Shinichi Nakagawa, National Research Centre for Growth and Development, Department of Zoology, University of Otago, PO Box 56, Dunedin 9054, New Zealand. Tel.: +64 3 479 5046; fax: +64 3 479 7584; e-mail:email@example.com
Dietary restriction (DR) extends the lifespan of a wide range of species, although the universality of this effect has never been quantitatively examined. Here, we report the first comprehensive comparative meta-analysis of DR across studies and species. Overall, DR significantly increased lifespan, but this effect is modulated by several factors. In general, DR has less effect in extending lifespan in males and also in non-model organisms. Surprisingly, the proportion of protein intake was more important for life extension via DR than the degree of caloric restriction. Furthermore, we show that reduction in both age-dependent and age-independent mortality rates drives life extension by DR among the well-studied laboratory model species (yeast, nematode worms, fruit flies and rodents). Our results suggest that convergent adaptation to laboratory conditions better explains the observed DR–longevity relationship than evolutionary conservation although alternative explanations are possible.
It is generally accepted that dietary restriction (DR, a reduction in food intake without malnutrition) has health benefits such as prolonging lifespan and protection from various diseases (diabetes, cancer and cardiovascular disease; Fontana et al., 2010; Partridge, 2010; Piper et al., 2011). Such benefits have been demonstrated in a wide diversity of species, across several animal phyla (Mair & Dillin, 2008). Much of the evidence concerning DR comes from five laboratory model species (Le Bourg, 2010): yeast (Sacchromyces cerevisiae), the nematode (Caenorhabditis elegans), the fruit fly (Drosophila melanogaster), the mouse (Mus musculus) and the rat (Rattus norvegicus), in which the effects of DR were first reported in 1935 (McCay et al., 1935; hereafter, we will refer to these five species as the ‘model species’). The phylogenetic diversity represented by these five species has underpinned a widely held belief in an evolutionarily conserved mechanism that mediates the relationship between DR and longevity (Le Bourg & Rattan, 2006; Mair & Dillin, 2008; Le Bourg, 2010). A recent report on the effect of DR on lifespan in long-lived rhesus monkeys (Macaca mulatta; Colman et al., 2009) reinforces this view, raising the hope that DR has the same beneficial effects for our own species (Fontana et al., 2010; Partridge, 2010). Nevertheless, the beneficial effects of DR do not appear to be universal: studies of houseflies and several species of rotifers (Kirk, 2001; Cooper et al., 2004), for example, failed to detect any life-extending effect of DR. Such studies call into question the possibility of any evolutionarily conserved mechanism.
Evolutionary explanations for DR effects on lifespan invoke a life history trade-off. Among and within species, fecundity negatively correlates with longevity, the so-called cost of reproduction (Williams, 1966). Dietary restriction results in increased longevity, but also in decreased fecundity, at least in laboratory animals (Partridge et al., 2005a). In the wild, many organisms encounter periods of starvation during which they should devote energy to somatic maintenance and repair, prolonging lifespan to survive until a nutritionally richer period when they can afford to reproduce (Kirkwood & Shanley, 2005). Findings on sex differences in the effects of DR support this cost-of-reproduction view. In Drosophila and several strains of mice, the life-prolonging effect is much more pronounced in females than in males (Partridge et al., 2005a). Such a sex difference is expected, because males are generally thought to invest less into reproduction.
Studies of single-gene mutations in the model species are now elucidating some of the molecular pathways of DR effects (Bartke, 2011). Mutations that prolong lifespan are usually involved in slowing down nutrient-signalling pathways. For example, down-regulation of the target of rapamycin (TOR) pathway extends lifespan in yeast, C. elegans, drosophila and mice (McCormick et al., 2011). Also, reduced activity in the insulin/insulin-like growth factor signalling (IIS) pathway leads to longer lifespan in C. elegans, Drosophila and mice (Kenyon, 2011). Dietary restriction is thought to influence one or more of these nutrient-signalling pathways, although how exactly DR acts on these pathways is still poorly understood (Fontana et al., 2010; Partridge, 2010). The mutation studies have been viewed as strong evidence against the DR effect arising from convergent adaptation and strong support for its evolutionary conservation (Mair & Dillin, 2008). Nevertheless, the unequivocal demonstration of the life extension by DR in the five model species does not prove the universality or conservation of the DR–longevity relationship (Le Bourg & Rattan, 2006; Le Bourg, 2010). For many generations, these model species have lived under laboratory conditions, which usually provide constant food supply, are free from pathogens, parasites and predators and select for fecundity and appetite but against longevity (Miller et al., 2002; Austad & Kristan, 2003). Thus, some researchers have controversially speculated that the effect of DR could be a laboratory artefact, for example, by alleviating the detrimental consequences of overfeeding (Le Bourg, 2010).
A second debate concerning DR studies is whether restriction of caloric intake per se can extend longevity (Masoro, 2006). DR is often referred to as ‘caloric restriction’ because a reduction in calories was believed to be the key factor prolonging an organism’s lifespan. A series of recent studies, however, suggests that the balance between macronutrients (the ratio between proteins and carbohydrate/fat) is more important than caloric restriction (Simpson & Raubenheimer, 2009). In several insect species (crickets and drosophila), fixed-calorie diets containing lower ratios of protein to carbohydrates (‘protein restriction’) extended longevity (e.g. Mair et al., 2005; Lee et al., 2008; Maklakov et al., 2008; Fanson et al., 2009). Furthermore, a study on drosophila suggests that adequate ratios of amino acids within protein intake are key for lifespan extension via DR (Grandison et al., 2009), but there is limited experimental evidence as to whether such clear effects of protein restriction apply to mammalian species. Some researchers argue that caloric restriction is crucial for the mammalian life-extending effect of DR (Masoro, 2006), although there is indirect support for the effect of protein restriction in humans (Fontana et al., 2008). It seems that both caloric and protein restriction may play a role in producing life extension by DR.
Here, we attempt to resolve these two debates by conducting a comprehensive and comparative meta-analysis on a wealth of published studies, investigating the relationship between DR and survival. Importantly, the comparative meta-analytic approach enabled us to combine a wide variety of species from a great number of studies and to extract general trend from what appears to be contradictory results while controlling for species-specific and study-specific effects (Hadfield & Nakagawa, 2010). Our main aims are the followings: (i) to determine the universality of DR effects on longevity between sexes and among species, especially focusing on the model and non-model species, and (ii) to quantify the importance of both caloric restriction and protein restriction in the effect of DR on longevity. Additionally, in the model species, we investigate whether DR affects either age-independent or age-dependent mortality rate or both (Partridge et al., 2005b; Phelan & Rose, 2005). An overall objective of this study is to quantitatively synthesize the current state of knowledge on this important topic for the first time, and thus, to present an overview of the empirical evidence.
Results and discussion
Universality of life-extending effect of DR
We located 145 studies investigating the relationship between DR and longevity in 36 species, which matched our selection criteria (see Experimental procedures). We extracted 529 effect sizes from these studies (Data S1); the effect size measure used is the natural logarithms of hazard ratio, ln(HR) (Table S1). In short, a set of three ln(HR) values were extracted from each pair of survival curves (consisting of the control group and the DR group) at three relative time intervals (during which 0–25%, 25–50% and 50–75% of the control group died), and the overall estimates from these three values constituted effect size values as ln(HR) (Fig. S1; Parmar et al., 1998; Williamson et al., 2002). Negative ln(HR) values mean that individuals in DR groups were less likely to die at a given point on average than ones in the control groups.
Overall, DR reduced the risk of death by 60% (Bayesian mixed-effects meta-analysis, BMM; Hadfield, 2010; Hadfield & Nakagawa, 2010; β[meta-analytic mean] = −0.434, 95% credible interval (CI) = −0.704 to −0.171; Table S2). This effect remains robust even when phylogenetic non-independence among 36 species was accounted for (Bayesian phylogenetic mixed-effects meta-analysis, BPMM: β[meta-analytic mean] = −0.515, CI = −0.953 to −0.093; Figs 1 and S2, Table S1 and Data S2). We observed moderate to high heterogeneity (Higgins & Thompson, 2002; BPMM: I2 = 53.73, CI = 41.15–66.00; Table S2; hereafter, results only from BPMM are presented, see Table S2–S6 for equivalent results from BMM); that is, the life-extending effect of DR is more apparent in certain species and/or studies.
In meta-analysis, significant heterogeneity calls for moderators (e.g. the effect of sex), which may explain such heterogeneity (Higgins & Thompson, 2002). Thus, we tested the controversial suggestion that the life-prolonging effect of DR is only true for the model species, along with a less contentious idea that DR has more influence on females than on males. We found that the life-extending effect of DR was 20% smaller for male organisms than for females and also that DR was nearly twice as effective in prolonging lifespan in the model species than in the non-model species (BPMM: β[female/male difference] = 0.218, CI = 0.038–0.411 and β[non-model/model difference] = −0.666, CI = −1.121 to −0.222; Fig. 2A,B and Table S3). Although the significant sex effect is more or less expected from previous work (Partridge et al., 2005a), our finding is, to our knowledge, the first quantitative proof for the generality of the sex effect in the DR–longevity relationship. The housefly study, where DR convincingly failed to induce life extension (Cooper et al., 2004), has often been cited as evidence against the universality of the DR–longevity relationship (Le Bourg, 2010). Nevertheless, we point out that all houseflies in this experiment were males, so that the negative result could be predicted from our meta-analytic result, and it is not conclusive evidence against a life-extending effect by DR in this species. In contrast to the sex effect, the significant and clear model species effect we discovered is unexpected and fascinating. This finding supports the idea that the life-extending effect of DR is related to living in peculiar laboratory conditions for many generations (Miller et al., 2002; Austad & Kristan, 2003; more discussion later).
Importantly, the validity of estimates from meta-analysis relies on the assumption that there is negligible publication bias in a particular research topic (Egger et al., 1997). Inspection of funnel plot symmetries of our data revealed no obvious signs of publication bias in our data set (Figs 2C–H and S3; for the results of a regression approach (Egger et al., 1997), consistent with the absence of publication bias, see Table S4 and Dialog S1). Therefore, our estimates are likely to be reliable.
Caloric restriction or protein restriction?
We now build upon the above analyses to investigate the relative importance of caloric and protein restrictions. It is noted that, on the one hand, the variable caloric intake (%) represents the relative percentage of caloric intake for the DR group in relation to the control group where caloric intake was 100%. One the other hand, the variable protein intake (%) is the percentage of total food energy coming from protein in relation to the other macronutrients, namely carbohydrate and fat for both groups (see Dialog S1).
We found significant quadratic effects of both caloric and protein intake on the risk of death, with the life-extending effect of caloric intake peaking around 50% and that of protein intake about 30% (BPMM: β2[caloric intake] = 1.785, CI = 0.664–2.907; β[caloric intake] = −1.702, CI = −2.815 to −0.651; β2[protein intake] = 5.352, CI = 3.219–7.358; and β[protein intake] = −3.088, CI = −4.440 to −1.389; Figs 3and S4 and Table S5). Our results indicate that the effect of protein intake is larger than that of caloric intake, illustrated in Fig. 3A,B. This result is remarkable because, while most studies included in our data set explicitly changed the caloric intake between control and DR groups, very few studies deliberately manipulated protein intake (Data S1 and Dialog S2). Nonetheless, different studies covered a wide range of protein intake (0% up to approximately 90%). The contours of the DR effect on longevity in Fig. 3C show the importance of the balance between caloric intake and protein intake for DR to be effective.
This bivariate action of DR may explain many equivocal results in the literature, where researchers usually focused on caloric intake rather than protein intake for their interpretation. For example, animals that were food-restricted by 50% would not necessarily show the benefit of DR, if they were also fed a high-protein diet. Our results strongly support a recently proposed protein restriction hypothesis (Simpson & Raubenheimer, 2007, 2009) with an implication that this phenomenon may be general across the animal kingdom.
Notably, when we used alternative values based on actual food consumption as a measurement of caloric restriction, where available (32% of all the data points; Model 14 in Table S6), the life-extending effect of caloric intake disappeared but that of protein intake remained virtually unchanged (BPMM: β2[caloric intake] = −0.087, CI = −0.441 to 0.227; β[caloric intake] = −0.059, CI = −0.580 to 0.431; β2[protein intake] = 5.119, CI = 2.937–7.264; and β[protein intake] = −2.976, CI = −4.467 to −1.386; Table S6). Although this particular result is difficult to explain and reconcile, caloric restriction may not be the main determinant for the life-extending effect of DR, as indicated in a series of insect studies (e.g. Mair et al., 2005; Lee et al., 2008; Maklakov et al., 2008; Fanson et al., 2009).
How are aging trajectories changed?
We next examined the five model species in which the effect of DR was apparent to reveal how DR exerted life-extending effects. There are two ways that mortality rates can be reduced: (i) they are reduced by a constant fraction across the lifespan of an organism (i.e. a change in the age-independent mortality rate, also known as initial mortality rate) and (ii) the rate at which mortality rates increase across the lifespan is reduced (i.e. a change in age-dependent mortality rate; Partridge et al., 2005b; Phelan & Rose, 2005). In our data set, the two mechanisms of life extension can be identified and distinguished by meta-analytically estimating intercepts and slopes of ln(HR) values over the three relative time intervals (0–25%, 25–50% and 50–75% of the control group being dead, as described in Dialog S1 and Fig. S5). Our data set for the model species included 290 estimates for both intercepts and slopes from 105 studies (Data S3). Statistically speaking, we should observe a negative intercept with a zero slope if only age-independent mortality change is at work, whereas a significantly negative slope with a zero intercept if only age-dependent mortality is occurring. If both types of mortality change occur, a negative intercept and negative slope should be observed.
Overall, both age-dependent and age-independent mechanisms contributed to the DR life-extending effects (BMM: β[intercept] = −0.516, CI = −0.686 to −0.354; β[slope] = −0.181, CI = −0.252 to −0.120; Fig. 4, Table S7). Previous studies have claimed that in drosophila, DR reduces the age-independent mortality rate but not the age-dependent mortality rate, whereas for rodents, much of the life extension by DR stems from decreasing the age-dependent mortality rate (Partridge et al., 2005b; Phelan & Rose, 2005). However, our meta-analytic results show that the DR effects generally result from the dual actions of age-dependent and age-independent mortality changes. How DR brings about these two effects simultaneously will be an important future question.
Peculiarly, yeast seemed to be the only species where we did not find convincing evidence for age-dependent mortality change (BMM; β[intercept] = −0.695, CI = −1.263 to −0.144; β[slope] = −0.006, CI = −0.254 to 0.225; Table S7). This result could be due to the limited sample size in this species (note that the slope for yeast is not statistically different from the other species). Alternatively, it is possible that a biological reason can account for this observation. Longevity of a single-celled organism, yeast, is measured by ‘replicative’ lifespan (i.e. how long a set number of mother cells can keep generating daughter cells; Lin & Sinclair, 2008) rather than actual age as in multicellular organisms. In other words, how DR ameliorates aging may differ between single and multicellular organisms.
Evolutionary conservation or laboratory convergence?
Bayesian phylogenetic mixed models can provide an index, termed phylogenetic heritability (H2; Hadfield & Nakagawa, 2010), which quantifies the amount of phylogenetic signal (Lynch, 1991) in the data set and ranges from 0 to 1 (for details, see Dialog S1). Our first BPMM, which only included the intercept (i.e. meta-analytic mean) as a fixed factor, had a reasonable amount of phylogenetic signal (BPMM: H2 = 0.351, CI = 0.004–0.512; Table S2), suggesting that the DR–longevity relationship was only apparent in some species but not others. However, our final BPMM, which accounted for the sex effect, the model species effect and the effect of protein and caloric intake, had much less phylogenetic signal (BPMM: H2 = 0.003, CI = <0.001–0.450; Table S5). Therefore, these four variables seem to explain most of the phylogenetic signal (i.e. between-species differences) in the original analysis. Given that the evidence for life extension by DR is weak in the non-model species (Fig. 2) and that the five model organisms represent four distantly related animal phyla, we suggest that the life-extending effect of DR is a response only evident when animals have been housed in laboratory conditions for a number of generations. There are a number of explanations why this may be the case.
It is possible that life extension by DR is a laboratory-induced convergence among the model species. This laboratory convergence explanation, however, runs counter to recent findings that the same or homologous nutrient-signalling pathways (e.g. TOR, IIS) are involved in prolonging lifespan among these distantly related model species (Fontana et al., 2010; Partridge, 2010; Bartke, 2011; Kenyon, 2011; McCormick et al., 2011). It is more parsimonious to think that such pathways are evolutionarily conserved and the life extension by DR is an ancestral response incurred by the nutrient-signalling pathways shared by contemporary animals. Therefore, we speculate that there may exist a neglected pathway, which reduces lifespan in animals, cancelling the life-extending effects associated with nutrient-signalling pathways. This neglected pathway may be less functional in the model species after many generations of cosseted laboratory life. It is also possible that such a pathway does not exist, but the life-extending effects of nutrient-signalling pathways are somehow more pronounced in the model laboratory species, which have been inadvertently selected for such a ‘positive’ response (Harper et al., 2006). A recent study revealed significant heterogeneity in the response to DR among 41 inbred straits of mice; DR even shortened lifespan in some stains (Liao et al., 2010). This observation indicates that laboratory selection of different sorts could result in a variety of responses to DR. Yet, another possibility is that researchers are able to control extraneous factors such as housing conditions and fine-tuning of experimental diets for the model species much better than for non-model species, so that the life-prolonging effect is exaggerated in the model species. If so, it would seem that life extension via DR is less effective and less general than previously supposed and, therefore, might be less applicable to humans.
Comparative and meta-analytic insights
There have been numerous qualitative papers discussing the universality and/or specificity of the life-extending effect of DR across species (Partridge & Brand, 2005; Phelan & Rose, 2005; Le Bourg & Rattan, 2006; Mair & Dillin, 2008; Le Bourg, 2010). Our meta-analytic approach goes significantly beyond these verbal assessments, however, and shows that, given the right dietary conditions, DR could extend lifespan of any species, although the effect may be modulated by important factors such as sex. Our work represents the first formal meta-analysis on the topic of the DR–longevity relationship. We see a number of fertile avenues for future meta-analytic research, which will offer quantitative answers to controversies in the field of nutrition-related gerontology (which are summarized in Table 1). Notably, the statistically robust results revealed by our meta-analysis cannot be extrapolated directly to individual species (Nakagawa & Cuthill, 2007); real differences among species remain.
Table 1. Potential meta-analytic topics in nutrition-related gerontology
These meta-analyses may be carried out within specific species or across different species.
Quantifying and ranking the usefulness of biomarkers as longevity indictors (e.g. oxidative damage, body temperature, glucose levels in blood)
Quantifying the effects of dietary restriction (DR) on health benefits (e.g. protection against cancer and cardiovascular disease)
Quantifying and comparing the effects of DR on a range of different traits, such as physiology, behaviour, life history and cognition
Quantifying the effects of and heterogeneity among different single-gene mutations, which mimic DR in extending lifespan
Quantifying the efficiency of life-extending compounds, such as the target of rapamycin (TOR) inhibitors, sirtuin-activating compounds (STACs; e.g. resveratrol) and antioxidants
Quantifying the effects of immune challenges and pathogen/parasite infections on immune responses and lifespan under DR
Quantifying the effects of stressors (e.g. heat shocks, pollutants) on lifespan and longevity biomarkers under DR
To conclude, our study reaches several significant findings that may resolve many conflicting results from previous studies: (i) DR is generally more effective at increasing female longevity than male longevity, (ii) the balance between protein and caloric restriction is a key factor to maximize the effect of DR in extending lifespan, (iii) DR induces life extension generally by reducing both age-independent and age-dependent mortality rates, and (iv) the effect of DR on longevity seems only apparent in the model species, but not in the non-model species, supporting the laboratory convergence explanation of the DR–longevity relationship. Although alternative explanations are possible for the last point (see above), future investigation into this hypothesis is warranted.
The data for meta-analysis were collected using ISI Web of Science and Google Scholar with the search string ‘calorie*/diet*/energy/food’ + ‘restrict*’ + ‘longevity/lifespan’. Backward and forward searching was also carried out primarily using review articles that focused on non-rodent or non-mammalian studies to locate smaller studies on less common species which may not have been conducted with a focus on calorie restriction and longevity in mind. Papers published at any date up to December 2009 were included. Of over 2000 studies this search yielded, papers were selected that contained a graphical survival curve or, in the case of some older studies, that gave the complete data set from which a survival curve could be constructed. Papers were also excluded from analysis based on a number of methodological criteria that are detailed in Dialog S1. From the included studies, we extracted effect size measures: we used survival curves to quantify the natural logarithms of hazard ratio, ln(HR) (Table S1). In short, a set of three ln(HR) values were extracted from each pair of survival curves (consisting of the control group and the DR group) at three relative time intervals (during which 0–25%, 25–50% and 50–75% of the control group died), and the overall estimates from these three values constituted effect size values as ln(HR) (Fig. S1; Parmar et al., 1998; Williamson et al., 2002). To analyse the collected data, we used standard procedures associated with meta-analysis and meta-regression, as described in Dialog S1. BMM and BPMM were implemented in the R package, MCMCglmm (Hadfield, 2010; Hadfield & Nakagawa, 2010).
We thank J. Hadfield, S. Morgan, D. Raubenheimer, L. Tain, S. Gronke, M. Piper, P. Dearden, the SN lab members and two anonymous referees for their comments and advice, which improved the earlier versions of this paper. We are also grateful to K. Miller for helping with figure production and J. Hadfield for statistical advice. This project was supported by funding from National Centre for Growth & Development, New Zealand, and funding from the Department of Zoology, University of Otago, New Zealand. SN is supported by the Royal Society of New Zealand Marsden Fund.
SN conceived the idea for the study, all authors contributed to design the study, SN, ML and KLH collected data, SN conducted data analysis, SN and ML wrote the first version of the manuscript, and all authors contributed to the final draft.