High-content live-cell imaging reveals co-regulation of cardiomyocyte decisions of hypertrophy and apoptosis

Cardiomyocyte hypertrophy and apoptosis underlie cardiomyopathies and heart failure. While previous studies have described hypertrophy or apoptosis at the level of cell populations, how individual cells commit to these distinct yet co-regulated fates is unclear. We used high-content imaging to track single-cell hypertrophy and apoptosis dynamics, revealing new features and unique subpopulation responses. Catecholamines isoproterenol and norepinephrine induced heterogeneous analog hypertrophy and digital apoptosis, which support a “grow and/or die” conceptual model for cell decisions. Multinomial log-linear models indicated that a cell’s initial size and DNA content predict its susceptibility to hypertrophy and apoptosis. This work integrates dynamic morphological and biochemical cell profiling to reveal that cardiomyocyte hypertrophic and apoptotic responses to catecholamines represent an incoherent feedforward loop in which hypertrophy enhances survival.

As such, it is unclear whether hypertrophy and apoptosis decisions occur concurrently, 40 sequentially, or in opposition to each other (Kang et al., 2015;van Empel and De Windt, 2004). 41 As the field lacks a conceptual model, we identify three possible conceptual models for joint 42 cardiomyocyte decision-making of hypertrophy and apoptosis. A "grow-and-die" model suggests 43 serial decision-making where cells hypertrophy but then later apoptose due to an inability to 44 handle growth stresses (Sawyer et al., 2002). The importance of caspase activation dynamics as 45 a regulator of apoptosis and the role of sub-apoptotic caspase activity in hypertrophy suggests a 46 distinct "grow-or-die" model, in which caspase activation dynamics regulate whether cardiomyocytes will either survive and hypertrophy or die prior to hypertrophic growth 48 (Putinski et al., 2013;Roux et al., 2015). An alternative to these previously implicated 49 conceptual models is a "grow-and/or-die" model, which suggests that mixtures of different 50 subpopulations of cells undergo both possibilities. In this model, a stimulus may induce a 51 subpopulation of cells to undergo hypertrophy. While this response may be protective against 52 apoptosis, hypertrophic cells may also reach a physical limit of viability and subsequently 53 undergo apoptosis. Fixed timepoint in vivo experiments cannot determine which of these three 54 conceptual models is most likely to be occurring dynamically and at a single-cell level. demonstrated that cardiomyocyte hypertrophy induced by catecholamines requires sub-66 apoptotic levels of caspase activation (Putinski et al., 2013). However, end point or cell lysis 67 assays do not capture initial heterogeneity, which may provide predictive or explanatory factors 68 for different phenotypes. Dynamic time course data is necessary to elucidate the mechanisms by 69 which individual cells decide to grow or die, and the interaction between those decisions (Paek 70 et al., 2016; Roux et al., 2015). To overcome these limitations, here, we designed a live-cell high-71 content imaging approach that monitors the dynamics of growth and death simultaneously at a 72 single-cell level. This approach incorporates multiplexed measurements of multiple phenotypes 73 to understand the relationship between hypertrophy and apoptosis; dynamic time course 74 measurements to elucidate the mechanisms by which individual cells decide to grow or die; the 75 interaction between those decisions; and single-cell measurements of decision-making. 76 Here, we characterize the cellular, nuclear, and biochemical single-cell dynamic hypertrophy 77 and apoptosis responses of cardiomyocytes to multiple perturbations. We identify a biphasic 78 cardiomyocyte concentration response to the catecholamine isoproterenol, resulting from 79 opposing digital apoptosis and analog hypertrophy responses. Classification of cell responses 80 into that of three subpopulations reveals differential concentration-dependent dynamic 81 responses. Finally, a multinomial log-linear data-driven model reveals new characteristics of a 82 cell that predicts its likelihood of committing to hypertrophy or apoptosis, which we then 83 validate. Our work thus develops a common experimental and statistical framework to reveal 84 signaling and phenotypic dynamics that underlie individual cell decision-making. 85 86

Assay design for single-cell phenotyping of hypertrophy-apoptosis dynamics 88
Our live single-cell assay uses multiple simultaneous fluorescent readouts of hypertrophic and 89 apoptotic activity ( Figure 1A). Cells were transduced with eGFP as a live-cell reporter of single-90 cell morphology and protein synthesis. Nuclei were labeled with Hoechst, and general cell death 91 and apoptosis were measured with propidium iodide (PI) and fluorescently-conjugated annexin 92 V (A5). To track temporal dynamics, individual wells in a 96-well microplate representing 93 different treatment conditions were imaged for 48 hours with a 1 hour resolution ( Figure 1B). 94 specific ɑ-actinin expression. We developed automated image analysis methods to measure dynamics of 96 multiple metrics related to cellular morphology, viability, and final ɑ-actinin expression at a single-cell 97 level (Methods). 98 To correct for cell-to-cell eGFP expression variability and uneven fluorescence that persisted 99 after flatfield illumination, we trained a random forest classifier to generate probability maps of 100 eGFP ( Figure 1C). The random forest model was robust to at least a 5-fold range in variations 101 in fluorescence signal-to-noise ratio, allowing optimization of fluorescence acquisition settings 102 to minimize phototoxicity (Supplemental Figure 1). We used this eGFP probability map for 103 thresholding, single-cell segmentation, and tracking for longitudinal morphology and 104 fluorescence measurements (Figure 1D and Methods). As our eGFP reporter was expressed 105 under a ubiquitous CMV promoter and thus could not differentiate between cardiomyocytes and 106 other cell types, we used an α-actinin fixed immunofluorescence probability map and live single-107 cell tracking to identify cardiomyocytes. Similarly, a live A5 probability map was used to 108 determine A5 positivity of each cell. 109 This assay was validated using cells treated with DMSO as a negative vehicle control. Even in the 110 absence of hypertrophic or apoptotic stimuli, cells showed a wide range of both hypertrophic 111 ( Figure 1A and D, orange outline and traces) and apoptotic trajectories (cyan outline and 112 traces). At the cell-population level, we found that A5 binding accumulated earlier than PI 113 uptake, suggesting that most of the death observed was due to apoptosis rather than other death 114 processes ( Figure 1F). This apoptotic lag between initial A5 binding and later PI uptake at the 115 population level was validated in individual A5/PI double-positive cells ( Figure 1G) , with the lag between those two events quantified (red distribution; * p = 0.019, one-sample 143 Wilcoxon signed rank test, n = 733 cells). 144 145

Cellular and nuclear profiling of staurosporine-induced apoptosis dynamics 146
Apoptosis is characterized by morphological and biochemical changes to both cell and nucleus. 147 As a representative cellular perturbation to examine apoptosis dynamics with our high-content 148 imaging assay, we selected stimulation with staurosporine (STS), an ATP-competitive inhibitor 149 with activity against a broad spectrum of kinases and a well-known inducer of apoptosis in  Treatment with STS induced large amounts of cell death marked by increased A5 binding and PI 159 uptake compared to DMSO vehicle-control treatment (Figure 2A and B). Bulk-population 160 dynamics of total accumulation of A5 and PI were again suggestive of apoptosis, with PI-uptake 161 occurring later than that of A5 ( Figure 2C). However, the dynamics of apoptosis were markedly 162 different between STS and DMSO treatments. DMSO treatment induced a biphasic response 163 with initially low levels of A5 and PI binding between 0 and 24 hours and a second phase of 164 increased apoptotic activity after 24 hours. In contrast, STS-treated cells largely completed 165 apoptosis within the first 36 hours of treatment. There was a significant lag from A5 to PI 166 binding in single cells that eventually became A5/PI-double positive (Figure 2D), suggesting 167 apoptosis and not other forms of cell death. 168 Next, we sought to characterize the temporal dynamics of additional cellular features that may 169 be associated with apoptosis ( Figure 2E). As expected, STS treatment induced global increases 170 of cell membrane A5 binding followed by nuclear PI uptake. Motivated by known apoptotic 171 nuclear and chromatin changes (Toné et al., 2007), we also investigated dynamic changes in 172 nuclear Hoechst staining. Mean, total (integrated), and median absolute deviation of the 173 Hoechst intensity all followed similar trajectories, with STS-treated trajectories increasing above 174 the DMSO-treated baseline, reflecting increased membrane permeability to Hoechst and 175 chromatin condensation (Ormerod et al., 1993). After approximately 24 hours, STS-treated 176 trajectories of Hoechst decreased below that of DMSO-treated trajectories, suggesting DNA 177 degradative apoptotic processes. We also investigated changes in nuclear morphology. During 178 the first 36 hours, nuclei of STS-treated cells had smaller nuclei, which later increased in size 179 relative to their DMSO-treated counterparts. Interestingly, nuclear form factor decreased and 180 eccentricity increased in STS-treated cells relative to DMSO-treated cells even at the very early 181 times, indicating more irregular and less circular nuclei consistent with apoptosis. Overall, these 182 analyses delineate a temporal sequence of apoptotic kinetics, starting with changes to nuclear 183 permeability and shape, followed by A5 binding, and finally loss of cell membrane integrity. 184 We previously identified divergent characteristics of cellular hypertrophy induced by different 220 hypertrophic agonists (Bass et al., 2012). While our previous approach segmented on ɑ-actinin 221 immunofluorescence, here we find greater fidelity when further incorporating cell-to-cell variation in 222 eGFP expression (Supplemental Figure 2). We therefore re-characterized the same metrics of 223 cell shape and size and found that treatment with PE led to significant changes to all metrics 224 observed ( Figure 3H). As expected, PE increased raw cell area and major axis length. 225 Elongation (axis ratio), eccentricity, and form factor all decreased, suggesting PE-treated cells 226 became more circular but more spiky ( Figure 3A, red outline).  showing induction of hypertrophy (red outline) and eGFP expression (gray

Isoproterenol induces biphasic dose-dependent hypertrophy and apoptosis 264 dynamics 265
In silico reconstructions of the in vitro cardiomyocyte hypertrophy signaling network have 266 identified the ꞵA signaling pathway as one that is potentially both hypertrophic and apoptotic (Kang et     Therefore we asked whether CaMKII inhibition would modulate single-cell decisions of 300 hypertrophy and apoptosis. Combination treatment with ISO and KN-93, a CaMKII inhibitor, 301 increased apoptosis relative to combination treatment with ISO and KN-92, an inactive 302 analogue of KN-93 (Supplemental Figure 6). 303 As a previous study showed that sub-apoptotic levels of caspase-3 were surprisingly required for showing induction of hypertrophy and A5 binding (magenta We hypothesized that the biphasic population-level responses to ISO may reflect heterogeneity 345 in single-cell decision making between cell growth and death, with potential for functional 346 crosstalk between these cell responses encompassed by three possible conceptual models 347 ( Figure 5A). To facilitate comparison of cell hypertrophy and apoptosis, we first classified cells 348 as "growers" , "stunned" , or "shrinkers". Strikingly, the hypertrophic dynamics of these three 349 subpopulations ( Figure 5B) are qualitatively distinct from the transient population-level 350 hypertrophic to 100 µM ISO ( Figure 4B). Thus, the transient population-level hypertrophy 351 results from the early increases in cell area of the "growers," followed by the later decreased cell 352 area of the "shrinkers," and some cells remaining "stunned". The "shrinker" subpopulation was 353 more prone to cell death and therefore contributed to the decrease in median cell population 354 area observed at later timepoints and higher ISO doses (Figure 4E and H). 355 As expected from the single-cell ISO concentration hypertrophy responses at 24 h, we saw a 356 trend towards an increased proportion of grower cells with ISO stimulation compared to DMSO. 357 In addition, the largest proportion of grower cells was seen with 1 µM ISO, which corresponds to 358 the dose at which we saw the greatest population-level hypertrophy response (Figures 5C and  359   4C,D). The relative decrease in proportion of grower cells as ISO concentration is increased 360 beyond 10 µM also correlated with the plateauing of the population-level hypertrophy ISO 361 concentration response (Figure 4C and D). 362 We hypothesized that growth and death responses may be related at the single cell level, with 363 apoptosis primarily occurring in "shrinker" cells. We quantified the proportions of cells from the 364 three growth subpopulations that contribute to the ultrasensitive response of A5 binding with 365 increasing ISO concentration (Figure 5D). At 100 µM ISO, shrinkers and stunned cells were 366 indeed statistically more likely than growers to become A5 positive, leading us to favor a "grow-367 or-die" model over the "grow-and-die" model. However, in contrast to our hypothesis, the 368 quantitative contributions to death responses were fairly similar ( Figure 5E). These 369 concentration-dependent hypertrophy and apoptosis decisions are summarised in Figure 5F. 370 Having found that apoptotic cells also tended to become smaller (Figure 4H), we further 371 analyzed the growth and death kinetics of the hypertrophic subpopulations to determine 372 additional causes for the biphasic hypertrophy concentration response observed at the 373 population level. We measured the time at which each cell achieved maximal cell area. 374 Furthermore, if the cell became A5 positive, we measured the time lag between those two events 375 ( Figure 5G). We found that the hypertrophic subpopulations differed not only in the 376 magnitude of their hypertrophy, but also the time at which they achieved their maximal 377 hypertrophy. Stunned cells grew to their maximal extent before 24 hours while growers 378 continued growing after 24 hours. As expected, shrinkers had maximal area at the start of the 379 time course (Figure 5H). Even though the three hypertrophic subpopulations differed 380 significantly in their hypertrophic kinetics, the dead cells in the grower and stunned populations 381 bound A5 with statistically different yet fairly similar delay after their maximal cell area. 382 However, the shrinker cells showed a striking difference from other subpopulations, with a 383 bimodal distribution in delay from maximal cell area to A5 binding. Indeed, almost half of the 384 shrinker cells persisted at least 24 hours before binding A5 (Figure 5I). From this analysis, we 385 identified that the source of the biphasic cell population hypertrophy ISO concentration 386 response was a subpopulation of cells that underwent cell shrinkage and then delayed apoptosis, 387 while cell hypertrophy may represent a protective mechanism against apoptosis. The presence of 388 these distinct subpopulation behavior may represent an incoherent feedforward network motif 389 between ꞵA stimulation, hypertrophy, and apoptosis. All data is the same as in Figure 4. as cells whose max fold change area was greater than that of the 75th percentile of the DMSO-399 treated population. Stunned (orange) cells were defined as cells whose max fold change area was 400 in the DMSO-treated interquartile range. Shrinkers (green) were defined as cells whose max fold 401 change area was less than the 25th percentile of the DMSO-treated cells. Solid lines indicate 402 population median, shaded areas indicate interquartile range, black line indicates population 403 median 404 (C) Concentration response of cell population makeup. Errorbars represent mean ± SEM. 405 Comparisons were made between each treatment's percent "grower" subpopulation; while the 406 overall mixed-effects ANOVA was not statistically significant (p=0.094) the more powered one-407 sided Dunnett's test showed statistically significant differences indicated with * p<0.05 vs 408 DMSO control, n = 3-4 biological replications. Moreover, the mixed-effects ANOVA between 409 the 3 hypertrophy subpopulations was not significant for the DMSO-treatment. However, there 410 was a significant increase in the proportion of the grower subpopulation relative to the other two 411 subpopulations with p<0.01, mixed-effects ANOVA followed by Tukey HSD.

412
(D) Concentration response of cell population makeup for all apoptotic cells (same data as 413 Figure 4F).

Initial cell morphology and nucleus predicts cellular decision-making 427
Given the considerable single-cell heterogeneity in response to ISO stimulation, we asked 428 whether pre-existing features of cardiomyocytes would predict their subsequent growth and 429 death. We used a multinomial log-linear regression model to identify cellular features that 430 predict hypertrophy-apoptosis decision-making (Methods, Figure 6A, Supplemental 431 Table 1). The response variable was one of 6 pairwise combinations of the 3 hypertrophic 432 trajectory classifications and the binary survival/death outcome. The predictor variables were 433 the 6 cellular and 6 nuclear size, shape and fluorescence intensity metrics identified by our assay 434 as potentially relevant to cardiomyocyte apoptosis and hypertrophy (Figures 2 and 3). showed that small cells had a hypertrophic growth response while larger cells did not 451 hypertrophy significantly (Figure 6C). While the initial area regression coefficient did not 452 specifically correlate with the survival status of a cell, it had an especially strong statistical 453 association with the shrinker/dier population. Indeed, cells in the largest quartile of initial cell 454 areas were more likely to die than those in the smallest quartile ( Figure 6D). A parallel random 455 forest regression model showed that the initial area was the most influential metric for model 456 accuracy (Supplemental Figure 9C). These initial cell size-dependent hypertrophy and 457 apoptosis decisions are summarised in Figure 6F. Our data suggest that while hypertrophic 458 signaling confers partial protection against apoptosis, ISO stimulation still generates a 459 population of hypertrophic cells that go on to die. Thus, our data best supports the "grow-460 and/or-die" model over the "grow-or-die" model. 461 Interestingly, this analysis identified nuclear Hoechst intensity and its median absolute 462 deviation as two nuclear metrics associated with both cell growth and survival (Figure 6A). The 463 median absolute deviation metric was also implicated as the second most influential metric for 464 model accuracy by the random forest model (Supplemental Figure 9C)  All data is the same as in Figure 4. After validating our cardiomyocyte assay for hypertrophy and apoptosis measurements 509 separately and identifying new features of those processes, we used our assay to examine the 510 dynamics of decision-making between hypertrophy and apoptosis in response to 511 catecholamines. Theoretical mechanisms underlying simultaneous regulation of analog and 512 digital phenotypic decision-making by a single signaling network have been proposed. To our 513 knowledge, however, the relative timing of early analog hypertrophy responses and late digital 514 apoptosis decisions represents the first experimental support of a single signaling network 515 regulating a mixture of analog and digital phenotypic decisions (Kovary et al., 2018). 516 Leveraging these heterogeneity dynamics in our cell populations, we identified a biphasic 517 hypertrophic ISO concentration response at the cell-population level, suggesting at least two 518 underlying, competing processes. We therefore hypothesized that different subpopulations of 519 cells contribute to the overall population response. 520 Our data best supports the "grow-and/or-die" conceptual model in which hypertrophic signaling 521 is moderately protective against apoptosis. Low-to mid-concentrations of ISO increased the 522 proportion of cells that hypertrophied and resisted apoptosis, while high ISO concentrations 523 increased the number of cells susceptible to death. We used cyclic data-driven hypothesis testing 524 and modelling to identify characteristic features of these different subpopulations. Our 525 multinomial log-linear regression models sought to identify nongenetic heterogeneities in initial 526 cell size, morphology, and fluorescence metrics at a single ISO dose that would predict dynamic, 527 single-cell decision-making. This modeling approach had an accuracy rate similar to the more 528 agnostic, black-box random forest model. This suggests that differences in model structure 529 (such as the presence of non-linear interactions between predictors) were less important than 530 other, non-characterized factors or biochemical sources of stochasticity that our high-content 531 imaging assay did not capture (Hilfinger and Paulsson, 2011;Swain et al., 2002). Even so, we 532 identified and validated the initial cell size as a predictor of hypertrophic trajectory and 533 susceptibility to death. Surprisingly, we identified two nuclear metrics, DNA content and 534 chromatin homogeneity, that predicted cell hypertrophy and survival, which suggest a potential 535 role for transcriptional regulation of differential cell decision-making (Liu and Tang, 2019; 536 Symmons and Raj, 2016). These three metrics were also predictive of hypertrophy/apoptosis 537 decision-making with other treatments, namely PE and other ISO concentrations. 538 We envision multiple avenues for extending the approach presented herein. Tracking in vitro 539 co-culture models of cardiomyocytes and fibroblasts and the relative local density of those cells 540 would provide new insight into paracrine signaling interactions between cardiomyocytes and 541 fibroblasts to yield new potential therapeutic targets (Piek et al., 2016). We have also previously 542 described a screen for cardiomyocyte proliferation using methods compatible with those 543 presented here, suggesting a potential third phenotype to be multiplexed with our assay (Woo et  Cells were transduced with 100μL of eGFP lentivirus and plating media supplemented with 580 8μg/mL polybrene (Santa Cruz Biotechnology, Cat#sc-134220) at the time of plating and after 581 24 hours. 36 hours after plating, the medium was replaced with low-serum media (phenol-red 582 free Dulbecco's modified Eagle media supplemented with 17% M199, 0.1% FBS, 1% ITSS, 20mM 583 HEPES, 50 U/mL penicillin, 50 μg/mL streptomycin, 4.8mM L-glutamine) for 24 hours of 584 serum starvation prior to treatment. 585

Dual hypertrophy/apoptosis assay 586
Cells were pre-stained and pre-treated with relevant small molecule inhibitors in low-serum 587 media supplemented with 27nM Hoechst 33342, 0.5nM Propidium iodide, and 0.5× Alexa Fluor 588 647-conjugated A5 (Invitrogen) for 2 hours. Cells were imaged on the Operetta CLS high-589 content imaging system (Perkin-Elmer) using a 10 × 0.3NA objective and the temperature and 590 CO2 control option set to 37°C and 5% CO2, respectively. Images were captured every 1 hour for 591 48 hours. 4 fields of view were acquired for each well. 592 After live-cell imaging, the cells were fixed with 4% PFA and stained with DAPI and monoclonal 593 anti-α-Actinin antibody (Sigma, Cat#A7811). The same fields of view acquired during live 594 imaging were again imaged.  Hoechst-and propidium iodide-positive nuclei were independently segmented using Otsu 605 thresholding. Binucleates were identified by merging touching nuclei and then shrunken by a 606 radius of 1 pixel. GFP positive cells were identified using a watershed segmentation algorithm on 607 the GFP probability map using the merged nuclei as starting seeds. These GFP positive cells 608 were tracked and measured for morphological and fluorescence intensity characteristics 609 including area, GFP intensity, A5 intensity and probability, final α-Actinin intensity and 610 probability. Nuclei were also measured for morphological and propidium iodide and Hoechst 611 intensities. 612

Statistical Analysis 613
All statistical analyses were done in R version 3.6.3 (R Core Team, 2020). The following 614 packages were additionally used for general data preparation and visualization: tidyverse 615  Modelling 620 Data were normalized prior to statistical modelling. For multinomial log-linear modelling, 621 predictor variables were filtered for multicollinearity on the basis of their variance inflation 622 factor (VIF) scores -removal of the raw major axis length metric reduced all VIF scores below a 623 maximum of 2.4. We also verified the remaining collinearities with a tableplot and that those 624 variables described sufficiently different metrics (Supplemental Figure 9A). For random 625 forest modelling, default values from the randomForest package were used with the exception of 626 1000 trees (Liaw and Wiener, 2002). 627