Cellular senescence—the permanent arrest of cycling in normally proliferating cells such as fibroblasts—contributes both to age‐related loss of mammalian tissue homeostasis and acts as a tumour suppressor mechanism. The pathways leading to establishment of senescence are proving to be more complex than was previously envisaged. Combining in‐silico interactome analysis and functional target gene inhibition, stochastic modelling and live cell microscopy, we show here that there exists a dynamic feedback loop that is triggered by a DNA damage response (DDR) and, which after a delay of several days, locks the cell into an actively maintained state of ‘deep’ cellular senescence. The essential feature of the loop is that long‐term activation of the checkpoint gene CDKN1A (p21) induces mitochondrial dysfunction and production of reactive oxygen species (ROS) through serial signalling through GADD45‐MAPK14(p38MAPK)‐GRB2‐TGFBR2‐TGFβ. These ROS in turn replenish short‐lived DNA damage foci and maintain an ongoing DDR. We show that this loop is both necessary and sufficient for the stability of growth arrest during the establishment of the senescent phenotype.
The phenomenon of cellular ‘senescence’—the permanent arrest of division in normally proliferating mammalian cells such as fibroblasts—is thought to be a central component of the ageing process. Senescence contributes both to age‐related loss of tissue homeostasis, as the loss of division capacity leads to impaired cell renewal, and also to protect against cancer, because it acts to block the uncontrolled proliferation of cells that may give rise to a malignant tumour. Replicative senescence is triggered by uncapped telomeres or by ‘unrepairable’ non‐telomeric DNA damage. Both lesions initiate the same canonical DNA damage response (DDR) (d'Adda di Fagagna, 2008). This response is characterized by activation of sensor kinases (ATM/ATR, DNA‐PK), formation of DNA damage foci containing activated H2A.X (γH2A.X) and ultimately induction of cell cycle arrest through activation of checkpoint proteins, notably p53 (TP53) and the CDK inhibitor p21 (CDKN1A). This signalling pathway continues to contribute actively to the stability of the G0 arrest in fully senescent cells long after induction of senescence (d'Adda di Fagagna et al, 2003). However, senescence is more complex than mere CDKI‐mediated growth arrest. Senescent cells alter their expression of literally hundreds of genes (Shelton et al, 1999), prominent among these being pro‐inflammatory secretory genes (Coppe et al, 2008) and marker genes for a retrograde response induced by mitochondrial dysfunction (Passos et al, 2007a).
There is a growing evidence that multiple mechanisms interact to underpin ageing at the cellular level (Kirkwood, 2005; Passos et al, 2007b) necessitating a systems biology approach if the complex mechanisms of ageing are to be understood (Kirkwood, 2008). With respect to cell senescence, the two major unanswered questions are (i) How does a DNA lesion that can be repaired, at least in principle, induce and maintain irreversible growth arrest? and (ii) How does a growth arrest trigger a completely different cellular phenotype as soon as it becomes irreversible?
To understand those questions, we performed a kinetic analysis of the establishment phase of senescence initiated by DNA damage or telomere dysfunction, focussing on pathways downstream of the classical DDR. Using an approach that combined (i) in‐silico interactome analysis, (ii) functional target gene inhibition, (iii) stochastic modelling, and (iv) live cell microscopy, we identified a positive feedback loop between DDR and mitochondrial production of reactive oxygen species (ROS) as necessary and sufficient for long‐term maintenance of growth arrest. Using pathway log likelihood scores calculated by a quantitative in‐silico interactome analysis to guide siRNA and small molecule inhibition experiments, and using results of sequential and combined inhibition experiments to refine the predictions from the interactome analysis, we found that DDR triggered mitochondrial dysfunction leading to enhanced ROS activation through a linear signal transduction through TP53, CDKN1A, GADD45A, p38 (MAPK14), GRB2, TGFBR2 and TGFβ(Figure 2D). We hypothesized that these ROS stochastically generate novel DNA damage in the nucleus, thus forming a positive feedback loop contributing to the long‐term maintenance of DDR (Figure 3A). First confirmation came from static inhibitor experiments as before, showing that nuclear DNA damage foci frequencies in senescent cells were reduced if feedback signalling was suppressed. To formally establish the existence of a feedback loop and its relevance for senescence, we used live cell microscopy in combination with quantitative modelling.
We transformed the conceptual model shown in Figure 3A into a stochastic mechanistic model of the DDR feedback loop by extending the previously published model of the TP53/Mdm2 circuit (Proctor and Gray, 2008) to include reactions for synthesis/activation and degradation/deactivation/repair of CDKN1A, GADD45, MAPK14, ROS and DNA damage. The model replicated very precisely the kinetic behaviour of activated TP53, CDKN1A, ROS and DNA damage foci after initiation of senescence by irradiation. Having established its concordance with the experimental data, the model was then used to predict the effects of intervening in the feedback loop. The model predicted that any intervention reducing ROS levels by about half would decrease average DNA damage foci frequencies from six to four foci/nucleus within about 15 h. It further predicted that this would be sufficient to reduce CDKN1A to basal levels continuously for at least 6 h in about 20% of the treated cells, thus allowing a significant fraction of cells to escape from growth arrest and to resume proliferation. This should happen even if the intervention into the feedback loop was started at a late time point (e.g. 6 days) after induction of senescence.
To analyse DNA damage foci dynamics we used a reporter construct (AcGFP–53BP1c) that quantitatively reports single DNA damage foci kinetics in time‐resolved live cell microscopy (Nelson et al, 2009). Foci frequency measurements quantitatively confirmed the prediction from the stochastic model. More importantly, we found that many individual foci in both telomere‐ and stress‐dependent senescence had short lifespans with half‐lives below 15 h. Feedback loop inhibition reduced only the frequencies of short‐lived DNA damage foci in accordance with the hypothesis that ROS production contributed to DDR by constant replenishment of short‐lived DNA damage foci.
Finally, we inhibited signalling through the loop at different time points after induction of senescence by ionizing radiation and measured ROS levels, DNA damage foci frequencies and proliferation markers. Treatments with the MAPK14 inhibitor SB203580 or the free radical scavenger PBN were used to block the loop. The results quantitatively confirmed the model prediction and indicated that the feedback loop between DDR and ROS production was both necessary and sufficient to maintain cell cycle arrest for at least 6–10 days after induction of senescence. Interestingly, the loop was still active at later time points and in deep senescence, but proliferation arrest was then stabilized by additional factor(s). This indicated that certain features of the senescent phenotype‐like ROS production that might be responsible for the negative impact of senescent cells into their tissue environment can be successfully inhibited even in deep senescence. This may prove relevant for novel therapeutic studies aiming to modulate intracellular ROS levels in both aging and cancer.
The sustained activation of CDKN1A (p21/Waf1/Cip1) by a DNA damage response induces mitochondrial dysfunction and reactive oxygen species (ROS) production via signalling through CDKN1A‐GADD45A‐MAPK14‐ GRB2‐TGFBR2‐TGFbeta in senescing primary human and mouse cells in vitro and in vivo.
Enhanced ROS production in senescing cells generates additional DNA damage. Although this damage is repairable and transient, it elevates the average levels of DNA damage response permanently, thus forming a positive feedback loop.
This loop is necessary and sufficient to maintain the stability of growth arrest until a ‘point of no return’ is reached during establishment of senescence.
Cellular senescence is defined by the irreversible loss of division potential of somatic cells and a variety of associated phenotypic changes (Campisi and d'Adda di Fagagna, 2007). Recent interest has been spurred by mounting evidence for major roles for cellular senescence in vivo: on the one hand, oncogene‐triggered senescence is a potentially very powerful tumour suppression mechanism (Ramsey and Sharpless, 2006; Bartek et al, 2007). On the other hand, cellular senescence might contribute to the loss of tissue homeostasis in mammalian aging. There is evidence that senescence‐marker‐positive cells increase with age in various tissues (Dimri et al, 1995; Krishnamurthy et al, 2004; Herbig et al, 2006; Wang et al, 2009) and in age‐related diseases including atherosclerosis (Minamino and Komuro, 2007) and diabetes (Sone and Kagawa, 2005). Although it is not known for how long senescent cells persist in vivo (Ventura et al, 2007; Krizhanovsky et al, 2008), there is a clear evidence that senescent check point activation can contribute to organismal aging (Rudolph et al, 1999; Tyner et al, 2002; Choudhury et al, 2007).
A DNA damage response (DDR), triggered by uncapped telomeres or non‐telomeric DNA damage, is the most prominent initiator of senescence (d'Adda di Fagagna, 2008). This response is characterized by activation of sensor kinases (ATM/ATR, DNA‐PK), formation of DNA damage foci containing activated H2A.X (γH2A.X) and ultimately induction of cell cycle arrest through activation of checkpoint proteins, notably p53 (TP53) and the CDK inhibitor p21 (CDKN1A). This signalling pathway continues to contribute actively to the stability of the G0 arrest in fully senescent cells long after induction of senescence (d'Adda di Fagagna et al, 2003). However, interruption of this pathway is no longer sufficient to rescue growth once the cells have progressed towards an established senescent phenotype (d'Adda di Fagagna et al, 2003; Sang et al, 2008).
Senescence is clearly more complex than CDKI‐mediated growth arrest: senescent cells express hundreds of genes differentially (Shelton et al, 1999), prominent among these being pro‐inflammatory secretory genes (Coppe et al, 2008) and marker genes for a retrograde response induced by mitochondrial dysfunction (Passos et al, 2007a). Recent studies showed that activated chemokine receptor CXCR2 (Acosta et al, 2008), insulin‐like growth factor binding protein 7 (Wajapeyee et al, 2008), IL6 receptor (Kuilman et al, 2008) or downregulation of the transcriptional repressor HES1 (Sang et al, 2008) may be required for the establishment and/or maintenance of the senescent phenotype in various cell types. A signature pro‐inflammatory secretory phenotype takes 7–10 days to develop under DDR (Coppe et al, 2008; Rodier et al, 2009). Together, these data suggest that senescence develops quite slowly from an initiation stage (e.g. DDR‐mediated cell cycle arrest) towards fully irreversible, phenotypically complete senescence. It is the intermediary step(s) that define the establishment of senescence, which are largely unknown with respect to kinetics and governing mechanisms.
Reactive oxygen species (ROS) are likely to be involved in establishment and stabilization of senescence: elevated ROS levels are associated with both replicative (telomere‐dependent) and stress‐ or oncogene‐induced senescence (Saretzki et al, 2003; Ramsey and Sharpless, 2006; Passos et al, 2007a; Lu and Finkel, 2008). ROS accelerate telomere shortening (von Zglinicki, 2002) and can damage DNA directly and thus induce DDR and senescence (Chen et al, 1995; Lu and Finkel, 2008; Rai et al, 2008). Conversely, activation of the major downstream effectors of the DDR/senescence checkpoint can induce ROS production (Polyak et al, 1997; Macip et al, 2002, 2003). Thus, cause–effect relationships between cellular ROS production and senescence are not sufficiently clear.
Here, we analyse the role of ROS for the establishment of senescence initiated by DNA damage or telomere dysfunction and show the existence of a positive feedback loop between DDR and ROS production. We show that telomere‐dependent or ‐independent DDR triggers mitochondrial dysfunction leading to enhanced ROS activation through a linear signal transduction through TP53, CDKN1A, GADD45A, p38 (MAPK14), GRB2, TGFBR2 and TGFβ. ROS contribute in a stochastic manner to the long‐term maintenance of DNA damage foci, and this is necessary and sufficient to maintain proliferation arrest in response to DNA damage during the establishment of an irreversible senescent phenotype. Our results provide experimental evidence that ROS‐dependent signalling is required for the establishment of irreversible senescence of cells with dysfunctional telomeres or damaged DNA in vitro and in vivo. This result might be relevant for therapeutic studies aiming to modulate intracellular ROS levels in both aging and cancer.
Delayed mitochondrial dysfunction and ROS production are a consequence of senescence
To measure the kinetics of ROS induction in senescence, we treated proliferation‐competent human MRC5 fibroblasts with ionizing radiation (IR, 20 Gy). This abolished cell growth and labelling indices for BrdU, Ki67 and caused expression of senescence‐associated β‐galactosidase (Sen‐β‐Gal) to an extent equal to the deep replicative senescence seen when cells had reached their normal proliferative limit (data not shown; Supplementary Figures S1A and B). After IR, DNA damage foci frequencies remained permanently elevated (Supplementary Figures S1C–E) but did not co‐localize with telomeres (Supplementary Figure S1F), together indicating that the cells were driven into stress‐induced premature senescence (SIPS). Importantly, the levels of mitochondrial superoxide (measured as MitoSOX fluorescence intensity in flow cytometry) and cellular peroxides (dihydrorhodamine 123/DHR fluorescence intensity) did not alter immediately after IR but after 24 h the levels of both ROS indicators increased and remained elevated over the entire observation period from day 2 onwards (Figure 1A). This delayed change in ROS indicators was accompanied by a corresponding increase in mitochondrial mass measured by nonyl acridine orange (NAO) fluorescence (Figure 1A) and decreased mitochondrial membrane potential measured by JC‐1 (5,5′,6,6′‐tetrachloro‐1,1′,3,3′‐tetraethylbenzimidazolylcarbocyanide iodide) fluorescence (MMP, Figure 1B). This was further associated with increased transcription of UCP‐2 (Supplementary Figure S1G), which codes for the main uncoupling protein in human fibroblasts. The proton leak‐dependent (oligomycin resistant) oxygen uptake increased about two‐fold after IR‐mediated arrest, similar to deep replicative senescence (Figure 1C), confirming mitochondrial uncoupling as an early event after DDR. Loss of MMP under DDR was also reflected by a diminished ability to maintain cellular [Ca2+]i homeostasis (Supplementary Figure S2).
To see whether ROS production would also be induced in telomere‐dependent senescence, we conditionally overexpressed a dominant‐negative version of the telomere‐binding protein TRF2 (TRF2ΔBΔM) by doxycycline removal (Supplementary Figures S3A–C). This induced purely telomere‐dependent senescence (van Steensel et al, 1998) and resulted in comparable kinetics of appearance of mitochondrial dysfunction and ROS production as well as loss of growth potential and induction of DNA damage foci containing activated H2AX (γH2AX, Figure 1D and E; Supplementary Figures S3D–G). Antioxidant treatment (growth of cells under low ambient oxygen and under treatment with the free radical scavenger PBN) reduced, but did not abolish DNA damage foci induction (Supplementary Figure S3H). Retroviral transduction of TRF2ΔBΔM into primary human MRC5 fibroblasts also induced a similar response (Supplementary Figure S4). Decreased MMP coupled with increased ROS levels is a hallmark of mitochondrial dysfunction that has recently been observed in senescent cells (Passos et al, 2007a). Our data now show that mitochondrial dysfunction is a delayed result of DDR regardless of how this is caused. We reasoned that such elevated ROS production might in turn contribute to DNA damage and DDR, thus forming a positive feedback loop.
Identification of a signalling pathway that induces ROS production and maintains DDR as part of a positive feedback loop
To test this idea and to delineate the signalling pathway between DDR and mitochondrial dysfunction/ROS production, we first modulated TP53 levels in MRC5 human fibroblasts and measured both ROS levels and DNA damage foci frequencies. TP53 overexpression increased cellular ROS levels and DNA damage foci frequencies, whereas siRNA‐mediated knockdown of TP53 before irradiation (Supplementary Figure S5) decreased both the parameters (Figure 2A). Inhibition of CDKN1A, MAPK14 (by siRNA or small molecule inhibitors, Supplementary Figure S6) and TGFβ (by small molecule inhibitor or blocking antibody against TGFBRII) equally reduced ROS and DDR (Figure 2B), showing that the induction of mitochondrial dysfunction and ROS in senescent fibroblasts is not because of a direct interaction of TP53 with the mitochondria, but mediated through CDKN1A and MAPK14/TGFβ. This is in accordance with recent data showing that TP53 is not necessary for the DDR‐dependent induction of senescence‐associated interleukine secretion in fibroblasts (Rodier et al, 2009). There is also published evidence that TGFβ and p38 (MAPK14) activation can induce cellular ROS production through mitochondrial and non‐mitochondrial (NADPHoxidase‐dependent) pathways (Torres and Forman, 2003; Koli et al, 2008). Both have repeatedly been associated with induction of the senescent phenotype (Davis et al, 2005; Debacq‐Chainiaux et al, 2005; Chretien et al, 2008) including DNA damage foci formation (Satyanarayana et al, 2004). However, the signalling hierarchy between TGFβ and MAPK14 with respect to mitochondrial ROS production and the connections, if any, to CDKN1A was not obvious.
To identify candidate pathways connecting CDKN1A and either MAPK14 or TGFβ, we performed an in‐silico interactome analysis. We first searched an interactome database combining Biogrid (Stark et al, 2006) and Phospho.ELM, a protein kinase target database (Diella et al, 2008) for all known paths between CDKN1A and either MAPK14 or TGFβ1/2 having not more than four intermediary nodes. We calculated the log likelihood score (LLS) of each pathway as the average of the LLS for each individual interaction involved. The 20 most probable pathways are shown in Supplementary Figure S7. Note that many of the interactions in Biogrid do not have an inherent directionality, so the resulting network is undirected. Path analysis therefore identified a number of false positive candidate pathways. However, all of the most probable pathways revealed a central position for the CDKN1A target gene GADD45A (Supplementary Figure S7). GADD45A binds directly to CDKN1A (Kearsey et al, 1995) and can activate MAPK14 both indirectly (through MAP3K4 and MAP2K3) and directly (Bulavin et al, 2003).
To analyse experimentally the role of GADD45A and the hierarchy of MAPK14 and TGFβ, we performed further sequential candidate inhibitions, both singly and in various combinations. All treatments reduced ROS formation and DDR by a similar amount, and there was no additive effect for any of the combined treatments (Figure 2C), strongly suggesting that all tested gene products are members of one single epistatic pathway. Calculation of LLS in alternative pathways suggested that the probability for MAPK14 being upstream of TGFβ is about two orders of magnitude higher than for its being downstream (Supplementary Table S1). The most probable pathway according to the interactome analysis is shown in Figure 2D. Overlaying the genes in these pathways with the results of an Affymetrix gene expression analysis of MRC5 fibroblast senescence (Passos et al, 2007a), we noted that the central genes in this pathway tend to be co‐ordinately upregulated in senescent human fibroblasts, resulting in a tight cluster when subjected to unsupervised hierarchical clustering (Supplementary Figure S8). We further confirmed signalling through CDKN1A‐MAPK14‐TGFβ as part of a positive feedback loop combining DDR and ROS production by showing that (i) inhibition of MAPK14 reduced the amount of secreted TGFβ (Supplementary Figure S9A), increased MMP and decreased mitochondrial mass after IR (Supplementary Figure S9B); (ii) inhibition of either MAPK14 or TGFβ or both reduced DNA damage foci containing activated ATM/ATR and 53BP1 (Supplementary Figure S10); (iii) treatment with the MAPK14 inhibitor SB203580 lowered the levels of activated TP53 (p53‐S15), CDKN1A and phosphorylated MAPK14 itself (Supplementary Figure S11); (iv) inhibition of MAPK14, but not of arachidonic acid metabolism, cytochrome P450 or PI3K signalling, specifically diminished the rise in ROS levels in telomere‐dependent senescence (Supplementary Figure S12); (v) inhibition of MAPK14 and TGFβ, alone or in combination, reduced nuclear CDKN1A levels in MRC5 fibroblasts after IR (Supplementary Figure S13A); and (vi) scavenging of ROS reduced DDR foci frequencies and CDKN1A induction after IR (Supplementary Figure S13B).
Together, these data strongly suggested that the DDR and ultimately growth arrest in senescent cells might be maintained by a positive feedback loop between DDR and mitochondrial dysfunction/ROS production through signalling through TP53‐CDKN1A‐GADD45A‐MAPK14‐GRB2‐TGFBRII‐TGFβ (Supplementary Figure 3A).
A stochastic feedback loop model predicts the kinetics of DDR and growth arrest at the single cell level
We quantified the conceptual model shown in Figure 3A to see whether it could sufficiently explain the kinetics of senescence induction and maintenance. To create a stochastic mechanistic model of the DDR feedback loop, we extended our previously published model of the TP53/Mdm2 circuit (Proctor and Gray, 2008) to include reactions for synthesis/activation and degradation/deactivation/repair of CDKN1A, GADD45, MAPK14, ROS and DNA damage (Supplementary Tables S2 and S3). We chose realistic values for reaction rate constants and the initial amounts of the variables (see Supplementary Tables S2 and S3) and ran stochastic simulations for 500 cells initially from 2 days before until 6 days after IR. We parameterized the model using experimental kinetic data for TP53‐S15, CDKN1A and MAPK14 protein levels (Supplementary Figure S11), DNA damage foci frequencies (Supplementary Figure S1E) and ROS levels (Figures 1A and 4A). The model replicated very precisely the kinetic behaviour of activated TP53, CDKN1A, ROS and DNA damage foci after irradiation. In simulations, the key variables stabilized after 2–3 days such that CDKN1A levels were maintained sufficiently above background to generate a stable growth arrest phenotype (Figure 3B). In contrast, a model without feedback would always return in less than 2–3 days to pre‐irradiation levels (Figure 3C).
Having established its concordance with the experimental data, the model was then used to predict the effects of intervening in the feedback loop. Suppression of MAPK14 signalling or antioxidant treatment at day 6 after IR reduced ROS levels by about half (Figure 3B). The model predicted significantly decreased DDR and, importantly, reduced CDKN1A levels to an extent that would allow a fraction of cells to escape from growth arrest (Figure 3B). Thus, the model predicts that a feedback loop is both necessary and sufficient for the stability of growth arrest after induction of senescence by DNA damage.
At higher time resolution, the model predicted a significant decline in DDR foci frequencies in IR‐arrested cells within a few hours after MAPK14 inhibition (Figure 3D). To test this prediction, we used a 53BP1–GFP fusion protein as DDR reporter in live cells. 53BP1 is ubiquitously expressed and homogeneously distributed throughout the nucleus, and redistributes into foci after DNA damage. We fused a large C‐terminal fragment of 53BP1, containing both the TUDOR and BRCT domains necessary for focus formation and interaction with TP53, to GFP, producing an AcGFP–53BP1c fusion protein that quantitatively reported foci dynamics in live cells (Nelson et al, 2009). DNA damage foci frequencies and levels of activated p53, CDKN1A and ROS all stabilized at around 2 days after IR and remained so over the following days (Figures 3B and 4A, B; Supplementary Figure S1E). Therefore, we visualized individual MRC5 fibroblasts stably expressing AcGFP–53BP1c from 3 day till up to 5 day after IR‐mediated arrest. While under observation, the cells were treated with the MAPK14 inhibitor SB203580 at 94 h. Time‐resolved live cell microscopy allowed the qualitative and quantitative identification of individual DNA damage foci (Figure 3E and Supplementary Movie SM1). Foci frequency measurements were in good agreement with static measurements by immunofluorescence (Supplementary Figure S10B) and quantitatively confirmed the prediction from the stochastic model (compare Figure 3D and F).
Closer observation of the time series revealed that the lifespan of many individual foci was much shorter than the observation period (see Supplementary Movie SM1). We ensured proper tracking of individual foci by using a wide z‐stack range (4.5 μm) and a high imaging frequency (every 7 min). Appearance or disappearance, respectively, of foci on at least two consecutive images was used as recording criterion.
We measured the lifespan of individual foci in young proliferating, irradiated and deep senescent MRC5 fibroblasts (Figure 3G). There were few foci in proliferating cells, most of them with lifespans below 5 h, probably caused by replication stress. Foci lifespan increased significantly in senescent cells. However, even in deep telomere‐dependent replicative senescence not all foci became permanent. Rather, about half of all foci were short lived with lifespans below 15 h (Figure 3G). Foci kinetics at >2 day after IR was very similar to that in replicative senescence (Figure 3G). Individual cell traces revealed that the reduction of total foci frequencies under SB203580 treatment was preferentially because of lower numbers of short‐lived foci (Supplementary Figure S14). Using a lifespan of 15 h as cut‐off, we found a significant decrease in the frequency of short‐lived, but not long‐lived foci as a result of SB203580 treatment (Figure 3H). These data show that ROS production, mediated by a delayed signalling through MAPK14, contributes to DDR by constant replenishment of short‐lived DNA damage foci.
The feedback loop between DDR and ROS production is necessary and sufficient to maintain senescent growth arrest during establishment of irreversible senescence
To analyse the relevance of the feedback loop for establishment and maintenance of the senescent phenotype we inhibited signalling through the loop at different time points after induction of SIPS by IR and measured ROS levels, DNA damage foci frequencies and proliferation markers. Treatments with the MAPK14 inhibitor SB203580 or the free radical scavenger PBN were used to block the loop, and a time frame from 30 min until 21 days after IR as well as deep replicative senescence was experimentally covered. Own (Supplementary Figure S1A) and published (Coppe et al, 2008; Rodier et al, 2009) data indicated that a full senescent phenotype develops over 7–10 days in human fibroblasts. Both treatments reduced ROS levels (Figure 4A) and γH2A.X foci frequencies (Figure 4B) equally, even if started weeks after induction of senescence, indicating that the feedback loop remains active in deep senescence. Treatments also enhanced MMP over the whole time frame (data not shown). Treatments that were begun within 1 week after induction of DDR effectively rescued growth arrest (Figure 4C–F). After 20 Gy IR, all fibroblasts were stably negative for the proliferation‐associated antigen Ki67, but this was rescued in up to 30% of the cells by treatment with either SB203580 or PBN, if the treatment was started within up to 9 days after induction of SIPS (Figure 4C). Almost all of the Ki67‐positive cells were negative for expression of CDKN1A (Figure 4D), indicating that these cells were not merely re‐entering G1 but had the potential to progress into the cell cycle. In fact, we observed robust recovery of proliferation in cells treated with SB203580 or PBN at 6 days after SIPS induction in both bulk culture (Figure 4E) and clonal growth (Figure 4F) assays. Treated cells grew at normal proliferation rates for at least several weeks whereas there was no growth at all in untreated cells after 20 Gy IR, indicating that the treatments not simply postponed but actually rescued senescence. Very similar results were obtained if cells irradiated with a lower dose of 5 Gy were treated either at 2 h (Supplementary Figure S15) or at 6 days after irradiation (Supplementary Figure S16). Moreover, post treatment of 5 Gy‐irradiated cells with another antioxidant, N‐acetylcysteine, rescued cell proliferation (data not shown).
We compared the experimentally observed rescue of growth with the predictions from the stochastic model. After single cell traces in the model, we observed that under SB203580 treatment some cells would for some time display CDKN1A values below the basal mean CDKN1A protein level as established in proliferating cells before IR (Supplementary Figure S17). We set a threshold level for CDKN1A in our stochastic model at 2 × standard deviation above basal (see Figure 3B) and assumed that cells would recover from growth arrest if CDKN1A levels would remain continuously below this threshold for at least 6 h (i.e. allowing for passage through the CDKN1A‐dependent G1/S border). Querying 500 individual cell tracks we found 96 cells (19.2%) under SB203580 treatment fulfilling this criterion, whereas none did so after IR alone (Supplementary Figure S16). A comparison with the experimental data (Figure 4C–F) shows that this model quantitatively predicts rescue of growth arrest by feedback loop inhibition.
Together, these results indicate that the feedback loop between DDR and ROS production is necessary and sufficient to maintain cell cycle arrest for at least 1 week after initiation of SIPS. However, at time points later than 9 days after initiation of senescence, the inhibition of feedback signalling became progressively less efficient in rescuing the arrest (Figure 4C), indicating that additional mechanisms not accounted for in our model stabilize growth arrest in deep senescence. Stabilization of the growth arrest in deep senescence might be due partly to gross changes in chromatin organization (Narita et al, 2003). Levels of senescence‐associated heterochromatin foci (SAHF) and HP‐1γ foci, two markers for senescence‐associated chromatin re‐modelling, were low at 6 days after IR but increased over the following weeks in parallel with the irreversibility of growth arrest (Supplementary Figure S18). Treatments of cells with either SB203580 or PBN at a late time point (10 days after IR), which did not effectively rescue growth, also did not change SAHF‐associated nuclear granularity (Supplementary Figure S18C).
DNA damage signalling through CDKN1A results in increased ROS‐mediated damage in vivo
To assess the relevance of feedback loop signalling for generation of oxidative damage in vivo we used primary embryonic fibroblasts (mouse embryonic fibroblasts, MEFs) and tissues from late generation (G4) TERC−/− and G4TERC−/−CDKN1A−/− mice. Loss of telomerase function over multiple generations induces telomere dysfunction triggering widespread DDR in tissues. Knockout of CDKN1A inhibited downstream signalling and partially rescued the shortened lifespan of G4TERC−/− mice (Choudhury et al, 2007). As in human cells, ROS levels increased in wt MEFs at 48 h after IR, and this was completely abolished in CDKN1A−/− MEFs (Figure 5A). Moreover, CDKN1A deletion suppressed the induction of ROS and DDR in telomere‐dependent senescence (i.e. under low ambient oxygen) in G4TERC−/− MEFs (Figure 5B). Interestingly, the co‐localization of telomeres and the remaining foci was closest in the double‐KO MEFs, confirming that loss of signalling through CDKN1A preferentially reduced non‐telomeric foci (Figure 5C). These data confirm the existence of a CDKN1A‐dependent feedback loop signalling in mice cells very similar to that in human cells.
Frequencies of senescent cells displaying DNA damage foci increase with age in various cells and tissues of control mice including the enterocytes in intestinal crypts. We established that these foci‐positive cells were not apoptotic and that very few of them were double positive for γH2A.X and proliferation markers (Wang et al, 2009). We obtained good quantitative agreement between estimates of Sen‐β‐Gal‐positive and γH2A.X‐positive, PCNA‐negative enterocytes, strongly suggesting that these cells are senescent (Lawless et al, 2009). To assess whether this can contribute to an increased oxidative load in vivo, we measured γH2A.X foci formation together with two markers of tissue oxidative damage (broad‐band autofluorescence and 8‐oxodG immunoreactivity) in intestinal crypts from age‐matched control, G4TERC−/− and G4TERC−/−CDKN1A−/− mice (Figure 5D–F). Frequencies of γH2A.X‐positive enterocytes per crypt were significantly increased in G4TERC−/− as compared with wild type or TERC+/− mice (Figure 5D, see also (Choudhury et al, 2007; Wang et al, 2009)). Although loss of CDKN1A only mildly decreased the numbers of crypts showing any γH2A.X positivity (Choudhury et al, 2007), it significantly reduced the frequencies of γH2A.X‐positive cells per crypt (Figure 5D and H). This effect was not related to apoptosis because frequencies of TUNEL‐positive crypt cells were not dependent on CDKN1A (Choudhury et al, 2007).
Broad‐band autofluorescence originates mainly from oxidized and cross‐linked cell components, such as advanced glycation end products and the age pigment lipofuscin and is thus related to oxidative stress (Gerstbrein et al, 2005). Autofluorescence intensity was increased in crypts from G4TERC−/− mice, but this was rescued by loss of CDKN1A (Figure 5E). A similar pattern was seen for oxidative DNA base modification (Figure 5F). Autofluorescence was significantly correlated with 8‐oxoG staining intensity (Figure 5G) and γH2A.X foci frequency on the single crypt level with only minor overlap between genotypes (Figure 5H). These data indicate that DNA damage signalling through CDKN1A contributes in vivo to oxidative damage in crypt cells.
CDKN1A‐dependent ROS production is not restricted to proliferative tissues: Brain neurons suffer from intense DNA damage (Rass et al, 2007) and thus display frequent DNA damage foci (Sedelnikova et al, 2004). The frequency of cerebellar neurons showing a DDR increases in late generation TERC−/− mice, and this increase is partially rescued in G4TERC−/−CDKN1A−/− mice (Supplementary Figure S19A). Autofluorescence intensity and 8‐oxodG positivity in these neurons changed in parallel (Supplementary Figures S19B–D). Similar data were obtained in the cortex and the hippocampus (not shown). These data show that telomere dysfunction induced oxidative damage in vivo through signalling through CDKN1A.
The free radical hypothesis of aging regards ROS‐mediated molecular damage as a potential cause of aging. Although there are conflicting results from intervention studies in vivo (Muller et al, 2007), an impact of ROS levels on senescence in vitro has been well documented (Passos et al, 2007b; Lu and Finkel, 2008). Our results show that mitochondrial dysfunction and ROS production can not only accelerate the onset of senescence (e.g. by accelerating telomere shortening, see (von Zglinicki, 2002)), but are also component and consequence of continuous DDR signalling and are thus an integral part of the senescent phenotype.
Takahashi et al (2006) were the first to convincingly show the existence of an ROS‐generating positive feedback loop. In their model of human fibroblasts expressing a temperature‐sensitive simian virus 40 large T antigen, the loop interconnected the CDKN2A pathway with ROS production and protein kinase C δ signalling (Takahashi et al, 2006). More recently, it was shown that oncogene‐induced senescence can lead to AMPK activation, mitochondrial dysfunction and ROS production, which in turn can trigger senescence on its own (Moiseeva et al, 2009), so suggesting yet another feedback loop involving ROS in senescence. The retinoblastoma protein pRb is an important downstream mediator of growth arrest for both the CDKN1A and CDKN2A pathways (Chang et al, 2000; Dai and Enders, 2000; Beausejour et al, 2003; Takahashi et al, 2006). However, in our system of primary human or mouse fibroblasts neither CDKN2A nor pRb was necessarily involved in the induction of mitochondrial dysfunction and ROS production.
Although we do not rule out the existence of other pathways stabilizing ROS production in senescent cells, we have identified signalling through CDKN1A, MAPK14 and TGFβ as the important link between telomere‐dependent and ‐independent DDR and ROS production in primary human and mice fibroblasts. This identification has been informed by in silico interactome analysis and quantitative stochastic modelling. Interactome analysis allows the integration of large amounts of data from a wide range of sources, even if they are available only in a dispersed set of databases, as is often the case with results of high‐throughput experiments. Possible relationships between genes that do not directly interact can be analysed by standard graph‐theoretic algorithms, such as the shortest‐paths analysis used here. The assignment of LLS to edges (Lee et al, 2004) provided a measure of strength to individual interactions, and permitted ranking of possible paths in order of biological plausibility. Major limitations of interactomes are their limited directionality and their static nature. As first steps to overcome these limitations, we combined Biogrid data with Phospho.ELM, a genome‐wide protein phosphorylation database (Diella et al, 2008) and overlaid the results with a query‐specific set of gene expression data.
A hallmark of aging processes, including cellular senescence, is their inherent stochastic heterogeneity (Kirkwood et al, 2005; Passos et al, 2007a). Therefore, we chose a quantitative model that can account for random processes and explain stochastic variation. This also enabled us to deal appropriately with molecular species present only at very low copy number (e.g. DNA damage foci) and with fractions of cells developing distinct outcomes (e.g. re‐enter proliferation after intervention).
Importantly, we found that ROS levels increase in senescent cells as a result of signalling through CDKN1A‐GADD45A‐MAPK14‐GRB2‐TGFβ and feed back into DNA damage induction and response, generating a stable, self‐sustaining feedback loop. This feedback loop persists even in irreversible deep senescence in vitro and in senescent cells in vivo. During the establishment phase of senescence, the loop is necessary and sufficient for maintenance of replication arrest. Thus, it is a necessary prerequisite for establishment of irreversible senescence.
Our data strongly suggest that during the early establishment phase of senescence a minimum frequency of DNA damage foci is necessary to maintain growth arrest long enough to allow cells to proceed towards irreversibility. This threshold frequency appears to be around five foci per nucleus in our experiments, but might be different in different cells and/or under different experimental conditions. Importantly, our data show that these foci do not have to be permanent themselves to induce permanent growth arrest. Cells might encounter permanent damage (i.e. uncapped telomeres) or high levels of, in principle, repairable damage (e.g. after high doses of IR). In any case, if this situation cannot be resolved within 1–2 days, a cell‐autonomous program is activated, which permanently generates ROS and ROS‐mediated DNA damage. The ensuing stochastic equilibrium of damage induction and repair maintains DNA damage signalling above threshold until further mechanisms (including but not necessarily limited to CDKN2A activation, autocrine signalling and chromatin remodelling) ensure permanent cell cycle arrest. Finding that about half of all foci in senescent cells are short lived, and that short‐lived foci are necessary and sufficient to maintain stability of growth arrest in our model of SIPS does not imply that persistent, irreparable DNA damage signalling might not be sufficient to induce senescence under some conditions. However, it shows that the assumption that DNA damage needs to be irreparable and persistent to generate a persistent proliferation arrest is not necessary. Rather, transient DNA damage can have an important role for senescence, as long as it is constantly replenished by a positive feedback loop that maintains a dynamic equilibrium between damage induction and repair.
An added layer of complexity is that foci composition can change with time. For instance, 53BP1, MRE11 and NBS1 are lost from DNA damage foci as cells enter mitosis, whereas γH2AX and MDC1 remain focally concentrated with chromatin (Nelson et al, 2009; Nakamura et al, 2010). Compositional changes of foci might also occur in tissues during aging (Panda et al, 2008). A better apprehension of foci dynamics will aid understanding of cellular damage responses and aging.
Oxygen levels in standard cell culture are unphysiologically high, and that causes increased ROS production at least partially from the mitochondria leading to acceleration of telomere‐dependent senescence (von Zglinicki, 2002; Passos et al, 2007a) and/or SIPS (Parrinello et al, 2003). This does not imply that the induction of an ROS‐generating feedback loop in senescence as described here is a cell culture artefact; however, ROS levels and DDR are still regulated in the same p21‐dependent manner in MEFs under physiologically low ambient oxygen concentration (see Figure 5). Moreover, DDR and oxidative stress are closely associated during telomere‐independent cell senescence in aging mice (Wang et al, 2009), and this association is dependent on p21 (this study). Telomere‐dependent senescence can be postponed by lowering mitochondrial ROS production and/or release (Saretzki et al, 2003; Passos et al, 2007a), which might be due to both slower telomere shortening and decreased levels of non‐telomeric DDR. Finally, recent data indicate that mitochondrial dysfunction and ROS production is triggered through p53‐ and pRb‐dependent pathways in oncogenic ras‐induced senescence and is able to contribute to ras‐dependent growth arrest (Moiseeva et al, 2009). Together, these data suggest that a feedback loop involving mitochondrial dysfunction and ROS production might well be important in various physiologically relevant forms of cell senescence.
We speculate that mitochondrial dysfunction and ROS production might be causal for the development of the senescent phenotype. Mitochondrial dysfunction including ROS production induces retrograde response, a major re‐programming of nuclear gene expression patterns, in senescent cells (Butow and Avadhani, 2004; Passos et al, 2007a, 2007c). Recent work implicated not only ROS (Bartek et al, 2007) but also factors involved in growth factor, chemokine and cytokine signalling as essential for oncogene‐induced senescence (Acosta et al, 2008; Kuilman et al, 2008; Wajapeyee et al, 2008; Kuilman and Peeper, 2009). Many gene products from these families take part in retrograde response (Butow and Avadhani, 2004), and they interact with TP53 and MAPK pathways (Acosta et al, 2008). A detailed examination of the role of mitochondrial dysfunction for the establishment of the secretory senescent phenotype is clearly warranted.
Although senescent cells can be cleared away efficiently from tissues under some conditions (Ventura et al, 2007; Krizhanovsky et al, 2008), cells bearing various senescence markers including DNA damage foci do accumulate in human (Dimri et al, 1995), primate (Herbig et al, 2006) and mouse (Wang et al, 2009) tissues with advancing age. The fact that DNA damage foci are associated with ROS production in vivo and in vitro (this paper) suggests that tissue‐resident cells with an activated DDR may disturb tissue function and homeostasis not only by secreting biologically active peptides (Campisi and d'Adda di Fagagna, 2007; Coppe et al, 2008) but also by inducing ROS‐mediated damage in their microenvironment. H2O2 is the main product of mitochondrial, cytoplasmic and extracellular superoxide dismutation. It is soluble in both water and lipid and is a relatively long‐lived ROS, allowing for easy diffusion between cells. Thus, H2O2 release from cells with activated DDR might contribute to the ‘bystander effect’, whereby senescent cells appear to ‘infect’ their originally unstressed neighbours(Sokolov et al, 2007).
Importantly, our results show that interventions against individual components of the senescent phenotype (e.g. ROS production) are possible in deep senescence. Even if these interventions do not rescue the growth arrest phenotype, they can ameliorate the phenotypic impact of senescent cells onto their microenvironment (the ‘bystander effect’). There is a preliminary evidence that ‘anti‐senescence’ interventions have potential to delay aging. However, interventions upstream of the cell cycle checkpoint machinery may significantly increase cancer risk (Baker et al, 2008; Tomas‐Loba et al, 2008). Improved understanding of the networks of signalling pathways that govern the senescent phenotype will undoubtedly help to identify the most promising targets to ameliorate the negative impact of aged cells on their environment.
Materials and methods
Cells and tissues
T19 cells containing a doxycycline inducible TRF2ΔBΔM and the retroviral expression vector pLPCNMyc‐TRF2 ΔBΔM were gifts from T de Lange, Rockefeller University, NY (van Steensel et al, 1998). MRC‐5 human embryonic lung fibroblasts (PD 38) were infected with pLPCNMyc‐TRF2 ΔBΔM retrovirus. Expression of TRF2ΔBΔM was monitored using an anti‐FLAG M2 mouse monoclonal antibody (Sigma) in T19 cells or the anti‐Myc antibody Myc‐FITC (#46‐0307, Invitrogen) in infected MRC5. MEFs were grown under 3% ambient oxygen.
Transgenic mice were generated and MEFs as well as brain and gut sections from 12–15‐months‐old mice were prepared as described (Choudhury et al, 2007). Experiments were approved by the local ethics committee.
MRC‐5 cells at PD 25‐30 were transiently transfected with either control untargeted siRNA, TP53 siRNA (SignalSilence® p53 siRNA kit, Cell Signalling Technology), CDKN1A siRNA (SignalSilence® p21 Waf1/Cip1 siRNA human specific), MAPK14 siRNA (SignalSilence® pool p38 MAPK siRNA) or GADD45A siRNA (GADD45A Validated StealthTM DuoPak, Invitrogen) using nucleofection (Amaxa) according to the supplier's protocols.
MAPK14 activity was inhibited using 10 μM SB202190 (Sigma) or SB203580 (Tocris Bioscience). Inhibition was confirmed by western blot using anti‐phospho‐p38 MAPK (mouse monoclonal, Cell Signalling) and anti‐p38 MAPK (rabbit polyclonal, Cell Signalling). Inhibition of the TGFβ pathway was performed using TGFβR2 antibody (rabbit polyclonal, Cell Signalling) or 10 μM SB431542 (Tocris Bioscience). Secreted human TGFβ1 was measured using Quantikine® Human TGF‐β1 Immunoassay (#DB100B, R&D Systems).
Transmission electron microscopy was performed using conventional techniques. All fluorescence microscopy was performed in confocal mode (Zeiss LSM510). Foci frequencies and fluorescence intensities were always measured under identical excitation and emission conditions using a 63 × (NA=1.4) objective and a 1 Airy unit pinhole. For broad‐band autofluorescence of 5 μm tissue sections a 20 × (NA=0.5) objective with a pinhole equivalent to an 11 μm z‐slice was used. The sample was excited at 458 nm and fluorescence emission captured above 475 nm, while simultaneously capturing a transmission image. Fluorescence intensities per cell or per crypt were quantified in ImageJ (http://rsb.info.nih.gov/ij/). MMP in live cells was measured by FACS (Passos et al, 2007a) or in a LSM510 equipped with a Solent incubator (Solent Scientific) at 37°C with humidified 5% CO2, using a 63 × (NA=1.4) objective; 100 nM Mitotracker Green and 16 nM TMRM (tetramethylrhodamine methyl ester, Invitrogen) were loaded in serum‐free medium for 30 min at 37°C. Mitotracker Green fluorescence (proportional to total mitochondrial mass) and TMRM fluorescence (proportional to MMP) were captured using 488 and 543 nm excitation and 505–530 nm bandpass or 560 nm longpass emission, respectively. Images of the entire cell depth were acquired using a 1.4 μm z‐slice every 400 nm as z‐stacks, deconvolved using Huygens (SVI) and rendered as surface projections using the same parameters for object size cut‐off and intensity between different treatments.
To obtain quantitative estimates of SAHF formation, the mean gradient amplitude was calculated from DAPI‐stained images in the masked region of interest and used as a measure of granularity using imageJ plugin http://bigwww.epfl.ch/thevenaz/differentials/ (Unser, 1999). Mean gradient amplitude was divided by mean DAPI intensity.
For live cell timelapse microscopy, cells were plated in Iwaki glass bottomed dishes (Iwaki) and imaged on an inverted Zeiss LSM510 equipped with a Solent incubator (Solent Scientific) at 37°C with humidified 5% CO2, using a 40 × 1.3 NA oil objective, as described (Nelson et al, 2002). Autofocus was performed at each timepoint before capturing a z‐stack to ensure the entire cell was captured (1.65 μm pinhole every 1.5 μm over 4.5 μm total z range). Cells and AcGFP–53BP1c foci were tracked manually using ImageJ (http://rsb.info.nih.gov/ij/).
Bioinformatics and quantitative stochastic modelling
A probabilistic functional integrated network of interactions was constructed using gene and protein interaction data from the BioGrid database (Stark et al, 2006), plus protein phosphorylation data from Phospho.ELM (Diella et al, 2008). To assess interaction likelihoods, an LLS was calculated for each dataset as described (Lee et al, 2004). Datasets larger than 100 interactions were analysed individually, whereas those with fewer than 100 interactions were grouped by evidence category. Pathway data, from the Kyoto Encyclopedia of Genes and Genomes (KEGG) (Kanehisa and Goto, 2000) Release 46.0, 1 April 2008, was used as the gold standard. The final LLS for an interaction between a pair of genes was calculated as the sum over the LLS of all the datasets containing that interaction.
Network analysis was performed using the Cytoscape platform (Shannon et al, 2003). An in‐house Python script was used to detect all paths between CDKN1A and either MAPK14 or TGFβ1/2 having no more than four intermediary nodes.
For cluster analysis of candidate pathway genes, raw MRC5 microarray data (Passos et al, 2007a) (four young confluent and five senescent cultures) were loaded into Bioconductor (http://www.bioconductor.org) and normalized using GCRMA method. Hierarchical cluster analysis was applied to the expression values relative to the mean of all arrays. The Pearson correlation was used as similarity measure and average linkage as cluster method.
To test whether the feedback loop between DDR and ROS production was necessary to explain the experimental data, we developed a stochastic mechanistic model of the DDR extending our previously published model of the TP53/Mdm2 circuit (Proctor and Gray, 2008) by the steps outlined in Figure 5A. Model variables, reactions, kinetic laws and parameter values are given in Supplementary Tables S2 and S3. The model is encoded in the Systems Biology Markup Language (Hucka et al, 2003). Simulations are run in the Biology of Ageing e‐Science Integration and Simulation (BASIS) system (http://www.basis.ncl.ac.uk; Kirkwood et al, 2003; Gillespie et al, 2006) using stochastic simulation based on the exact Gillespie algorithm (Gillespie, 1977). The model is available in the public space of the BASIS database (urn:basis.ncl:basis:6178) and is also obtainable from Biomodels (MODEL 5989624192), http://www.ebi.ac.uk/biomodels/ (Le Novere et al, 2006).
Further experimental procedures
Flow cytometric and confocal imaging techniques to measure mitochondrial superoxide, cell peroxides, mitochondrial mass, cytoplasmic calcium, γH2A.X. immunofluorescence, telomere immunoFISH, RT–PCR and sen‐β‐Gal staining were described (Passos et al, 2007a). All flow cytometry was performed on at least 30 000 cells per measurement. Anti‐p21 mouse monoclonal IgG (U2OS, Calbiochem Inc.), anti‐p16 rabbit polyclonal IgG (C20, Santa Cruz Biotechnology Inc.), anti‐p38 MAPK rabbit polyclonal (Cell signalling) were used for immunofluorescence, and anti‐γ‐H2A.X rabbit monoclonal (Cell Signalling) and mouse monoclonal 8‐oxodG (Japan Institute, with mouse‐on‐mouse kit) were used for immunohistochemistry with vectastain ABC kit (PK‐6101, Vector).
Cellular oxygen uptake was determined using high‐resolution respirometry (Oxygraph‐2K; Oroboros Instruments, Insbruck, Austria). The measurements were taken at 37°C. After the recording of the rate of steady‐state basal oxygen uptake, the following chemicals were sequentially added: oligomycin (1 μg/ml), carbonyl cyanide p‐(trifluoromethoxy) phenylhydrazone (FCCP, 3–4.5 μM, depending on cell state), rotenone (0.5 μM) and antimycin A (2.5 μM). Antimycin‐resistant oxygen uptake was registered as non‐mitochondrial oxygen uptake and subtracted from the raw data to calculate the mitochondrial oxygen uptake.
Group means were compared by Mann–Whitney rank sum test, if data were not normally distributed. Multiple comparisons were performed by ANOVA followed by Tukey's test for comparison of individual subgroups. γH2A.X‐telomere co‐localization was tested by pixelwise Pearson correlation analysis as described (Passos et al, 2007a).
We thank Drs G Lei, A Tsolou, A Oakley and Mrs M Maddick and M‐C Fawcett for technical help and T de Lange, Rockefeller University, NY, for the TRF2ΔBΔM cells and the pLPCNMyc‐TRF2ΔBΔM expression vector. The study was supported by grants from BBSRC/EPSRC (CISBAN) to TK and TvZ and from Research into Ageing UK to TvZ. JP was partially supported by the Fundação para a Ciência e Tecnologia through the GABBA Programme, University of Porto, Porto, Portugal. CP was funded on a Fellowship from the Alzheimer Scotland and the Alzheimer's Research Trust.
Conflict of Interest
The authors declare that they have no conflict of interest.
Supplementary Movie SM1
Supplementary Figures S1–19, Supplementary Tables S1–3
This is an open‐access article distributed under the terms of the Creative Commons Attribution License, which permits distribution, and reproduction in any medium, provided the original author and source are credited. This license does not permit commercial exploitation without specific permission.
- Copyright © 2010 EMBO and Macmillan Publishers Limited