Dynamic map illuminates Hippo-cMyc module crosstalk driving cardiomyocyte proliferation

ABSTRACT

Numerous regulators of cardiomyocyte (CM) proliferation have been identified, yet how they coordinate during cardiac development or regeneration is poorly understood. Here, we developed a computational model of the CM proliferation regulatory network to obtain key regulators and systems-level understanding. The model defines five modules (DNA replication, mitosis, cytokinesis, growth factor, Hippo pathway) and integrates them into a network of 72 nodes and 88 reactions that correctly predicts 74 of 81 (91.35%) independent experiments from the literature. The model predicts that in response to YAP activation, the Hippo module crosstalks to the growth factor module via PI3K and cMyc to drive cell cycle activity. This predicted YAP-cMyc axis is validated experimentally in rat CMs and further supported by YAP-stimulated cMyc open chromatin and mRNA in mouse hearts. This validated computational model predicts how individual regulators and modules coordinate to control CM proliferation.

Summary: A computational model of cardiomyocyte proliferation is developed and validated against experiments from the literature. A predicted YAP-cMyc axis is validated in data from cells and mice.

INTRODUCTION

Cardiovascular diseases are the leading causes of death worldwide and are the highest cause of morbidity and mortality (https://www.cdc.gov/heart-disease/data-research/facts-stats/). While the survival rate from cardiovascular disease has increased over the past few decades, heart failure morbidity and mortality have remained staggering. A large part of this decline is imparted by cardiomyocyte (CM) loss and no supporting CM renewal capacity. Currently, proposed therapies can only reverse or slow certain aspects of cardiac dysfunction and disease, failing to replace lost CMs. Instead, the heart undergoes wound healing to replace dead CMs with fibrotic scar tissue, which impairs normal cardiac function (Leach and Martin, 2018). Therefore, there is a strong need for therapeutic strategies that increase CM renewal and, in turn, restore the contractile function of the heart and considerably improve survival and disease outcomes (Bertero and Murry, 2018; Eschenhagen et al., 2017; Hashimoto et al., 2018; Mann and Bristow, 2005).

One therapeutic strategy is to induce the proliferation of endogenous CMs (Zhao et al., 2020) by manipulating signaling molecules and regulatory proteins. This strategy requires an understanding of both the molecular mechanisms that promote CM cell division and those that result in a cell cycle exit (Payan et al., 2020). Over the past few years, it has been found that manipulating cell cycle regulatory and inhibitory proteins can promote CM proliferation. Previous studies have also shown that combining four cell cycle regulators could stimulate adult CM proliferation in vivo (Mohamed et al., 2018). In addition to these cell cycle proteins, several signaling pathways, such as YAP and neuregulin signaling pathway proteins, stimulate CM proliferation (Eschenhagen et al., 2017; Hashimoto et al., 2018; Hashmi and Ahmad, 2019; Mann and Bristow, 2005; Oyama et al., 2013; Yester and Kühn, 2017; Yutzey, 2017). MicroRNAs have also been implicated in stimulating CM proliferation (Eulalio et al., 2012; Ouyang and Wei, 2021; Tian et al., 2015; Torrini et al., 2019). Like other complex biological systems, innovative models and quantitative analyses are needed to unravel specific pathways and crosstalk between various regulators affecting CM proliferation.

In this study, we constructed and validated what is, to our knowledge, the first computational model of the CM proliferation regulatory network. The model integrates five modules that incorporate multiple signaling pathways. With a virtual knockdown screen, we predicted how the influence of particular species changes with different signaling contexts. Additionally, the model elucidated crosstalk between the growth factor and Hippo modules. We also identified key hubs for which the Hippo signaling pathway regulates CM proliferation. Predicted drivers of CM proliferation were further validated experimentally in cells and published data from mice.

RESULTS

A predictive computational model of the CM proliferation regulatory network

Although there has been substantial progress in discovering key regulators of CM cell cycle progression, their interconnections have yet to be defined at a network level. For this reason, we manually constructed a regulatory network model using ∼23 literature articles that described protein and gene interactions that regulate cell cycle progression. We also used resources such as Signor, which represents the cell cycle G1/S and G2/M phase transition pathways in humans (Licata et al., 2020); KEGG, which describes the mitotic cell cycle progression (https://www.genome.jp/pathway/hsa04110); and Cell Signaling, which also describes cell cycle regulation pathways (https://www.cellsignal.com/pathways/by-research/cell-cycle-regulation-pathways). Literature descriptions of these interactions and mechanisms were categorized based on whether they described a direct interaction (e.g. RB1 inhibits Cyclin D) or an indirect relationship (e.g. CDC25A overexpression enhances DNA replication). Direct interactions were used for model development, whereas indirect relationships were reserved to later validate the model. For pathways with inadequate support in CMs, literature for central proliferation nodes from other cell types were used where necessary.

The CM proliferation regulatory network integrates six modules that describe the interactions between signaling pathways and cell cycle regulators (Fig. 1A). The network links regulatory modules of the cell cycle: core cyclin-dependent kinases (DNA replication module) regulating the G1/S checkpoint; mitosis and DNA damage responses regulating the G2/M checkpoint (mitosis module); as well as well-studied signaling pathways in the heart: growth factor and Hippo modules (Xin et al., 2011, 2013; Zheng et al., 2020). Each module contains proteins, transcription factors, or genes that were key to the overall process. For example, the DNA replication module incorporates major cyclin and cyclin-CDK complexes and cyclin inhibitors that play a role in the cell cycle progression through activating E2F family members (E2Fa) and the G1/S checkpoint (Fig. 1B). The mitosis module incorporates DNA damage cues (ATM and ATR) that activate parallel pathways that inhibit the primary regulator of mitosis, cyclin B/CDK (Fig. 1C). The main effectors of cytokinesis in the heart have not been well studied. The cytokinesis module relies on species and interactions that have been primarily studied in other cell types. This included those examined in CM regulation, such as anillin, shown by Engel et al. as a crucial process for cytokinesis in CMs, and Ect2, shown to contribute to cytokinesis and increased binucleated cells (Fig. 1D) (Bergmann, 2021; Engel et al., 2006; Jiang et al., 2019; Liu et al., 2019). Extracellular factors have been identified to activate intracellular receptors involved in CM proliferation: FGF1, Nrg1, and IGF1 (Rebouças et al., 2016). The growth factor module shows how these factors interlink with cell cycle progression regulators (Fig. 1E). The Hippo signaling pathway is among the best-established modules for regulating CM proliferation. The Hippo module incorporates regulators such as YAP and TEAD that substantially affect CM proliferation (Xin et al., 2013) (Fig. 1F). The output module interlinks the phenotypic outputs of the other modules, representing how experimentally measured aspects of cell cycle activity (DNA replication by EdU or Ki67), mitosis by phospho-Histone 3 (pHH3), abscission by cytokinetic midbody converge on polyploidy, binucleation, or cytokinesis (e.g. completed proliferation) (Fig. 1G). Species or regulators with dashed borders that overlap between modules were used to connect the modules to create the CM proliferation network model. For example, cMyc is found in the Hippo, growth factor, and DNA replication modules, so those reactions were combined (Fig. S1). Fig. S1 shows the network connected with overlapping species, with the network activity simulated in the basal state.

Fig. 1.

A computational model of the cardiomyocyte proliferation regulatory network. (A) Schematic describing the model's five regulatory modules and their interactions. Arrows between modules represent one or more reactions that link species from one module to species in another module. The table (bottom) describes the number of species and reactions within each module. (B-G) Activity flow diagram for each module, with dashed boxes representing those species that appear in multiple modules. Phenotypic outputs are represented as gray hexagons.

To convert the network into a predictive computational tool, network reactions were modeled using logic-based differential equations (LDEs), a well-validated computational approach (Ryall et al., 2012; Tan et al., 2017; Zeigler et al., 2016). In an LDE model, the normalized activation of each node is represented by ordinary differential equations with saturating Hill functions and continuous logical AND or OR logic gates to characterize pathway crosstalk. OR gating is used when each input to a node is sufficient but not necessary for activation, whereas AND gating is used when each input is required. A previous study identified default values of the parameters (Y_max, EC₅₀, w, etc.) that most accurately predict the results of knockdown screens compared to a model where all biochemical parameters were measured experimentally (Kraeutler et al., 2010). Subsequent studies started from these default values and further demonstrated that model accuracy was robust to random variation in the parameters (Tan et al., 2017; Zeigler et al., 2016). Consistent with these previous models, we performed robustness analysis that demonstrated that the CM proliferation model accuracy (compared against 78 experiments) is maintained at >80%, with up to 35% variation in Y_max, 30% variation with w, and a variation of >50% with EC₅₀ (Fig. S4). LDEs were generated in Netflux and implemented in MATLAB to observe model dynamics and simulate perturbations to the model (Movie 1). Baseline conditions of model inputs were adjusted based on literature information about their presence in CM phenotypes (neonatal, adult, and embryonic validation). More specifically, if they are inhibitors or activators of cell cycle progression. For example, MST1, a major YAP/TAZ/TEAD signaling inhibitor, is thought to have a high activity within the adult phenotype, so the node input was set between 0.8 and 1 activation.

To examine the predictive accuracy of the model, we simulated perturbations previously reported in CM literature not used to construct the model. For example, the model predicted that adding growth factors Nrg and FGF with p38 inhibition increases DNA replication (Fig. 2A). These predictions are consistent with the published data in CMs from Engel et al., which found that p38 activity augments growth factor-mediated DNA synthesis in neonatal CMs (Engel et al., 2005). We compared model predictions to 81 published experimental observations of in vivo or in vitro CMs from ∼33 literature articles not used for model construction. Overall, the model accurately predicts 74 of the 81 (91.35%) observations. Fig. 2B summarizes the validation of three different validation relationship types: input-output (no change), inhibition (decrease), and overexpression (increase). We first focused on validating the phenotypic outputs of the model by searching for descriptions of experiments that measured: EdU incorporation or Ki67 expression (indication of DNA replication), pHH3 (indicating mitosis), and Aurora B at the midbody (cytokinesis). We included validations of individual species by looking for gene expression measurements in literature to expand our validation. We validated important cyclins and their CDK complexes (cyclin A, B, and E) as well as well-characterized cell cycle inhibitors (RB1, p27, p107, p21, and p53) that affect cell cycle progression (Fig. 2B). The model predicts 91.35% of observations found in the literature, but there are some exceptions. For example, the three model predictions and experimental observations for polyploidization did not match, indicating the complexity surrounding the possibility of cell phenotypes in CM populations.

Fig. 2.

Validation of model predictions against independent experimental data from the literature. (A) Model prediction of DNA synthesis in response to p38 inhibition and a combination of p38 inhibition with growth factors, compared to experimental data from Engel et al. (2005). (B) Qualitative comparison of model predictions in response to the noted stimuli, compared with published experimental observations of in vitro and in vivo cardiomyocyte proliferation. The model validates 74 of 81 (91.35%) predictions.

Influential drivers of CM proliferation with Nrg1 stimulation

After validating the model's predictive capacity, we performed a network-wide sensitivity analysis to characterize the functional roles in the CM proliferation regulatory network. We first simulated a complete virtual knockdown screen of the model under baseline conditions (Fig. S3A), which showed that no knockdowns caused substantial decreases in DNA replication or cytokinesis. This is consistent with a low baseline proliferation rate described in CM literature. As Nrg1 has previously been found to increase the proliferation of CMs (Bersell et al., 2009; D'Uva et al., 2015; D'Uva and Tzahor, 2015; Rupert and Coulombe, 2015), we then sought to identify the species most influential in driving Nrg1-dependent proliferation. We identified the most influential species as those for which knockdown produced the highest summed change in the phenotypic outputs of the network, along with the top measured species (Fig. 3A). This analysis allowed us to predict the network inhibition response of specific proteins, receptors, or genes. The topmost changed or sensitive species included key cell cycle regulatory proteins such as cyclins, cyclin-CDK complexes, and cell cycle inhibitory proteins. Additionally, phenotypic outputs were highly influenced by the knockdown of these most influential species in the activated Nrg1 context.

Fig. 3.

Virtual knockdown screen to predict regulators of CM proliferation with Nrg1 stimulation. (A) Knockdown of individual nodes (columns) in the context of high Nrg1 (input=0.8) and the effect on network node activities (rows). (B) Network visualization of the predicted effect of node knockdowns on cytokinesis.

Notably, this analysis revealed complex relationships between cell cycle outputs. For example, knockdown of genes upstream of DNA replication [e.g. p38 (Mapk1)] was predicted to increase all five outputs because they act in series. In contrast, knockdowns of genes specific to the cytokinesis module (e.g. Bub1, Apc) differentially affect cytokinesis versus polyploidy or binucleation because these are competing parallel paths.

As regulators of bone fide CM cell division are not well characterized, we further simulated a global knockdown screen for cytokinesis. We found that the influence of particular genes was correlated with topological features of the network, such as eccentricity and shortest path length (Fig. S5, Table S1). In the baseline state, the model predicts that knockdown of only a few species (PKA, Lats1, RB1, and SMAD3) are sufficient to increase cytokinesis (Fig. S6A). But in the context of high Nrg1 (Fig. S6B), the virtual knockdown screen (Fig. 3A,B) predicted many negative regulators of DNA replication (e.g. RB1, p107, GSK3β, and other CKIs) and mitosis (e.g. ATR/ATM, APC, p53, and p21; see outputs in Fig. 3A). In addition to these negative regulators, knockdown simulations predicted positive regulators of Nrg-dependent cytokinesis such as CyclinD/CDK4, a direct regulator of DNA replication, PI3K/AKT, and CDC25BC, an upstream regulator of mitosis.

Context-dependent regulation of CM proliferation

Next, we asked whether the influence of species in the CM proliferation network differs depending on the proliferative stimulus. We compared the influence of individual species on the DNA replication, mitosis, and cytokinesis modules in the context of increased Nrg1, YAP activity, or baseline conditions. We first performed a sensitivity analysis under baseline conditions (Fig. S2). We identified the 25 species under basal conditions with the highest influence over network activity and the phenotypic outputs. In the basal state, the most influential species included those mediating signals from the growth factor and Hippo signaling module to the rest of the network: PI3K/AKT, GSK3B, SMAD3, Nrg1/ErbB, and Lats1 (Fig. S3A).

We next visualized the network under YAP stimulation (Fig. S3C). With an increase in activated YAP, numerous species in the DNA replication module changed in influence compared with the basal state or Nrg1 context (Fig. S3A,B). For example, with increased YAP activity, CDC25A and key cyclins were the most influential species, while cell cycle inhibitors were lowered in their influence (Fig. 4A). Unexpectedly, several species associated with ‘upstream’ growth factor signaling, such as AKT and cMyc, were highly influential in mediating YAP-dependent proliferation. Species that were highly influential with high Nrg1 decreased in influence with high YAP, whereas species that were not considered as influential in the Nrg1 state became highly influential in the high YAP state, such as CyclinA/CDK2, CDC25BC, and FOXM1. Overall, regulation of the DNA replication module is highly context-dependent, while regulation of mitosis and cytokinesis modules were relatively context-independent (Fig. 4A).

Fig. 4.

Influence of node knockdowns shifts with context, revealing crosstalk from Hippo to growth factor modules. (A) Total influence of node knockdowns on the DNA replication, mitosis, and cytokinesis modules, compared across multiple signaling contexts: baseline (top), high Nrg (middle), and high YAP (bottom). Total influence sums the overall effect of a node knockdown on a network module. (B) The total influence of each network module varies depending on whether a basal state, high Nrg, or high YAP signaling context is applied. (C) Capillary electrophoresis western blot for phosphorylated ERK, β-actin, and GAPDH from neonatal CMs treated with Nrg or TT10 for 30 min. (D) Model predictions of AKT and ERK activity of acute response to Nrg or TT10 (time constants for gene expression set to 100). (E) Quantification of effects of Nrg or TT10 treatment on p-ERK (from western blot in panel C, n=3) or p-AKT [from western blot from Hara et al. (2018), n=1].

To gain a more global view of the relationships between network modules, we computed the total influence of each module and quantified how the total influence of each module varies with the signaling context (Fig. 4B). As expected, in the context of increased Nrg1, the growth factor module increased in influence, while the Hippo module slightly decreased compared to the basal state. This is followed by an increase in the influence of DNA replication, mitosis, and cytokinesis modules. In contrast, and supporting the above predictions for AKT and cMyc in Fig. 4A, stimulation of YAP increases the influence of crosstalk from Hippo to growth factor modules. This leads to an even greater increase in the influence of the DNA replication, mitosis, and cytokinesis modules compared to the basal context or high Nrg1 context. Overall, these simulations identify an unexpected crosstalk from YAP to the growth factor module that is predicted to be important for context-dependent CM cell cycle progression.

From inspection of the network structure, we hypothesized that crosstalk from the YAP module to the growth factor module occurs via slow transcriptional pathways rather than fast post-transcriptional signaling. Supporting this hypothesis, the model predicted that species in the growth factor module ERK and AKT would acutely respond to Nrg but not to YAP (Fig. 4C). To experimentally validate this prediction, we performed capillary electrophoresis western blots of p-ERK from neonatal CMs treated with either Nrg or the YAP-activating compound TT10 for 30 min. Indeed, these experiments validated the model prediction that acute stimulation with Nrg but not YAP is sufficient to induce p-ERK (Fig. 4D). Similarly, Hara et al. (2018) found increased p-AKT with acute Nrg but not TT10 treatment, further validating the model predictions (Fig. 4E). While the model did not predict the observed acute TT10-induced decrease in p-AKT, Hara hypothesized that this could be explained by the Wnt/β-catenin regulation of YAP nuclear translocation, which is not in the current model (Hara et al., 2018).

Crosstalk between modules of the CM proliferation regulatory network

The role of the Hippo pathway in cardiac regeneration and as a potential therapeutic target is well established (Wang et al., 2018; Zheng et al., 2020). To identify potential mediators of YAP signaling that crosstalk to the growth factor module, we simulated long-term YAP stimulation (in contrast to acute YAP stimulation in Fig. 4D) and found species with greater than 15% change in activity (Fig. 5A). About 70% of the DNA replication module had a significant increase in response to an increase in YAP activity and network outputs. Interestingly, FoxM1, PI3K, and cMyc almost doubled their activation with an increase in YAP, and all three are directly regulated by the YAP target transcription factor TEAD in the model (Fig. 5B). The transcriptional nature of these predicted interactions is consistent with the lack of crosstalk from YAP to growth factor module at short timescales in Fig. 4D.

Fig. 5.

Prediction and experimental validation of cardiomyocyte cell cycle progression mediated by the Hippo pathway via PI3K, cMyc, and FoxM1. (A) Network species affected by at least 15% in response to high YAP, grouped by module. (B) Network hypothesis of key hubs in the Hippo signaling pathway and how they affect the phenotypic outputs. (C) Quantification of model predictions of inhibitors for PI3K, cMyc, and FOXM1 alone and in the context of high YAP (TT10+). Please note the PI3K inhibition is very close to zero. (D) Quantification of Ki67 and pHH3 from experiments in rat neonatal CMs labeling Ki67 (cell cycle activity) or pHH3 (mitosis) in response to 10 µM TT10; PI3K inhibitor (1 µM Ly294002), cMyc inhibitor (20 µM 10058-F4), or FOXM1 inhibitor (1 µM RCM1) (n=8). Data are mean±s.e.m. (E) Representative images from the experiment described in D. Statistical significance was determined using a two-way ANOVA on all experimental groups with a post-hoc Dunnett's test comparing the two different isolations. *P≤0.05, ***P≤0.001.

We further explored whether PI3K, FoxM1, and cMyc mediate the proliferative effects of YAP. Previously, Lin et al. (2015) found, in neonatal rat ventricular CMs, that YAP functions directly upstream of PI3K to enhance its expression. In addition, they found that Pik3cb, which links the Hippo-YAP pathway to the PI3K-AKT signaling pathway and promotes CM cell cycle progression and survival, is a crucial direct target for YAP (Lin et al., 2015). Other systems have shown the connection between YAP and FoxM1 in malignant mesothelioma cells, showing that YAP regulates FoxM1 transcription directly through TEAD (Mizuno et al., 2012). We simulated the addition of TT10, a YAP-activating drug (Hara et al., 2018), coupled with inhibition of PI3K, FoxM1, or cMyc, on CM cell cycle progression. As in past experiments, the model predicted that increasing YAP with TT10 greatly increases the percentage of cells in the DNA replication and mitosis phases (Hara et al., 2018). Further, the model predicted that inhibition of PI3K or cMyc would reduce the proliferative effect of TT10, with a smaller effect for FOXM1 (Fig. 5C). To test these predictions experimentally, we cultured neonatal rat CMs with combined activation of YAP (10 µM TT10) and inhibition of PI3K (1 µM Ly294002), cMyc (100 µM 10058-F4), or FoxM1 (1 µM RCM1). As shown in Fig. 5D,E, the increased cell cycle activity (Ki67) and mitosis (pHH3) induced by YAP activator TT10 were significantly attenuated by inhibition of cMyc and PI3K but not FoxM1 inhibition. These neonatal CM in vitro results validate the model prediction that YAP-induced cell cycle progression is mediated by module crosstalk involving transcriptional regulation of cMyc and PI3K.

We next asked if YAP may regulate the predicted targets of PI3K, cMyc, and FOXM1 in vivo. In Monroe et al. (2019), an active version of YAP, termed YAP5SA, was overexpressed in CMs in adult mice (Fig. 6A). To examine whether YAP induced the expression of PI3K, cMyc, and FOXM1 in vivo, we analyzed RNA-seq data from Monroe et al. (2019) (Fig. 6B). Indeed, cardiac overexpression of YAP5SA increased mRNA for Myc and Foxm1. To examine whether there is coincident remodeling of the chromatin within the promoter regions of these genes, we then analyzed ATAC-seq data (Monroe et al., 2019) from the same mouse model of YAPS5A (Fig. 6C-E). Promoters of both Myc and PI3Kca exhibited an altered chromatin open state. Only cMyc demonstrated fully consistent effects on YAP-dependent CM proliferation, YAP-dependent mRNA, and YAP-dependent chromatin opening, as predicted by the model.

Fig. 6.

In vivo RNA-seq and ATAC-seq further validate the model-predicted role of cMyc in YAP regulation of CM cell cycle. (A) Schematic of experimental design of cardiac constitutively active YAP5SA transgenic mice (Monroe et al., 2019). (B) RNA-seq data showing expression of genes with and without the YAP5SA transgene. Statistical significance was determined by performing an unpaired two-tailed Student's t-test. **P≤0.01, ***P≤0.001. Data are mean±s.e.m. (C-E) ATAC-seq data describing open chromatin remodeling between genes PIK3ca, Myc, and Foxm1 with and without the YAP5SA transgene. Gray bars denote statistically significant regions of accessible chromatin (P<0.00001). Scale bars represent the genome browser track height and are proportional to the number of reads mapped to each genomic position.

DISCUSSION

In this study, we developed a dynamic map of the CM proliferation network to identify drivers and network principles of CM proliferation. By creating and validating a predictive model of the CM proliferation network, we showed how the crosstalk between signaling pathways and core cell cycle regulators allows CMs to progress through key cell cycle checkpoints (DNA replication, mitosis, and cytokinesis). Our review of the literature indicated multiple complex molecular pathways that regulate CM proliferation, including growth factors, Hippo signaling, G1/S transition, G2/M transition, or cytokinesis pathways (Díaz Del Moral et al., 2021; Hashmi and Ahmad, 2019; Johnson et al., 2021; Johnson and Halder, 2014; Payan et al., 2020; Wang et al., 2018). Several review articles (Besson et al., 2008; Díaz Del Moral et al., 2021; Mia and Singh, 2019; Wang et al., 2018; Zheng et al., 2020) also organized the literature based on these distinct pathways or processes, which we used to define the boundaries of the six modules. However, how these molecular pathways work together is not well characterized. Therefore, we designed the model to incorporate each of these established modules but also how they work together to drive CM proliferation. Overall, our model defines CM proliferation through 72 CM proliferation species connected by 88 reactions. The model was validated at 91.35% compared to independent, published experiments in CMs.

Because this is a literature-based network model, each component or interaction has been studied individually. Beyond the ∼30 papers used to validate the model in Fig. 3, the model makes broader predictions of how these components coordinate proliferation. For example, Fig. S2 provides ∼5000 predictions of how each protein responds to the knockdown of every other protein. These computational screens identified that YAP regulates CM cell cycle activity via cMyc, which we experimentally validated in Fig. 5 in vitro and Fig. 6 in vivo. There are many other insights that can be validated in future studies or used to expand the model based on its open-source availability on GitHub (https://github.com/saucermanlab/Cardiomyocyte-Proliferation-Network).

Identification of knowledge gaps in CM proliferation

While the model validated 91.35% of input-output relationships from observations not used in model development, five input-output relationships were incorrectly predicted. For example, MDM2, which targets p53 (Fig. 1C), was shown by Stanley-Hasnain et al. (2017) to increase DNA synthesis in CMs when knocked down. However, our model predicted that it decreased DNA synthesis activity. MDM2 plays a diverse role in the cardiovascular systems, and studies to elucidate its interaction with upstream and downstream regulators are needed (Hauck et al., 2017; Lam and Roudier, 2019; Stanley-Hasnain et al., 2017).

Our model validation is notably weakest in predicting experiments on polyploidization, indicating a need for more experimental data to characterize polyploidy and cytokinesis pathways. For example, the model predicted a decrease in polyploid cells in response to Lats1 knockdown, whereas experimental data from the literature showed an increase (Heallen et al., 2013). In a review, Derks and Bergmann (2020) found that, despite numerous explanations for CM polyploidization that have been proposed, more work needs to be done to identify distinct mechanisms regulating CM ploidy. Similar findings were reported in Kirillova et al. (2021) and Pandit et al. (2013), where one of the outstanding questions was which genes are essential for polyploidization and by which mechanisms do they regulate this process. These incorrect predictions highlight areas for which future model revision and experiments are necessary. Because such data are limited in CMs, we performed an additional validation against polyploidization experiments from other cell types as summarized in Pandit et al. (2013). The CM proliferation model predicted 85% (6 of 7) of non-CM experiments (Fig. S7).

This network model largely represents the canonical roles of proteins and does not distinguish between family members, which sometimes have distinct roles in proliferation. For example, the E2Fa node represents only the pro-proliferative role of E2F family members 1/2/3/4, without capturing their distinct effects on apoptosis (Ebelt et al., 2005) or the anti-proliferative roles of E2F7 or E2F8 (Yu et al., 2023). This model may provide a foundation for focused study of particular members or isoform-specific roles.

Additionally, there is limited information on the pathways that regulate cytokinesis in CMs. The ‘AND’ gate structure leading to cytokinesis depends on immediate upstream signals MYH7, Factin and Septin (Fig. 1D). The inhibitory effect of APC controls these signals (the only input into this module). The knockdown of any of these cytokinesis elements is predicted to block cytokinesis. This shows that these upstream elements are important but, in themselves, are not sufficient. Future experiments are needed to identify conserved or differential mechanisms of polyploidization and cytokinesis in CMs.

Network crosstalk and its control of CM proliferation

The main focus of experimental studies on CM proliferation has been to identify strategies that enhance cell cycle entry. This predictive model provides a framework to identify effective ways to induce CM proliferation in response to specific stimuli. Nrg1 is well described to promote CM proliferation and is a prime molecular target to promote cardiac regeneration. In addition, YAP, a major effector of Hippo, is a central regulator of CM proliferation and survival (Del Re et al., 2013; Heallen et al., 2011, 2013; Lin et al., 2015; von Gise et al., 2012; Xin et al., 2013). Using the influence metric from our sensitivity analysis, we compared total module influence in the basal, high Nrg1, and high YAP contexts. Analysis of our model identified crosstalk between the growth factors and Hippo signaling modules. When YAP is activated, not only does the influence of the Hippo module itself increase, but the model predicted a surprising crosstalk in which the influence of the growth factor module also increased. For this reason, we investigated the role of YAP in the downstream regulation of CM proliferation.

Previous studies of YAP in CM proliferation focused on upstream inhibitors (Flinn et al., 2020) via MST1/2, SAV, or Lats 1 (Heallen et al., 2011, 2013; Monroe et al., 2019). However, the direct downstream targets that mediate YAP-induced CM proliferation are unclear. Our model predicted that three effectors (PI3K, cMyc, and a smaller effect of FoxM1) connect YAP to the growth factor and mitosis modules. In both model predictions and experimental validations in cultured CMs, inhibition of PI3K or cMyc significantly attenuated the effect of YAP activation on cell cycle. Others have implicated YAP in increased expression of PI3K (Lin et al., 2015). Our analysis of in vivo RNA-seq data from mice expressing constitutively active YAP5SA only in CMs (Monroe et al., 2019) identified a significant increase in Myc and FoxM1 expression but not PIK3ca. Further, in ATAC-seq data from hearts overexpressing YAP5SA (Monroe et al., 2019), we found increased chromatin accessibility at promoters of PIK3ca and Myc but not FoxM1. Together, these in silico, in vitro, and in vivo data support a YAP-cMyc axis for CM cell cycle activity.

Overexpression of Myc induces CM proliferation in vitro and in vivo in several contexts, with open chromatin and Myc binding near mitotic genes (Bywater et al., 2020). But to our knowledge, crosstalk of YAP with Myc has not been reported in the heart. Our model prediction and experiments in neonatal CMs support a YAP-TEAD-Myc pathway for CM cell cycle activity. Further, our analysis of ATAC-seq and RNA-seq data from Monroe et al. (2019) validate that YAP induces Myc chromatin availability and gene expression in adult mouse hearts. In MDA-MB-231 breast cancer cells, YAP/TAZ/TEAD bind directly to Myc enhancers through chromatin looping, with decreased acetylation of H3K27 and cell proliferation upon YAP/TAZ knockdown (Zanconato et al., 2015). YAP-TEAD-Myc signaling regulates the proliferation of cancer cells (Zanconato et al., 2015), tumorigenesis (Chen et al., 2018), and the growth of Drosophila imaginal discs (Neto-Silva et al., 2010). In the future, computational models and experiments are needed to better resolve how YAP promotes proliferation via Myc in the adult heart, including regulation by Mycn (Singh et al., 2018) and cyclin T1 (Bywater et al., 2020). Additionally, further model revision is needed based on these molecular mechanisms of YAP-TEAD-Myc interactions to distinguish between chromatin accessibility, transcription factor binding, and gene expression.

Limitations and future directions

This model was developed using logic-based differential equations and default parameters, which we have previously shown to exhibit a strong predictive accuracy for other networks (Kraeutler et al., 2010; Tan et al., 2017; Zeigler et al., 2016). As more information becomes available in the literature, we can adjust parameters to replicate different phenotypes, including the adult CM phenotype. From these changes, we can identify unique mechanisms present in CMs and, specifically, the adult heart. While this model's predictions are most relevant to immature CMs, it is the first molecular network model of CM proliferation.

In the future, we hope that we and others may extend this model to identify how factors like species, age, and experimental design regulate proliferation. For example, this model does not include newly-identified regulators of CM proliferation such as Nrf1 or complex in vivo environments such as myocardial infarction (Cui et al., 2020, 2021; Kuppe et al., 2022). However, these endeavors would span multiple manuscripts and the field currently lacks sufficient stage-specific data. For example, a previous foundational computational model of CM electrophysiology (Luo and Rudy, 1994) focused on adult guinea pig. This model became the foundation for a range of developmental and species-specific models in electrophysiology (Courtemanche et al., 1998; Paci et al., 2013; ten Tusscher et al., 2004). We believe the open availability of our code will enable similar dissemination and extension for additional factors.

The network model identifies gaps in the current understanding of CM proliferation signaling. Even though the five phenotypic outputs of the model have been extensively studied and verified by experimental data, the potential regulators that lead to binucleated or polyploid cells are still unclear. Our model structure focuses on the current information on cell cycle signaling networks that meet specified criteria for inclusion in a CM-specific model. This model provides an initial network framework for integrating additional discoveries in CM proliferation. As more information becomes available in CM proliferation literature, the model can be adapted. Additionally, the field can use our open-sourced model to adapt this model to other developmental stages or species

Conclusions

We developed a predictive model of the CM proliferation regulatory network that identifies the species and network structures that regulate the cell cycle of CMs. A sensitivity analysis of our model identified drivers of CM proliferation and showed that the drivers for DNA replication vary in influence with changes in signaling contexts. In contrast, node influence within mitosis and cytokinesis modules was more robust. Based on total module influence in different signaling contexts, the model predicted crosstalk from the Hippo module to the growth factor module involving a YAP-cMyc axis, which was validated experimentally and in vivo.

MATERIALS AND METHODS

Model construction

A predictive computational model of the cell cycle-proliferation regulatory network in CMs was manually constructed from experimental literature and resources such as KEGG, SIGNOR, and Cell Signaling (https://www.cellsignal.com/pathways/by-research/cell-cycle-regulation-pathways; https://www.genome.jp/pathway/hsa04110; Licata et al., 2020). We built the model in a modular way, starting with pathways that have been extensively studied in terms of proliferation and regeneration, such as the growth factor and Hippo signaling pathways. We then focused on creating modules that described the progression of the cell cycle like DNA replication, mitosis and cytokinesis. Literature search began by identifying references that indicate a role for specific proteins, whether by a diagram of the activity or the role of specific proteins in the heart. Direct molecular reactions between proteins were included if multiple sources supported the interaction. However, some interactions were not found in CMs. All papers involving in vitro or in vivo experiments performed in rat CMs or human induced pluripotent stem cells (hiPSC)-CMs were reserved for validation during the literature review. Once all modules were built, we examined each module for nodes that were common across all modules. We then made the necessary connections to create the overall model. Model outputs were set as general cell cycle phases that are measurable through experiments such as DNA replication (measured by EdU assay), general cell cycle activity (fluorescent staining of Ki67), mitosis (measured by pHH3-positive cells), or cytokinesis (measured by Aurora B presence at the midbody of the cell). The overall network contains 72 nodes and 88 reactions. The model includes 14 inputs, including receptor inputs [Nrg1, insulin-like growth factor (IGF), and fibroblast growth factor (FGF)] and five phenotypic outputs (DNA replication, mitosis, polyploid, cytokinesis, binucleation). This network integrates five modules that increase CM proliferation: DNA replication module, mitosis module, cytokinesis module, growth factor module, and Hippo module. Full documentation supporting model reactions (interactions) is provided in Table S2, and all references are provided in Table S4.

Signaling dynamics were predicted using a previously described LDE approach (Kraeutler et al., 2010). In this method, the activation of one node in the model by another is modeled using a normalized Hill function. Logical AND or OR operations were used to represent pathway crosstalk. OR gates are used when each input to a node is sufficient but not necessary for activation, whereas AND gates are used when each input is necessary. Default reaction parameters included reaction weight (w=1), Hill coefficient (n=1.4) and half-maximal effective concentration (EC₅₀=0.5), and default node parameters included initial activation (Y_init=0), maximal activation (Y_max=1) and time constant (τ=1). The reaction weight for model inputs (oxygen, HIPK2, AurkABora, SMAD3, AurB, Bub1, Nrg1, FGF2, IGF1, PKA, Mst1, Rho, and PRC1) was chosen to maximize or minimize the number of species activated between 25% and 90%, preventing a drastic increase or decrease in node activity to obtain the most information from the sensitivity analysis. The system of logic-based differential equations was auto-generated from Table S2 using Netflux (Clark et al., 2024) and implemented in MATLAB. As in past logic-based differential equation network models (Kraeutler et al., 2010; Tan et al., 2017; Zeigler et al., 2016), species (or nodes) refer to a small molecule, gene, protein, or process. Reactions (or edges) are activating or inhibiting relationships between network species.

Model validation

The literature used to validate the network's input-output relationships was identified by searching for each network relationship together with the term ‘cardiomyocyte’ or ‘cardiomyocyte proliferation’ in the PubMed or Google Scholar database. To create quality or reproducible models, validations used were only from studies using rats, mice, or human CMs. All supporting studies used were independent of those used to develop the model network. Validation was performed by comparing the qualitative increase, decrease, or no change in output activity of the model simulation to the experimental results. Based on statistics from the original studies, observations from the literature were encoded as increase, decrease, or no change. At baseline, input reaction weight parameters (w) were set to specific values based on network structure and available literature data. To simulate experiments with biochemical stimuli, input reaction weights were increased to 0.8 or 1. To simulate experiments with inhibition or knockdown, the corresponding maximum species value (Y_max) was set to 0.1 or 0. Changes of less than 0.5% were categorized as ‘no change’. All model validations, along with complete annotations for all validation simulations and supporting literature, are provided in Table S3.

Virtual knockdown screens

Functional analysis of the model was performed by simulating individual knockdowns for each of the 72 nodes in the network and predicting the corresponding change in the activity of every other node in the network. First, the steady-state activity of model nodes was obtained by running the model until there was a <0.05% change in activity for all nodes, or about 100 time units. Fig. S1 shows the predicted steady-state activity of all nodes under basal conditions. This predicted baseline steady state is consistent with experimental data, for example, low basal DNA replication, mitosis, and cytokinesis, while cell cycle inhibitors (RB1, p107, and p53) are high. We then knocked down the activity of each node one at a time and subtracted the basal activity values from the values in the knockdown state to calculate ‘Δ Activity’. Influence is measured as the number of nodes with a 25% change or more significant change in activity following the knockout of the perturbed node, and sensitivity is the number of nodes that will affect the target by a 25% change or greater when knocked out.

The CM proliferation regulatory network was exported from Netflux into Cytoscape (Shannon et al., 2003) for topological analysis. The Network Analyzer (Assenov et al., 2008) plug-in was used to calculate topological properties. The correlation coefficient for matching topological to functional metrics was computed using the fitlim function in MATLAB. The functional metrics as defined by the sensitivity analysis were: (1) influence, the number of nodes with an activity change more significant than 25% with the knockdown of node n; (2) sensitivity, the number of nodes that change the activity of node n by more than 25% when knocked down.

Robustness analysis

Network robustness to variation in model parameters was tested using a validation threshold of 5% absolute change (Tan et al., 2017). For each parameter (Y_max, w, n, and EC₅₀), new values were generated by sampling from a uniform random distribution with indicated half-width about the original parameter value. One hundred new parameter sets were created for each distribution range for each parameter, and simulations were run to compare model predictions with literature observations. No changes in validation accuracy resulted from varying τ or Y_init.

Cell culture

Cardiac myocytes were isolated from 1- to 2-day-old Sprague-Dawley rats using a Neomyt isolation kit (Cellutron). The cells were cultured in plating media (low-glucose Dulbecco's Modified Eagle Media, 17% M199, 10% horse serum, 5% fetal bovine serum, 1% L-Glutamine, 10 U/ml penicillin, and 50 mg/ml streptomycin) at a density of 30,000 cells per well of a 96-well Corning CellBIND plate. After 48 h of incubation with daily media change, the medium was replaced with Williams E Medium (Life Technologies, A1217601) supplemented with cocktail B supplement (Life Technologies, CM4000) serum-starved for at least 4 h before treatment.

Proliferation assay

DMSO (0.2%) and TT10 drug (10 µM, Aobius, 2230640-94-3) were used as the negative and positive controls, respectively. Cells were treated with a PI3K inhibitor (1 µM, Selleck Chem, Ly294002-S1105), cMyc inhibitor (20 µM, Abcam, 10058-F4), and FoxM1 inhibitor (1 µM, Tocris, 6845) in serum-free media for 2 days with and without the addition of TT10.

Immunofluorescence and imaging

After 2 days of treatment, the cells were fixed in 4% paraformaldehyde and fluorescently labeled with DAPI (1:1000, Life Technologies, 62248) and primary antibodies: monoclonal anti-α-actinin (1:200, Sigma Aldrich, A7811, RRID:AB_476766) or cardiac troponin T (cTnT) (1:200, Abcam, 45932, RRID:AB_956386), anti-Ki67 (1:200, Thermo Fisher Scientific, 14-5698-80, RRID:AB_10853185), and phospho-Histone 3 (1:200, Cell Signaling Technology, 9706, RRID:AB_331748). Cells were then treated with secondary antibodies: AlexaFluor 568 (1:200, Thermo Fisher Scientific, A-11031, RRID:AB_144696), AlexaFluor 488 (1:200, Thermo Fisher Scientific, A-11008, RRID:AB_143165), and AlexaFluor 680 (1:200, Thermo Fisher Scientific, A32788, RRID:AB_2762831). Cells were then imaged with the Operetta CLS high-content imaging system (Perkin-Elmer) using a 10×0.3 NA objective.

Image processing

Image analysis scripts were developed in MATLAB to segment and classify the cells automatically. Nuclear segmentation methods adapted from Woo et al. and Bass et al. apply to images of nuclei stained with DAPI (Bass et al., 2012; Woo et al., 2019). Briefly, a median filter with a window size of three pixels was applied to smooth the images, and the blurred nuclei were segmented using an Otsu threshold. Next, we identified clumps of nuclei by measuring the circularity of the segment's nuclei (<0.75). Objects with a low circularity factor were further separated by applying an erosion operation and then a watershed transform. Finally, objects outside the range of CM nuclei and those touching the image borders were removed. A mask of these objects was used to measure the integrated intensities of DAPI, α-actinin, EdU, Ki67, and pHH3. Thresholds for determining whether a nucleus was positive for a label were calculated as a fraction of the standard deviation from the model of the integrated intensities.

Statistics

Primary rat CMs were isolated from two different isolations. For each isolation, the treatment groups had at least four replicate wells for a total of eight replicates per experimental group. We performed a two-way ANOVA on all experimental groups with a post-hoc Dunnett's test comparing the two different isolations. There were 200-800 nuclei counted per well in the experiments. Error bars represent the standard error of the mean. Statistical significance was set at P<0.05 or P<0.01.

mRNA sequencing analysis

RNA-seq data was retrieved from the GEO repository database (GSE123457) consisting of RNA extracted from Bead-bound PCM1⁺ nuclei from two control samples and two samples of YAP5SA-overexpressing CMs. Relative mRNA abundance values for PI3Kca, Myc, and FoxM1 were averaged for the two control samples and the two YAP5SA-overexpressing samples. Significance was estimated by performing an unpaired two-tailed Student's t-test.

Assay for transpose accessible chromatin sequencing analysis

ATAC-seq data was retrieved from the GEO repository database (GSE123457) consisting of two control samples of CM nuclei extracted from an Myh6-MCM mouse strain and two YAP5SA-overexpressing samples of CM nuclei extracted from a YAP5SA gain-of-function transgenic mouse strain. Approximately 50,000 CM nuclei were extracted for each sample as input for ATAC-seq. Paired end 2×75 bp sequencing was performed using the Illumina Nextseq 500 instrument and reads were mapped to the mouse genome (mm10) using Bowtie2. We downloaded the processed BigWig files to use for peak-annotation at the Pik3ca, Myc, and Foxm1 loci. BigWig files were converted to bedgraph files using the UCSC-tools/3.7.4 (Kent et al., 2010) conversion package bigWigToBedGraph. Bedgraph files for each experimental condition were then merged using BEDTools unionbedg (Quinlan and Hall, 2010). Finally, we called the accessible chromatin region peaks for each of the two merged files using MACS2 bdgpeakcall (Zhang et al., 2008) with the default P-value cutoff threshold of 1e-5. Bedgraph and narrowPeak files were visualized using the Integrative Genomics Viewer (Robinson et al., 2011).

Protein analysis

For protein analysis, we used an auto-western blot service (RayBiotech). Cell samples of control, Nrg-treated and TT10-treated neonatal rat CMs were sent to RayBiotech, where they used their high-throughput automated capillary electrophoresis western blotting system.

Peer review history

The peer review history is available online at https://journals.biologists.com/dev/lookup/doi/10.1242/dev.204397.reviewer-comments.pdf