GENOMIC BASIS OF AGING AND LIFE HISTORY EVOLUTION IN DROSOPHILA MELANOGASTER

Selected Populations

We applied selection on survival and late-age reproduction in replicated, paired selection (S) and control (C) populations of D. melanogaster. The ancestral population was derived from ~8000 offspring of 400 wild-caught females collected in New Jersey in 1998. This population was maintained in the lab at large population size (>1000) until 2004, when we obtained >1600 individuals from Dr. Alan Gibbs. We then maintained the population with overlapping generations (flies transferred to new food every 14 days, but without clearing of parental flies) until the beginning of the selection experiment in late 2005. During that period, we maintained the population in vials (100 vials, >25 flies per vial) and on media identical to those used during the selection experiment.

We established three independent S-C pairs by taking three independent samples of founders from the ancestral population. Offspring of 320 single-pair matings were used to establish both the S and the C population within a replicate. A different set of 320 pairs was used to establish each S-C pair, such that each pair represented an independent sample of the genetic variation existing within the progenitor population. After the initial generation, the S and C lines of a pair were maintained independently for the remainder of the experiment. Each generation, S and C populations were propagated from 220–320 single-pair matings. Parents were cleared after 7 days, but we collected offspring on days 10–11, to ensure that they were derived from eggs laid early in the mating stage. Equal numbers of offspring were collected from each pair and allowed to age in single-sex vials until 14 days of age for C lines or N days of age for S lines (at the beginning of the experiment N = 28; N was gradually increased to 40 as flies evolved longer reproductive life spans). At the appropriate age, flies were paired with other surviving individuals from their population, but from a different family, to produce single-pair matings for the next generation. That is, in S populations, only flies that survived and were fertile at >N days of age contributed to the next generation; in C populations, flies that survived and were fertile at 14 days contributed to the next generation. We chose reproduction at 14 days for C populations to mimic average age at reproduction in the overlapping-generation rearing protocol for the ancestral population and to avoid artifactual selection for too-early reproduction (Linnen et al. 2001). Approximately 50% of individuals in S populations survived and contributed offspring to the next generation. We avoided differential selection on development time by collecting newly emerged adults within a 24-hour period beginning 10 days after initiation of egg laying. In early 2009 (~ generation 35), the single-pair mating strategy was changed to three males and three females per vial to increase productivity of S lines. Because of the different generation times, S and C lines were in the experiment for different numbers of generations. The genomic and phenotypic data reported here are based on the 50th generation of selection in the S lines (about 80 generations of selection in the C lines). S- and C-line generations were not perfectly synchronized, but phenotypic assays of the S and C lines within a pair were begun within 7 days of each other.

The selection scheme was chosen to minimize directional selection in C lines relative to the laboratory conditions under which the founding population had been maintained, but we do not claim that no selection was imposed on C lines during laboratory evolution in this experiment. Our goal was to impose differential selection on age at first reproduction in S and C lines and then to identify genomic regions that diverged during the period of differential selection.

Phenotype Assays

Sequencing

To evaluate effects of selection, we prepared genomic DNA from 100 female flies for each of the six populations into six libraries (one per population) and sequenced them as paired-ended 74-bp reads on an Illumina GAII. Each library was individually barcoded and used to create a pooled sample of barcoded DNA. Use of an aliquot of the pooled sample in each of six lanes of the GAII minimized lane effects. This process generated 514 million reads and 3.8 × 10¹⁰ bp of sequence. Sequenced reads were aligned as pairs, with BWA 0.5.7 (Li and Durbin 2009), to the complete D. melanogaster 5.29 genome build downloaded from Flybase. Reads were allowed up to six mismatches throughout the 74 bp per end, and unique reads were mapped to the genome. All other BWA alignment parameters were set to default values. Approximately 80% of the sequencing reads mapped uniquely and with high stringency to the Flybase 5.29 genome build, resulting in ~244X coverage of 126 Mb of the D. melanogaster genome (95% of euchromatic and 52% of heterochromatic DNA). Median coverage ranged from 40 to 50X across all six population samples. Sequence data are archived at the NCBI Short-Read Archive under accession number SRA038471/SRP007248.

Allele Frequency Differences

Initially, reads from all six populations were pooled to identify multiallelic sites across the genome. For this comparison, 2,106,064 positions were considered, all having at least two alleles observed in the pooled data and each allele called at least five times with frequency of at least 5% in the pooled data. To identify genomic positions with differences in allele frequencies at these sites, we used a Fisher exact test to compare the counts of all base calls at each position between S and C populations (either as independent replicate pairs or pooled). We applied a Bonferroni cutoff of 0.05 to call allele differences significant.

Signatures of Selective Divergence

We compared the sequences of selection and control lines to detect genome regions that had differentiated significantly. Differences in allele frequency between S and C populations could arise from selection, from the stochastic effects of genetic drift, or from binomial sampling error. To detect regions where divergence was more probably driven by selection in S populations, we scaled the heterozygosity in S populations

a measure of allelic variation, by the allelic divergence between S and C populations (D), where D was calculated as the average number of pairwise differences between two populations. This statistic is analogous to the HKA test of deviation from neutral evolution (Hudson et al. 1987); it is expected to be small in genome regions that experienced directional selection on a sequence variant in the S but not the C population. Selection on a sequence variant should carry the variant (along with flanking genome regions) to high frequency, increasing D and decreasing Hs in a “selective sweep.” H_S/D is therefore expected to be small in regions that experienced a sweep in S populations. Although we designed the experiment to minimize evolution of C lines during the selection phase, this method of evaluating divergence of allele frequencies between S and C populations does not depend on any assumptions about selection or neutrality in the C populations.

To identify genomic signatures of selection, we calculated expected H_S/D for all nonoverlapping 1-kb windows. To calculate likelihood scores for each statistic, we summed individual scores across 100 1-kb regions and across all three replicate pairs and used the empirical genomewide distribution of this statistic to calculate the probability of observing more extreme values by chance across 100-kb windows using Equation 1:

where F is the empirical genomewide distribution of the appropriate statistic h and H = (H_1j, H_2j, …, H_100j) are the 100 individual 1-kb measure of h within a 100-kb window for the jth replicate. Because the overlapping 100-kb regions were not independent, we used permutations to identify a significance level. Bootstrap resampling of 100 random 1-kb regions for each replicate indicated that between 2.5 and 4.6% of samples had scores that exceeded 100. We chose this cutoff of 100 and assumed this false positive rate. After identifying 100-kb regions with scores greater than 100 for each statistic, we collapsed overlapping regions.

Although the H_S/D statistic is sensitive to classic selective sweeps, it is not sensitive to “incomplete sweeps,” in which selection in S lines favored an allele that was initially at low frequency and carried it to intermediate frequency during selection (Przeworski et al. 2005). Under this scenario, genome regions with high H_S would also exhibit increases in D. To identify these incomplete sweeps, we therefore scaled H_S (expected to be high) by H_C (expected to be low), using a likelihood calculation equivalent to that in Equation 1.

We also assessed divergence among replicate C lines and among replicate S lines. No genomic regions had log likelihood (LL) > 17 in any of these comparisons (results not shown). We therefore restrict our attention here to differentiation between the two selection regimes.

Gene Expression

To identify genes within sweep regions that also diverged in gene expression, we evaluated mRNA abundance using microarrays. Comprehensive analysis of genomewide expression patterns will be described in a future publication. We extracted RNA from two independent groups of female flies for each population/age/tissue combination (six populations, five ages, two tissues for a total of 120 samples). Each replicate consisted of tissue from seven individual flies. We used the PicoPure RNA isolation kit (Arcturus); RNA was amplified with Ambion’s MessageAmpII aRNA Amplification Kit (Ambion) and reverse transcribed to cDNA by the SuperScript III protocol (Invitrogen Life Technologies). Samples were labeled by means of Roche-Nimblegen’s One Color DNA Labeling Kit, and cDNA from each replicate was hybridized to its own partition of Roche-Nimblegen’s D. melanogaster 12×135K expression arrays. Each array consisted of 12 partitions, and each partition contained 16,637 target genes with eight probes per gene (DM 5.7 genome build). Arrays were scanned with a GenePix 4000B Scanner and Roche NimbleScan software.

For analysis, we used GCRMA-normalized expression values as the dependent variable in gene-specific mixed models (Chu et al. 2002) implemented in SAS Proc Mixed v. 9.2. Selection regime, age, and selection-by-age interaction were fixed effects; population within selection regime was treated as the level of independent replication. Biological replication within population-age-tissue combinations was treated as a random effect nested within population-age-tissue to account for the expected correlation between samples drawn from the same population. Chip (array) was also treated as a random blocking effect. Except as noted, we applied a false-discovery-rate cutoff of 0.05 to call effects significant (Benjamini and Hochberg 1995). Gene enrichment analysis was conducted with DAVID (Dennis et al. 2003), and statistical tests of enrichment were performed within FlyMine (Lyne et al. 2007). We determined over- and underrepresentation of ontology terms by comparing the number of genes falling within each category to the number of transcripts represented on the Nimblegen Drosophila array that fall into that category. FlyMine uses a hypergeometric test to determine whether an ontology category is significantly overrepresented in a list of differentially expressed genes. Ages at which gene expression was measured did not correspond exactly to ages at which fecundity was measured because we wished to capture gene expression in newly eclosed (prereproductive) flies and because budgetary constraints prevented us from assaying expression at more ages. For logistical reasons, the tissue used in expression analysis was extracted from a different generation than the tissue used for DNA analysis. Expression data was based on flies from generation 40 of selection. Microarray data are archived at NCBI Gene Expression Omnibus under accession number GSE38106.

Phenotypic Response to Selection

Figure 1 illustrates the results of 50 generations of selection in the S populations. In each replicate, S flies had longer virgin and mated life spans (all log-rank tests P < 0.005; mean over all replicates ± standard error: S virgin females = 69.0 ± 0.4 days, C virgin females = 60.5 ± 0.4, S virgin males =72.0 ± 0.5, C virgin males = 62.2 ± 0.4;) and longer mated life spans (S mated females = 34.1 ± 0.5; C mated females = 29.9 ± 0.4, S mated males = 44.2 ± 0.6; C mated males = 40.7 ± 0.7). A generalized linear model fit to fecundity data collected from five adult ages (7, 14, 35, 44, and 60 days after eclosion) indicated highly significant differences between selection regimes (χ² = 15.5, P < 0.0001) and ages (χ² = 1619.2, P < 0.0001) and a highly significant interaction between selection regime and age (i.e., a difference in how S and C fecundity changed with age, χ² = 48.9, P < 0.0001). Post-hoc tests indicated that S lines had higher fecundity at 44 days after eclosion (S = 8.6 ± 0.4, C = 6.1 ± 0.4, P < 0.005) and lower fecundity at 14 days after eclosion (S = 26.9 ± 0.7, C = 30.0 ± 0.8, P < 0.01). Means and standard errors for the complete fecundity data are shown in Table 1. These phenotypic results parallel those seen in other studies of selection on late-age reproduction (Rose and Charlesworth 1980; Rose 1984; Partridge et al. 1999), but see Partridge and Fowler (1992).

Genetic Divergence and Tests for selective Sweeps

Figure 2a shows H_S/D calculated at 1-kb intervals across the genome and illustrates seven genome regions where the statistic is significantly below background levels in a comparison of all S and C populations (and two regions significantly above background, see Supplemental Table S1 for LL values for genes within these regions). Regions with low values of H_S/D were similar in this pooled comparison and in each of the three S-C population comparisons individually, suggesting that the same regions responded to selection in each replicate (Supplemental Fig. S1). Four of these sweep regions are near centromeres (on X, 2L, and 3L). We do not know of any methodological biases that could lead to this pattern. These regions fall within chromosomal neighborhoods characterized by low recombination rates (Fiston-Lavier et al. 2010) and might therefore be subject to greater hitchhiking effects.

We also observed several additional genome regions with H_S significantly elevated relative to H_C, consistent with the pattern expected under an incomplete sweep. Six chromosomal regions had H_S/H_C values significantly higher than background levels (Fig. 2b, which also shows three regions with values significantly above background LL values, see Supplemental Table S1). Again, regions of extreme values for this statistic were similar in each replicate S-C pair of populations (Supplemental Fig. S1). These regions also exhibited significantly reduced frequency in S populations of the most common allele in C populations and high F_ST values (Fig. 2c,d, and Supplemental Fig. S1). This pattern is similar to that reported in a recent study of response to selection on development rate in D. melanogaster (Burke et al. 2010). These incomplete-sweep areas occupy relatively broad chromosomal regions, as would be expected if the sweeping allele derived from only one or a few ancestral haplotypes and adjacent loci evolved by genetic hitchhiking (Hudson 1994).

Candidate Genes within Sweep Regions

We identified the loci most strongly implicated in sweeps as those with the highest likelihood scores for H_S/D or H_S/H_C within each candidate region (Supplemental Table S1), using an LL difference ≥ 2 to indicate a meaningful difference in scores. Using this criterion, we identified 156 genes, ranging from one to 21 genes per region. These genes are enriched for Gene Ontology biological-function categories related to oogenesis, immune response, and proteolysis (Table 2). In an attempt to identify candidate causal nucleotides, we evaluated these 156 genes’ changes in coding regions. Fifteen genes exhibited nonsynonymous changes (or start-codon changes) that were unique to either S or C samples (Table 3). Moreover, nine of these 15 genes (60%) have been implicated in aging, reproduction, immune response, or proteolysis in previous studies: CG12004 (response to microbial infection; Tanji et al. 2006); Doc3 (age-related motor impairment; Jones et al. 2009), Fhos (response to microbial infection and aging; Boutros et al. 2002; Lai et al. 2007), fs(1)M3 (oogenesis; Degelmann et al. 1990; Cernilogar et al. 2001), Paramyosin (oviduct and ovary response to mating; Apidianakis et al. 2005; Kapelnikov et al. 2008), Tequila (response to aging, oxidative stress, and microbial infection; Zou et al. 2000; Munier et al. 2004; Lai et al. 2007), Ten-a (a direct target of doublesex; Luo et al. 2011), and CG6041 and CG6048 (serine-type proteases). Because significant differential expression was observed for only two of these nine genes (Ten-a and Paramyosin, see below), the nucleotide differences in the other 13 genes may encode structural-protein variation that contributes to effects on late-life health and performance.

Expression and Phenotypic Effects of Genes Implicated in Sweeps

We next looked for variation in the expression of candidate genes that could indicate cis-regulatory evolution. Of the 156 genes most strongly implicated in sweeps, 25 were differentially expressed between S and C populations at false discovery rates < 0.10 (Supplemental Table S2). We found these differentially expressed genes within three of the seven complete-sweep regions and five of six incomplete-sweep regions. The differentially expressed genes included three with functions related to reproduction (Cct1, karst, and Ten-a), two encoding antimicrobial peptides (dro-2 and dro-4), and one encoding a peptidase (Psa). Cis-regulatory change in these genes is therefore implicated in extended life span and late-age fertility of the selected lines. Functional studies support this hypothesis for at least two genes. Cct1 encodes a membrane phospholipid and the rate-limiting enzyme in phosphatidylcholine synthesis. Experimental up-regulation of this gene increases fly life span by 8% (Landis et al. 2003). Our expression data are consistent with that result: Cct1 was up-regulated in S relative to C flies at all ages beyond the day of eclosion (Supplemental Figure S2a). The aminopeptidase Psa has also been associated with effects on age-related health and performance in flies and in mammals. In humans, the PSA protein can protect brain cells from degeneration caused by aggregations of abnormal tau proteins, which is thought to have a causative role in Alzheimer’s, Parkinson’s and other age-related neurodegenerative disorders (Karsten et al. 2006). In flies, experimental up-regulation of Psa protects against tau-induced neurodegeneration (Karsten et al. 2006). In our experiment, Psa was up-regulated in S flies relative to C flies at all ages except 50 days (Supplemental Figure S2b), suggesting that S flies would be less susceptible to tau-induced neural defects.

Natural Genetic Variation Underlying Life-History Evolution and Life-History Trade-offs

Using experimental evolution, high-throughput whole-genome resequencing, and transcriptome profiling, we found 156 genes that responded significantly to selection on improved late-life performance. By focusing on genes with strong signatures of selection and either (1) changes in nonsynonymous sites or (2) changes in gene expression, we constructed a narrower list of 38 candidate genes demonstrating structural and regulatory variation. Moreover, 14 of these genes (37%) have demonstrated effects on functions related to reproduction and aging (eight from the nonsynonymous change list, five from the differential expression list, and one gene, Ten-a, that appeared on both lists). Our focus on these genes does not mean that we exclude the other 118 strongly differentiated genes as candidates, only that we were not able to identify obvious functional differentiation. We would not have found changes in noncoding regions that regulate mRNA stability or translocation or that influence posttranslational modifications.

We imposed only 50 generations of selection, so the divergence we observed is likely to reflect changes in frequency of alleles that were already segregating in the ancestral population and not new mutations that arose during the course of the experiment. Alleles contributing to segregating variation have been filtered by natural selection and are maintained in a population either by some form of balancing selection or because weak purifying selection is balanced by recurrent mutation (Charlesworth and Hughes 2000). Either initially rare alleles (more likely to be transient alleles maintained by mutation-selection balance) or intermediate-frequency alleles (more likely to be maintained by balancing selection) could have contributed to near-complete or incomplete sweeps. Alleles that were at intermediate frequency in the ancestral population would have occurred on many different genetic backgrounds. Change in frequency of those alleles would not be expected to create broad chromosomal regions exhibiting sweep signatures, as would be expected for alleles that were rare in the ancestral population (Hermisson and Pennings 2005). The large sweep regions we observed, especially those in locations with high recombination rates (e.g., the middle of X, 2L, and 3L), therefore support the involvement of at least some rare alleles, but we cannot rule out the possibility that some balanced polymorphisms were initially rare in our founding population because of sampling or laboratory selection. Patterns consistent with contributions from both rare and common variants have also been reported from selection on body size in D. melanogaster (Turner et al. 2011).

The very broad region of high heterozygosity in S lines on chromosome arm 3R falls within the known breakpoints of the cosmopolitan inversion polymorphism In(3R)Payne, which could explain the large expanse of that signature of selection, but the other candidate sweeps are not associated with known inversion polymorphisms. Together, complete and near-complete sweep signals and broad and narrower sweep regions suggest that rare, formerly deleterious alleles and alleles that were maintained by balancing selection contributed to the life-history evolution that we observed.

Because selection on longevity and late-age reproduction resulted in increased life span, improved late-age fertility, and decline in early fertility, the candidate genes we identified are also candidates for regulating a major life-history trade-off between early- and late-life fitness components. Such trade-offs are key components of life-history theory and of some models of the evolution of aging (Rose 1991; Roff 2002), yet genetic mechanisms underlying these trade-offs are virtually unknown (Charlesworth and Hughes 2000; Hughes 2010b). Our results suggest that genes regulating oogenesis, immune function, and proteolysis are important in conferring improved late-age performance at the cost of early-age fecundity. Perhaps the best candidate for directly mediating a trade-off between early fertility and life span is Cct1, which is known to participate in ovarian development (Gupta and Schupbach 2003) and has been shown to increase life span when up-regulated (Landis et al. 2003); effects of up-regulation on fecundity are not known, but the prediction based on our results would be that changes that lead to life-span increase should also lead to fecundity decline. fs(1)M3 and Prm are involved directly in oogenesis and ovary function, respectively (Degelmann et al. 1990; Cernilogar et al. 2001; Apidianakis et al. 2005; Kapelnikov et al. 2008), but effects on life span have not been reported.

“Repeatability” of Evolution

The extent to which evolution is “repeatable” is being actively debated (Arendt and Reznick 2008; Stern and Orgogozo 2008; Hughes 2010a; Rosenblum et al. 2010). The phenotypic response to selection was similar in all replicates, and the regions we identified as responding to selection were the same across three different replicate S-C pairs (Supplemental Fig. 1S). This pattern supports the hypothesis that evolution was repeatable at the molecular level in our experiment. The three population pairs were initiated from the same founding population but represent three independent samplings of the genetic variation present in that population. Results similar to ours have been reported in a recent study of natural populations of sticklebacks undergoing repeated adaptation to fresh water (Jones et al. 2012) and in a genomewide sequencing study of D. melanogaster populations selected for development time (Burke et al. 2010). Together, these results suggest that repeated adaptation from a common ancestral population leads to similar underlying genetic changes, despite the potential influence of stochastic processes.

Targets of Selection

That late-age female fertility was a direct target of selection in our S populations could account for the prominence of genes with oogenesis-related functions among the strongest candidates for causal sequence changes and differential expression. Late-age male fertility was also a target of selection, but genes related to male fertility were not overrepresented among candidates for causal sequence changes (χ[1]² =0.004, P > 0.9) or among differentially expressed genes (χ[1]² = 0.29, P > 0.5). The D. melanogaster genome contains 2.5 times more genes annotated with functions related to oogenesis than to spermatogenesis (296 and 117, respectively), so failure to detect changes in male fertility genes might be due to underdescription of the genetic basis of spermatogenesis relative to that of oogenesis. If annotation is unbiased, then our results suggest that oogenesis presents more targets for selection and possibly higher levels of standing genetic variation. Alternately, selection on male fertility might be weaker because of the ability of females to store sperm for long periods.

Although not a direct target of selection in this experiment, immune response has been linked to longevity and late-age morbidity in both fruit flies (Sarup et al. 2011) and humans (Bonafe et al. 2001; Franceschi et al. 2007). In flies, immune-response genes exhibit large increases in expression with age, and the expression of antimicrobial peptide reporters predicts subsequent life span in individual flies; long-lived individuals have lower expression of these genes when they are young (Landis et al. 2004). Again, this pattern is consistent with our results: drosomycin genes dro-2 and dro-4 were down-regulated in S relative to C flies at early and intermediate ages (Supplemental Fig. S2c, d).

Novel Candidates for Regulating Variation in Aging

Identifying and characterizing alleles responsible for natural variation in aging-related traits is relevant to biomedicine because these alleles are representative of the variation that segregates in human populations. Moreover, mechanisms of aging appear to be conserved in vertebrates and invertebrates (Partridge and Gems 2002), so human variation for late-age health and performance might be regulated by genes and pathways that are homologous to those we report here.

One surprising result of our experiment was the relatively weak support for involvement of the insulin/insulin-like signaling (IIS) or Target-of-rapamycin (TOR) pathways in the evolution of late-life performance. Mutations in genes within these pathways can alter life span and fertility in flies and other organisms (Partridge and Gems 2002); natural genetic variation in expression of IIS/TOR-pathway genes has been reported to predict aging-related phenotypes (Nuzhdin et al. 2009), and natural clinal variation in the insulin receptor gene InR has been associated with variation in stress resistance and fecundity (Paaby et al. 2010). We therefore expected that some of these genes would contribute to the evolution of life span and late-life fecundity in our experiment. Only one gene previously annotated with the Gene Ontology biological function “determination of adult life span” (Cct1) was among the genes bearing the strongest signature of selection, no more than would be expected by chance (1/96 of the candidate genes that had some biological process annotation, compared to 116/10792 of all genes with some biological-process annotation, χ[1]² = 0.002, P > 0.96). Genes annotated with the functions “aging” or “determination of adult life span” were also significantly underrepresented among differentially expressed genes (43/215 transcripts with these annotations had P < 0.05 for line or line-by-age effects, compared to 4488/13258 of all annotated transcripts, χ[1]² = 18.1, P < 0.0001). Most of the genes we identified are therefore novel candidates for the regulation of life span and late-age performance.

These results do not necessarily contradict the empirical evidence that IIS/TOR pathways are key regulators of aging that are capable of sensing and responding to environmental conditions. The allele-frequency changes we observed necessarily reflect segregating variation that was present in the founding population. If genes in these pathways were functionally invariant in that population, then they could not contribute to the evolution of aging and fertility patterns that we observed. Functional variants could have been lost during the initial sampling from the wild, particularly if they segregated at low frequency in the wild population, or they could have been lost as a result of selection or drift during laboratory husbandry. Two other studies of natural variation have identified allelic variation in InR and associated the polymorphism with variation in life span (Geiger-Thornsberry and Mackay 2004; Paaby et al. 2010). We note, however, that both these analyses focused on phenotypic variation among inbred lines of flies, and we cannot be sure that homozygous effects of alleles predict their contribution to heritable variation in outbred populations. To our knowledge, the present study is the first to identify naturally occurring allelic variation associated with a life span–fecundity trade-off in outbred populations.

With the above caveat, our results support the hypothesis that genes in these critical metabolic pathways are under selective constraint in natural populations and might not be the most important determinants of naturally segregating variation (Hughes and Reynolds 2005; Partridge and Gems 2006). IIS/TOR pathways are thought to regulate longevity and aging in part by up-regulating response to stress and to immune challenge (Becker et al. 2010); they also participate in a negative feedback loop involving proteolysis (Rui et al. 2001). Our results are therefore more consistent with evolutionary change in downstream targets of IIS/TOR than with change in the upstream regulators of these pathways.

The high proportion of novel candidates that we identified is consistent with the results of a recent screen of lines homozygous for single P-element insertions (Magwire et al. 2010). Only one of 58 mutations that increased life span in that screen had been known previously to affect aging. That analysis also showed that a very large fraction of insertions affected life span and that mutations that increased life span had extensive pleiotropic, sex-specific, and epistatic effects, many of which were antagonistic. Although we focused on natural variants and Magwire et al. (2010) on P element insertions, taken together these data imply that the regulation of aging involves interacting networks of genes, many of which have not yet been identified.