Abstract
Induced pluripotent stem cells (IPSCs), with their unlimited regenerative capacity, carry the promise for tissue replacement to counter age-related decline. However, attempts to realise in vivo iPSC have invariably resulted in the formation of teratomas. Partial reprogramming in prematurely aged mice has shown promising results in alleviating age-related symptoms without teratoma formation. Does partial reprogramming lead to rejuvenation (i.e. “younger” cells), rather than dedifferentiation, which bears the risk of cancer? Here we analyse cellular age during iPSC reprogramming and find that partial reprogramming leads to a reduction in the biological age of cells. We also find that the loss of somatic gene expression and epigenetic age follow different kinetics, suggesting that rejuvenation can be achieved with a minimised risk of cancer.
Introduction
The human ageing process is accompanied by multiple degenerative diseases. Our understanding of such ageing related disorders is, nevertheless, fragmented, and the existence and nature of a general underlying cause are still much debated (Faragher 2015; Gladyshev & Gladyshev 2016). A breakthrough technique, the generation of induced pluripotent stem cells (iPSCs), allows the reprogramming of somatic cells back to an embryonic stem cell (ESC) like state with an unlimited regenerative capacity. This has led to multiple strategies for tissue replacement in degenerative diseases (Takahashi et al. 2007). Clinical application of iPSCs however, is at its infancy (Takahashi & Yamanaka 2016; Singh et al. 2015; Soria-Valles et al. 2015), and the potency of iPSCs bears risks, not least cancer induction. For example, in vivo experiments with iPSCs have shown that continuous expression of Yamanaka factors (Oct4, Sox2, Klf4 and c-Myc, thus OSKM) in adult mice invariably leads to cancer (Abad et al. 2013; Ohnishi et al. 2014).
To avoid this risk, a parallel concept of epigenetic rejuvenation has been proposed: the ageing process in cells can be reversed whilst avoiding dedifferentiation (Singh & Zacouto 2010). In other words, an old dysfunctional heart cell could be rejuvenated without the need for it to be passed through an embryonic/iPSC state. The concept of epigenetic rejuvenation requires that rejuvenation and dedifferentiation each follow a distinct pathway. Nevertheless, it is not well understood whether rejuvenation and dedifferentiation are invariably intertwined, or instead whether it is possible to manipulate age without risking dedifferentiation.
Ocampo et al. demonstrated that partial reprogramming by transient cyclic induction of OSKM ameliorates signs of ageing and extends lifespan in progeroid mice, with no resulting teratoma formation (Ocampo et al. 2016). Ocampo et al. thus established partial reprogramming as a promising candidate intervention for age-related disease. However, accurately determining biological age in mice was not possible at the time of the Ocampo study. Therefore, the nature (i.e. dedifferentiation/rejuvenation) of the cellular changes remain unexplored:
Does the epigenetic remodelling seen truly reflect rejuvenation (i.e. a reduction in cellular/tissue age)? If so, can we observe rejuvenation in human cells?
What is the extent of the rejuvenation after a cycle of partial reprogramming (e.g. years/cyclic induction)?
Are there signs of dedifferentiation in early reprogramming?
A major obstacle in understanding the relation between differentiation and ageing has been our inability to accurately measure cellular age with a high correlation to the chronological age of the organism. However, over the last five years a number of age predictors have been developed, the most accurate of which utilise DNA methylation (known as epigenetic clocks) (Horvath 2013; Hannum et al. 2013), with the “Horvath clock” being the most widely used (r=0.96). The Horvath clock predicts the age (or epigenetic age, eAge) of multiple tissues with a median error of 3.6 years (Horvath 2013). A predicted eAge older than the chronological age (“epigenetic age acceleration”) is associated with a higher risk of all-cause mortality (Marioni et al. 2015; Christiansen et al. 2016; Perna et al. 2016), premature ageing syndromes (Down and Werner) (Maierhofer et al. 2017; Horvath et al. 2015), frailty and menopause (Breitling et al. 2016; Levine et al. 2016). Epigenetic age is distinct from other biomarkers, such as senescence and telomere length (Lowe et al. 2016). All of these studies suggest that eAge truly measures biological age.
Does partial reprogramming lead to a reduction in cellular age? Calculating eAge dynamics over the course of iPSC reprogramming of human dermal fibroblasts (HDFs) allows us to address this question. We observe onset of a continuous decline of eAge after day 3 from induction with OSKM. Our results suggest that partial reprogramming leads to a reduction in the eAge of cells and is therefore a rejuvenation mechanism. Comparing eAge decline to loss of somatic gene expression indicates a window within which rejuvenation might occur uncoupled from dedifferentiation.
Results
Epigenetic age shows a steady decline in early reprogramming
To understand the dynamics of eAge within a reprogramming time-course, we calculated eAge using Horvath’s multi-tissue age predictor over a previously published 49-day reprogramming time-course on HDFs (Ohnuki et al. 2014; Horvath 2013). Epigenetic rejuvenation, i.e. decrease of eAge, commenced between days 3 and 7 after OSKM transduction and continued until day 20, when it was stably reset to zero (Fig. 1a). A broken stick model with two linear sections starting from day 3 showed a good fit to the observed data and measured a steady decrease with 3.8 years per day until day 20 (SE 0.27, P = 3.8×10−7) (Fig. 1a). Our data suggest that partial reprogramming does indeed result in a reduction of eAge in human cells and can be considered a rejuvenation mechanism.
Partial reprogramming in Ocampo et al. was achieved after just two days of OKSM induction in mice carrying an inducible OSKM transgene. However, kinetics are likely to be different with in vitro fibroblasts where OSKM is induced by viral transduction. To associate the eAge with intermediate states in the reprogramming trajectory we compared it to gene expression measured in the same samples. We analysed corresponding microarray gene expression data for 19 well-established pluripotency genes (Table 1 and Supplementary fig.1) as a proxy for reaching a mature pluripotent state (Ginis et al. 2004; Cai et al. 2006; Mallon et al. 2013; Galan et al. 2013; Boyer et al. 2005). We clustered expression patterns of those genes (Genolini et al. 2015), which resulted in two composite trajectories. These followed previously described expression dynamics of early (cluster 1) and late (cluster 2) activated pluripotency genes (Fig. 1a) (Tanabe et al. 2013; Chung et al. 2014; Buganim et al. 2012; Takahashi & Yamanaka 2016). Pluripotency gene cluster 1 included NANOG, SALL4, ZFP42, TRA-1-60, UTF1, DPPA4 and LEFTY2, and their expression increased dramatically within the first 10 days and then established stable pluripotency expression levels by day 20. In contrast, pluripotency gene cluster 2 (containing late expressing genes such as LIN28, ZIC3 and DNMT3B) elevated expression more slowly and reached stable pluripotency levels by day 30 (Tanabe et al. 2013; Chung et al. 2014). eAge is reset to zero at the same time that the genes in cluster 1 reach their pluripotent state levels, which temporally precedes completion of the full iPSC time course. In summary, eAge decline is observed well within the first wave of pluripotency gene expression.
Loss of somatic gene expression is uncoupled from rejuvenation dynamics and occurs step-wise
Therapeutic partial reprogramming will depend on rejuvenation with minimal dedifferentiation, which carries the risk of malignancies. Here we studied the dynamics of fibroblast gene down-regulation as a proxy for the loss of somatic cell identity. The individual trajectories of commonly used fibroblast marker genes (Kalluri & Zeisberg 2006; Zhou et al. 2016; Janmaat et al. 2015; Pilling et al. 2009; Chang et al. 2014; Goodpaster et al. 2008; MacFadyen et al. 2005) (Table 1 and Supplementary Fig. 2) clustered into three composite expression patterns, two of which (clusters 2 and 3) went into an immediate decline after OSKM induction (Fig. 1b). However, one fibroblast-specific cluster (cluster 1) remained stable in its expression for the first 15 days. Interestingly, after day 7, fibroblast-specific gene expression in clusters 2 and 3 stopped declining and plateaued until day 15. Vimentin (VIM), for example, remained at 60% of maximal expression until day 15 of reprogramming, similarly to FAP, CD248 and COL1A2 in cluster 2. After day 15, fibroblast gene expression declined rapidly in all three clusters, and only by day 35 had all reached ESC expression levels, marking a complete loss of somatic identity. Of the three fibroblast gene clusters, cluster 1 showed the slowest decline, and was also the last to reach ESC expression levels. This cluster contains FSP1, COL3A1 and TGFB2/3 (Supplementary fig. 2), which are well described indicators of fibroblast identity (Kalluri & Zeisberg 2006). In summary, we found several fibroblast specific genes (cluster 1) that maintained fibroblast expression levels until day 15 (by which time a significant drop in eAge has been observed), and the pattern of the decline observed in all three fibroblast clusters indicates that loss of tissue specific expression occurs in a step-wise manner.
Discussion
Ground-breaking work by Ocampo et al. showed that partial reprogramming can alleviate age-related pathologies in prematurely ageing mice, highlighting it as a rejuvenation strategy to counteract age-related disease (Ocampo et al. 2016). The authors suspect that epigenetic rejuvenation is the driver behind the improved age-associated phenotype both in their in vivo and in vitro experiments. Epigenetic rejuvenation, reversal of cellular age, is a promising concept as it could avoid the oncogenic risks associated with dedifferentiation. However, to determine whether partial reprogramming indeed leads to epigenetic rejuvenation (or other cellular changes that might improve age-related pathologies) requires an accurate measure of biological age. This is currently only feasible through profiling DNA methylation, which remained unexplored in Ocampo et al. Here, analysing a reprogramming time-course on HDFs, we show that eAge indeed declines early in reprogramming suggesting that the improvements Ocampo et al. observed might be due to epigenetic rejuvenation.
A deep understanding of the kinetics of rejuvenation will be required to master therapeutic partial reprogramming, since dedifferentiation carries the risk of malignancies. We analysed the dynamics of rejuvenation along the iPSC time-course and compared it to fibroblast specific gene expression as a proxy for dedifferentiation. Within the iPSC reprogramming time-course, partial reprogramming happens within an early, reversible phase involving stochastic activation of pluripotency genes. It is followed by a more deterministic maturation phase with predictable order of gene expression changes, where cell fate is firmly bound towards pluripotency (Takahashi & Yamanaka 2016; Smith et al. 2016). For example, Tanabe et al. showed that ~50% of cells with TRA-1-60 expression, a marker of human ESC/iPSC cells, spontaneously revert to a TRA-1-60 negative state at reprogramming days 7 and 11 (Tanabe et al. 2013). Transcriptome clustering in Tanabe et al. also suggests that a reversion for 9 days within the uncommitted phase places the cells closer to the original fibroblast than day 4 OSKM (Tanabe et al. 2013). Indeed, it has been shown that mouse fibroblasts fail to become iPSC and revert to their original state if OSKM expression is discontinued during the initial stochastic phase (Brambrink et al. 2008; Stadtfeld et al. 2008). Our data suggest a window of opportunity (3-7 days in the iPSC experiment) within the uncommitted reprogramming phase where a pronounced decline of eAge happens alongside a partial maintenance of fibroblast gene expression (Fig. 1b) (Tanabe et al. 2013). Further experiments are needed to determine the safe rejuvenation boundary in different reprogramming systems.
Different dynamics between the step-wise decline in fibroblast expression and the linear decline in eAge further indicate that dedifferentiation and epigenetic rejuvenation can be uncoupled. It remains to be shown how stable the rejuvenated phenotype is. How long or how often would cyclic induction of OSKM need to be conducted to stabilise the young phenotype? Further analysis is also needed regarding the effect of partial reprogramming on adult stem cells or premalignant cells, which have already shown a higher propensity of transforming to malignancy (Abad et al. 2013; Ohnishi et al. 2014). It is possible that a premalignant phenotype could be attenuated or amplified by partial reprogramming.
Methods
Overview of the Ohnuki et al dataset
450K DNA methylation array and gene expression microarray data of full HDF reprogramming time-course was obtained from GSE54848. Microarray data (LOG2 transformed) was available for three to four replicates per data point, whilst DNA methylation data was available for three replicates.
Predicting eAge
For each time point, methylation data for 26987 CpG sites was uploaded to the online DNA methylation age calculator to assess eAge: https://labs.genetics.ucla.edu/horvath/dnamage/ (Horvath 2013).
Methylation Age Trajectories
A ‘broken stick’ model with two linear sections was constructed to chart overall change in DNA methylation age over time between the three HDF cell lines. A linear mixed model was specified with a random intercept term for each replicate. A variable break point was set between the minimum and maximum day, plus and minus a small constant (3 days), respectively. The predicted values from the regression models were plotted against the measurement day.
Gene clusters and trajectories
For each gene in a category (e.g. pluripotent gene list), a loess curve with a span of 0.5 was fitted with the predicted values extracted at each time point. The predicted values were then normalised within each gene to a value of 1 at the first time point and a value of 0 and the last time point (and vice versa for the pluripotent genes). K-means clustering for longitudinal data was applied to determine the optimal number of trajectories within each gene category.
All analyses were performed in R, using the kml (Genolini et al. 2015), lme4 (Bates et al. 2014), and lmerTest (Kuznetsova et al. 2016) packages.