence via a small molecule inhibitor. Aging Cell 23, e14083 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R32] 32.Nacarelli T., Liu P., Zhang R., Epigenetic basis of cellular senescence and its implications in aging. Genes (Basel) 8, 343 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R33] 33.Brandl A., Meyer M., Bechmann V., Nerlich M., Angele P., Oxidative stress induces senescence in human mesenchymal stem cells. Exp. Cell Res. 317, 1541–1547 (2011). [DOI] [PubMed] [Google Scholar]

[R34] 34.Xu M., Su X., Xiao X., Yu H., Li X., Keating A., Wang S., Zhao R. C., Hydrogen peroxide-induced senescence reduces the wound healing-promoting effects of mesenchymal stem cell-derived exosomes partially via miR-146a. Aging Dis. 12, 102–115 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R35] 35.Davalos A. R., Kawahara M., Malhotra G. K., Schaum N., Huang J., Ved U., Beausejour C. M., Coppe J. P., Rodoer F., Campisi J., p53-dependent release of Alarmin HMGB1 is a central mediator of senescent phenotypes. J. Cell Biol. 201, 613–629 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R36] 36.Freund A., Laberge R. M., Demaria M., Campisi J., Lamin B1 loss is a senescence-associated biomarker. Mol. Biol. Cell 23, 2066–2075 (2012). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R37] 37.Carpenter A. E., Jones T. R., Lamprecht M. R., Clarke C., Kang I. H., Friman O., Guertin D. A., Chang J. H., Lindquist R. A., Moffat J., Golland P., Sabatini D. M., CellProfiler: Image analysis software for identifying and quantifying cell phenotypes. Genome Biol. 7, R100 (2006). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R38] 38.Victorelli S., Salmonowicz H., Chapman J., Martini H., Vizioli M. G., Riley J. S., Cloix C., Hall-Younger E., Machado Espindola-Netto J., Jurk D., Lagnado A. B., Sales Gomez L., Farr J. N., Saul D., Reed R., Kelly G., Eppard M., Greaves L. C., Dou Z., Pirius N., Szczepanowska K., Porritt R. A., Huang H., Huang T. Y., Mann D. A., Masuda C. A., Khosla S., Dai H., Kaufmann S. H., Zacharioudakis E., Gavathiotis E., LeBrasseur N. K., Lei X., Sainz A. G., Korolchuk V. I., Adams P. D., Shadel G. S., Tait S. W. G., Passos J. F., Apoptotic stress causes mtDNA release during senescence and drives the SASP. Nature 622, 627–636 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R39] 39.Wu P. H., Phillip J. M., Khatau S. B., Chen W. C., Stirman J., Rosseel S., Tschudi K., Van Patten J., Wong M., Gupta S., Baras A. S., Leek J. T., Maitra A., Wirtz D., Evolution of cellular morpho-phenotypes in cancer metastasis. Sci. Rep. 5, 18437 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R40] 40.Noren Hooten N., Evans M. K., Techniques to induce and quantify cellular senescence. J. Vis. Exp. 2017, 55533 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R41] 41.Beretta L., Santaniello A., Nearest neighbor imputation algorithms: A critical evaluation. BMC Med. Inform. Decis. Mak. 16, 74 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R42] 42.Cubillos M., Wøhlk S., Wulff J. N., A bi-objective k-nearest-neighbors-based imputation method for multilevel data. Expert Syst. Appl. 204, 117298 (2022). [Google Scholar]

[R43] 43.López-Otín C., Blasco M. A., Partridge L., Serrano M., Kroemer G., Hallmarks of aging: An expanding universe. Cell 186, 243–278 (2023). [DOI] [PubMed] [Google Scholar]

[R44] 44.Phillip J. M., Aifuwa I., Walston J., Wirtz D., The mechanobiology of aging. Annu. Rev. Biomed. Eng. 17, 113–141 (2015). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R45] 45.Van Deursen J. M., The role of senescent cells in ageing. Nature 509, 439–446 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R46] 46.Gasek N. S., Kuchel G. A., Kirkland J. L., Xu M., Strategies for targeting senescent cells in human disease. Nat. Aging 1, 870–879 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R47] 47.Heinemann T., Kornauth C., Severin Y., Vladimer G. I., Pemovska T., Hadzijusufovic E., Agis H., Krauth M. T., Sperr W. R., Valent P., Jäger U., Simonitsch-Klupp I., Superti-Furga G., Staber P. B., Snijder B., Deep morphology learning enhances ex vivo drug profiling-based precision medicine. Blood Cancer Discov. 3, 502–515 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R48] 48.Xu M., Pirtskhalava T., Farr J. N., Weigand B. M., Palmer A. K., Weivoda M. M., Inman C. L., Ogrodnik M. B., Hachfeld C. M., Fraser D. G., Onken J. L., Johnson K. O., Verzosa G. C., Langhi L. G. P., Weigl M., Giorgadze N., LeBrasseur N. K., Miller J. D., Jurk D., Singh R. J., Allison D. B., Ejima K., Hubbard G. B., Ikeno Y., Cubro H., Garovic V. D., Hou X., Weroha S. J., Robbins P. D., Niedernhofer L. J., Khosla S., Tchkonia T., Kirkland J. L., Senolytics improve physical function and increase lifespan in old age. Nat. Med. 24, 1246–1256 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R49] 49.Cohn R. L., Gasek N. S., Kuchel G. A., Xu M., Xu M., The heterogeneity of cellular senescence: Insights at the single-cell level. Trends Cell Biol. 33, 9–17 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R50] 50.Phillip J. M., Wu P. H., Gilkes D. M., Williams W., McGovern S., Daya J., Chen J., Aifuwa I., Lee J. S. H., Fan R., Walston J., Wirtz D., Biophysical and biomolecular determination of cellular age in humans. Nat. Biomed. Eng. 1, 0093 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R51] 51.Bray M. A., Singh S., Han H., Davis C. T., Borgeson B., Hartland C., Kost-Alimova M., Gustafsdottir S. M., Gibson C. C., Carpenter A. E., Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nat. Protoc. 11, 1757–1774 (2016). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R52] 52.Tavakol M., Wetzel A., Factor analysis: A means for theory and instrument development in support of construct validity. Int. J. Med. Educ. 11, 245–247 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R53] 53.Adamer M. F., Brü Ningk S. C., Tejada-Arranz A., Estermann F., Basler M., Borgwardt K., reComBat: Batch-effect removal in large-scale multi-source gene-expression data integration. Bioinform. Adv. 2, vbac071 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

[R54] 54.L. McInnes, J. Healy, J. Melville, UMAP: Uniform manifold approximation and projection for dimension reduction. arXiv:1802.03426 [stat.ML] (2018).

PERMALINK

Single-cell morphology encodes functional subtypes of senescence in aging human dermal fibroblasts

Pratik Kamat

Nico Macaluso

Yukang Li

Anshika Agrawal

Aaron Winston

Lauren Pan

Teasia Stewart

Bartholomew Starich

Nicholas Milcik

Chanhong Min

Pei-Hsun Wu

Jeremy Walston

Jean Fan

Jude M Phillip

Roles

Abstract

INTRODUCTION

RESULTS

Multiple induction conditions and biomarkers enable the robust profiling of cellular senescence

Fig. 1. Cellular and nuclear morphologies are biomarkers for in vitro cellular senescence.

High-content imaging enables the profiling of senescence across biological conditions

Single-cell morphology encodes heterogeneous senescence phenotypes in dermal fibroblasts

Fig. 2. Single-cell morphology encodes heterogeneous senescence phenotypes.

Single-cell morphology robustly encodes senescence biomarker expression

Imputing senescence biomarker expressions based on single-cell morphologies

Machine learning strategies reveal morphological senescence subtypes

Fig. 3. Machine learning and computational techniques to identify morphological senescence subtypes.

Emergent senescence phenotypes depend on inducer-specific kinetics

Fig. 4. Cellular senescence is a dynamic phenotype.

Baseline senescence burden predicts susceptibility to senescence induction with DOX

Fig. 5. Baseline morphology more effectively encodes senescence susceptibility than chronological age.

Senescence subtypes exhibit differential responses to D + Q

Fig. 6. Senescence subtypes encode differential responses to senotherapies.

DISCUSSION

MATERIALS AND METHODS

Cell culture

Induction of senescence

Profiling cellular secretions

Immunofluorescence staining

High-content microscopy and quantification of morphological features

Analysis of protein-based biomarkers

UMAP and k-means clustering

Protein biomarker imputation

Kinetics of senescence experiments

Machine learning approach to quantifying senescence subtypes using the Xception algorithm

Evaluating the effects of aging on senescence experiments

Senotherapy response and live-cell imaging

D + Q single-cell responses and viability

Statistics and reproducibility

Software

Acknowledgments

Supplementary Materials

The PDF file includes:

Other Supplementary Material for this manuscript includes the following:

REFERENCES AND NOTES

Associated Data

Supplementary Materials

ACTIONS

PERMALINK

RESOURCES

Similar articles

Cited by other articles

Links to NCBI Databases