To better understand the changes caused by CR, mice were fed diets high (“HIGHCAL”) or low (“CR”) in calories for 14 days and total liver RNA was used to produce expression profiles using spotted oligonucleotide DNA microarrays. Food was administered twice daily to reduce the complications of inducing a starvation response, including prior to sacrifice. Body weights of most mice stabilized after the first week and even the weights of most CR mice reached plateau by 7 to 9 days after initiation of CR. After background threshold adjustment and normalization our microarrays detected 8,347 non-redundant genes. To statistically test which genes are associated with CR relative to high calorie feeding, we used Significance Analysis of Microarrays (SAM) [26]. SAM performs a moderated t-test on two user-selected groups of microarray datasets, and returns a set of discriminant genes ranked in descending order by d, the t-statistic, and calculates a false discovery rate (FDR) based on random permutations of the data labels (see Methods). SAM also reports the q-value for each ordered gene (the lowest FDR at which the gene is called significant) and several other measures including fold-change values for each gene.

At a restrictive cutoff of q<0.01, 1,897 of 8,347 detected genes (22.7%) differed between CR and HIGHCAL (3,855 genes (46.2%) at the permissive cutoff of q<0.1). As shown in Table S1 , Quantitative RT-PCR (Q-RT-PCR) confirmed direction of change (up or down) of 18 transcripts called by SAM as differentially expressed, and fold change values in most cases were larger for Q-RT-PCR (subsequent fold changes given are microarray values). The 90^th percentile number of median false positives at the permissive cutoff is about 634 genes. This suggests that even without considering the false negative rate the number of true positive genes changed significantly by CR can be conservatively estimated to be over 3,000. All genes called as significant at both cutoffs are listed in Table S2 . Of the 1,897 genes changed by CR at q<0.01, 827 (43.6%) are upregulated and 1,070 (56.4%) are downregulated relative to HIGHCAL (at q<0.1, 1,693 (43.9%) are upregulated and 2,162 (56.1%) are downregulated). Figure 1 shows all 3,855 genes in the permissive list (cutoff q<0.1) in a plot of significance (absolute value of d, |d|) versus log₂ fold change. The shape of this plot shows that even though the majority of genes are downregulated, many upregulated genes show greater fold-changes and |d| scores than downregulated genes. Overall, 27 genes were upregulated more than 10-fold, while only 4 genes were downregulated more than 10-fold. This general asymmetry held true irrespective of the normalization method used.

As highlighted in Figure 1 , the single most significant and one of the most fold-changed discriminant genes is DNA-damage inducible transcript 4 (Ddit4), a p53-regulated negative regulator of the Frap1 pathway [27], which is increased by CR by over 33-fold (80.1-fold as measured by Q-RT-PCR). In addition, brief visual inspection of the significance-ranked list of genes reveals that many other top-ranked genes are involved in a variety of processes related to aging, hormonal synthesis and signaling, cell cycle regulation and apoptosis. To begin to further analyze these classes of significant genes in a systematic way we determined the intersection of these genes with the human and mouse genes listed in the GenAge database [28]. As described below, analyses for enriched Gene Ontology (GO) terms confirmed many of the broader general categories observed by visual inspection, however, GO terms and even GenAge annotations currently are too coarse to capture certain aging-related classes of genes and processes. One important excluded class among GO terms is genes that are known members of aging-related pathways but which have not been isolated as mutants that alter lifespan; key regulators of these pathways are of particular importance. The GenAge database includes many such genes. Table 1 lists the intersection of our set of 1,897 significant genes (q<0.01) with the set of 265 non-redundant mouse and human genes listed in GenAge plus seven selected genes of this type that are not listed in GenAge but are among the most significant differentially expressed genes in our data.

Among aging-related genes upregulated by CR are leptin receptor (Lepr), cathepsin L (Ctsl), plasminogen activator urokinase (Plau), cyclin-dependent kinase inhibitor 1A (Cdkn1a) (a.k.a. p21/Cip), nicotinamide N-methyltransferase (Nnmt), peroxisome proliferative activated receptor, gamma, coactivator 1 alpha (Ppargc1a) (a.k.a. Pgc-1 alpha), serum/glucocorticoid regulated kinase 1 (Sgk1), and two Cebp paralogs: CCAAT/enhancer binding protein delta (Cebpd) and CCAAT/enhancer binding protein beta (Cebpb). Downregulated genes include aprataxin (Aptx), telomerase reverse transcriptase (Tert), insulin-like growth factor binding protein, acid labile subunit (Igfals), and growth hormone receptor (Ghr). Lepr was upregulated by over 58-fold by CR (49.3-fold by Q-RT-PCR) and was third in the list of genes ranked by significance. An increase in Lepr transcript abundance was somewhat expected since this is a reported result of CR [29] and of starvation [20] but the increase we observed was about 10-fold greater than seen in these previous reports. All of the genes in these lists are of possible importance in the life-extension response to CR but the large fold changes, high statistical ranks, and known roles in key aging-regulatory pathways make Ddit4, Nnmt1 and Cdkn1a especially interesting.

The large increase in expression of Ddit4 suggests inhibition of the Frap1 (TOR) pathway by CR. Frap1 is a protein kinase whose currently known substrates include the kinases Akt1 and S6K1, and the translation regulatory protein Eif4ebp1 [9], [10]. Recent data suggest Frap1 enhances initiation of 5′-cap-dependent translation of most mRNA species by phosphorylation of Eif4ebp1, promoting its dissociation from the translation initiation factor Eif4e [30], [31]. Several sites on Eif4ebp1 are phosphorylated but phosphorylation of Thr69 (human Thr70) is a primary determinant of this dissociation [31]. Phosphorylation of S6K1 on Thr390 (human Thr389) is a primary activator of its kinase activity and appears to regulate translation of ribosome-related transcripts and ribosome biogenesis, although the mechanism underlying differential regulation of phosphorylation of Eif4ebp1 and S6K1 by Frap1 is currently unknown [31]. Ddit4 has been shown to strongly inhibit Frap1 activity [32] and simply overexpressing Ddit4 is sufficient for this inhibition [33]. Therefore, the over 33-fold increase of the Ddit4 transcript in CR mice suggested a reduction of phosphorylated isoforms of substrates of Frap1 in livers of these mice.

To test this hypothesis we performed Western blots on Eif4ebp1 and S6K. Two primary isoforms of S6K1, p70 and p90, were detected in our liver samples. We assayed Thr390 of p70 (Thr412 of p90), and for Eif4ebp1 we assayed Thr69, which are recognized targets of phosphorylation by Frap1. Consistent with the current model of the regulation of Eif4ebp1 by Frap1, Thr69 phosphorylation of Eif4ebp1 was increased more than 2-fold by high calorie feeding relative to CR ( Figure 2 ) (P = 6.0×10⁻⁴, one-sided Wilcoxon rank sum test). In contrast, the Thr390 phosphorylated p70 isoform of S6k1 was undetectable in our liver samples, including those from TAL-fed mice, while Thr412 of p90 was detected but no significant change was found in HIGHCAL relative to CR mice. These results suggest that CR results in reduced efficiency of initiation of cap-dependent translation through phosphorylation of 4EBP1 but does not significantly alter processes dependent upon phosphorylation of S6K1, such as translation of ribosomal transcripts and ribosome biogenesis [34].

To further understand and quantify enriched classes of genes significantly upregulated or downregulated by CR we employed the GO analysis tool FuncAssociate [35], which returns all overrepresented GO terms for a given set of input genes. We separately analyzed upregulated, downregulated, and the combined upregulated and downregulated sets of genes using both permissive and restrictive cutoffs for a total of 6 separate analyses. The number of terms returned for each of the upregulated (U), downregulated (D), and combined upregulated and downregulated (C) sets of genes are as follows: at the restrictive cutoff q<0.01 U = 57, D = 61, and C = 103; at the permissive cutoff q<0.1 U = 94, D = 99, and C = 170. All GO terms and statistics can be seen in Table S3 .

To more clearly understand the changes in biological function that result from changes in gene expression, for each GO term we plotted in a stacked bar graph the relative fraction of upregulated and downregulated genes associated with that GO term from the SAM output (q<0.1). Figure 3 shows a selection (culled to eliminate excessive redundancy) of the GO terms with the largest fractions of upregulated or downregulated genes. CR results in a clear relative upregulation of genes involved in protein biosynthesis, rRNA processing, mRNA metabolism and splicing, ribosome, and regulation of protein translation. Many GO categories consist mostly of downregulated genes including those involved in mitochondria, carboxylic acid metabolism, DNA replication, steroid, cholesterol and lipid metabolism, lysosome, peroxisome, and glutathione S-transferase activity. Several downregulated GO classes are involved in protein turnover including endoplasmic reticulum/ER, isomerase activity, amino acid derivative biosynthesis, protein positioning, and the proteasome.

Since genes involved in hormone biosynthesis are generally downregulated in our data and several genes previously demonstrated to display sexually dimorphic expression are found at the top of our statistical and fold-change rankings, we suspected CR might have an overall feminizing effect on gene expression. To test this hypothesis we compared our data to the male versus female liver expression microarray data of Yang and colleagues which were collected using 334 total microarrays (one microarray per liver sample; 165 male and 169 female mice) [36]. Of the 1,897 genes in our restrictive list, 872 were found in their study to be more highly expressed in either male or female (1,711 of 3,855 genes in our permissive list).

Figure 4 displays the frequencies (at both permissive and restrictive cutoffs) of each of four classes of genes: increased by CR and more highly expressed (UP) in females, increased by CR and UP in males, decreased by CR and displaying lower expression (DOWN) in females, and decreased by CR and DOWN in males. The percentage of genes displaying feminized expression is higher among the set of genes upregulated in CR than among genes downregulated in CR. Separate Chi-square analyses of CR-upregulated genes and CR-downregulated genes at both the restrictive and permissive cutoffs show a significant bias toward a female pattern of expression. We also used FuncAssociate to determine the overrepresented GO category associations of each of these four gene sets. As shown in Table 2 , genes downregulated and masculinized by CR are involved in immune response; genes upregulated and feminized by CR are involved in general metabolism and protein biosynthesis; and genes downregulated and feminized by CR are involved in enzymatic catalysis, and protein modification and transport through the Golgi and endoplasmic reticulum. These results suggest that CR has a generally feminizing effect on gene expression, which involves protein metabolism and turnover, which might contribute to the lifespan extension effects of CR.

Nine-month-old retired breeder ICR male mice were obtained from Harlan Breeders (Indianapolis, IN). Animals were kept in a positively pressurized HEPA air-filtered animal room with 12-hour light/dark cycles. A total of 35 mice were fed 3 specified dietary regimens. Six mice were fed a True Ad Libitum (TAL) diet of which they ate as much as they desired, 12 mice were fed calorie-restricted diets of 73 kcals per week (CR), and 17 mice were fed a diet of 110 kcals per week (Fixed High Calorie, FHC). Recovered food from TAL mice suggested an average consumption plus loss of 149 kcal/week, however, a small amount of the food clearly was lost in bedding, and a more accurate estimate is probably between 120 and 130 kcal/week. Therefore, we estimate the approximate percent calorie reduction for each fixed-calorie diet relative to TAL is 10% to 15% for FHC and 40% to 45% for CR. Calorie reduction of CR relative to FHC is 34%. All mice were given non-acidified water ad libitum throughout regimen treatment.

Mice were fed freshly hydrated food every twelve hours at the beginning and end of the light cycle. Purified diets differing in caloric composition but similar in macronutrient ratios to AIN-93M (Diet No. F05312, Bioserv, Frenchtown, NJ) were assembled from individual ingredients and vitamin and mineral mixes (Bio-Serv, Frenchtown, NJ). These diets consisted of the following approximate caloric composition: 10% fat, 15% protein, and 75% carbohydrate. Animals were fed 12 to 16 hours prior to sacrifice and remaining food was removed after feeding FHC and TAL mice. Animals were sacrificed by rapid decapitation. Tissues were immediately harvested and flash frozen in liquid nitrogen, placed in polypropylene tubes and placed in cryogenic storage for future processing. All animal housing and experiments were performed in compliance with prevailing local, state, and federal regulations of Waltham, MA, USA, and according to the guidelines of Suckow et al. [63] and the Institute for Laboratory Animal Research (ILAR) [64].

To assess the possibility of minimizing expenses over the long-term some RNA samples were combined into pools of 3 prior to cDNA synthesis and hybridization to single microarrays. Pooling of RNA samples for microarray analysis has been reported to reduce intrasample variance and results in averaging of most genes, but does not otherwise affect the analysis [65]. In our experiments all microarrays of pooled samples were analyzed identically with microarrays from individual samples. The 8 CR microarrays consist of 6 microarrays from individual samples and 2 microarrays each produced using pooled RNA samples from 3 other individual mice. The 15 high calorie microarrays consist of 3 individual and 1 pool of TAL samples and 8 individual and 3 pooled FHC samples.

Spotted, long oligonucleotide microarrays representing ∼19,000 genes, duplicates and controls (17,752 non-redundant genes and ESTs) were used to interrogate mRNA abundance (GEO platform accession GPL6761; GEO Samples GSM283874 through GSM283896). The oligonucleotide library was designed by Compugen and manufactured by Sigma-Genosys. Liver samples were taken from cryogenic storage and fragmented in liquid nitrogen. Total RNA was purified by Qiagen RNeasy Mini kit. Briefly, several small pieces totaling 200 mg taken from different locations from an individual liver were homogenized in Qiagen guanidine isothiocyanate lysis buffer, and then subsequent steps were performed according to the Qiagen protocol. All pools of RNA were made by combining equal amounts of each individual RNA sample subsequent to purification and quantification of RNA samples from the individual livers. Microarray design and preparation, sample labeling, and microarray hybridization and scanning services were performed by the Harvard Partners Center for Genetics and Genomics (now Partners Center for Personalized Genetic Medicine). Stratagene Universal Mouse Reference RNA was chosen as a control for two-color hybridization [66]. Fluorescent labels were incorporated directly during cDNA synthesis using the ChipShot direct labeling and cleanup system according to the manufacturer's instructions (Promega, Madison, WI). Total liver RNA was labeled with Cy3 and control RNA was labeled with Cy5. Following dye incorporation and sample cleanup, sample cDNA and control cDNA were mixed together with spiked controls that aid in hybridization calibration, and then the mixed sample was hybridized to the microarray surface using a Maui Hybridization System (Biomicro, Salt Lake City, UT). Microarrays were washed and scanned at 532 nm and 635 nm wavelengths on an Axon 4000B microarray scanner (Axon Instruments, Union City, CA).

Scanned data were collected and processed using GenePix 4.1 software (Axon Instruments, Union City, CA). Genes with median spot intensities below 50 for liver expression for both CR and HIGHCAL were culled from the data set to give a set of 9,550 genes and controls for normalization. These culled data were normalized using the CARMAweb implementation of the limma package for R. Median background was subtracted from median foreground signal using the normexp method followed by loess within-array normalization and then scale between-array normalization [67], [68]. To further eliminate genes near background in these normalized data we set a median expression threshold value of 90 for either CR or HIGHCAL. These adjustments gave a final normalized and culled set of 8,347 genes and controls for statistical testing. For testing of significance we used the software package, Significance Analysis of Microarrays (SAM) and implemented the two class unpaired option [26]. SAM computes and returns a ranked list of significant genes and several statistics, including the modified t-statistic d, and the median false discovery rate (FDR). The FDR increases with the number of genes called significant and is the median number of genes falsely called significant from all permutations divided by the number of genes called significant at a user-specified cutoff. The q-value is the lowest FDR at which a given gene is called significant. For group 1 we chose all 15 microarrays from all high-calorie-fed mice (HIGHCAL) and for group 2 we chose all 8 CR microarrays. We included gene duplicates and controls in the SAM procedure but removed them after, leaving a non-redundant set of 8,347 genes.

For GenAge associations we used only mouse genes and mouse orthologs of human genes listed in GenAge. For human genes we conservatively used only mouse orthologs that share the same function and assigned Entrez gene name listed in the Homologene database (release 59). For GO classification we used the web tool FuncAssociate [35]. FuncAssociate uses a Monte Carlo simulation approach and returns all overrepresented GO classes for a given set of input genes relative to the gene universe from which this subset is derived, along with several descriptive statistics, including P-value based on Fisher's exact test, and adjusted P-value based on 1000 null hypothesis simulations to correct for simultaneous testing of multiple non-independent hypotheses. We analyzed six different sets of genes using FuncAssociate: upregulated, downregulated, and all genes, from both permissive and restrictive data sets. FuncAssociate analyses were run using the unordered input option. The input gene universe for FuncAssociate consisted of all genes on our microarrays, which FuncAssociate culled down to a final set of 17,851 unique genes.

For Western blots mouse livers (100 mg) were homogenized in 1× NuPAGE LDS loading buffer (cat. #NP0007, Invitrogen, Carlsbad, CA) with 1× protease inhibitor mix (cat. #8340, Sigma-Aldrich, St. Louis, MO), 1× Phosphatase Inhibitor Cocktail Set 1 (cat. #524624, Calbiochem, San Diego, CA) and 1× Phosphatase Inhibitor Cocktail Set 2 (cat. #524625, Calbiochem, San Diego, CA). The homogenates were sonicated for 15 seconds to reduce viscosity, and centrifuged at 15,000×g for 10 minutes to remove insoluble material. Samples were subjected to standard SDS-PAGE and probed by Western blot using phospho-specific antibodies against phospho-4E-BP1 (Thr70) (cat. #9455, Cell Signaling Technology, Danvers, MA), phospho-p70 S6 kinase (Thr389) (cat. #9205, Cell Signaling Technology, Danvers, MA) and phospho-p70 S6 kinase (Thr421/Ser424) (cat. # 9204, Cell Signaling Technology, Danvers, MA). Membranes were stripped using Restore Western Blot Stripping Buffer (cat. #21059, Pierce Biotechnology, Inc., Rockford, IL) for 30 minutes at 37°C, and were reprobed with antibodies directed against total p70 S6 kinase (cat. #9202, Cell Signaling Technology, Danvers, MA) or total 4E-BP1 (cat. #9452, Cell Signaling Technology, Danvers, MA). Membrane bound HRP-conjugated secondary antibody was detected using SuperSignal West Femto Maximum Sensitivity Substrate (cat. #34095, Pierce Biotechnology, Inc., Rockford, IL). Quantitation of protein bands was performed using Quantity One software (Bio-Rad Laboratories, Inc., Hercules, CA).

We used quantitative RT-PCR (Q-RT-PCR) to verify microarray results for specific genes of interest. A listing of primers used in Q-RT-PCR assays can be found in Table S4 and all fold-changes can be found in Table S1 . Two control transcripts (Masp2 and D3Ucla1) were chosen; of all transcripts with high but non-saturable expression levels, they were among the transcripts with the lowest variance across all tested regimens. Primers that amplify a predicted amplicon of at least 100 bases and less than 250 bases were favored over others. All of the primers for Q-RT-PCR were designed to span exon junctions so as to avoid amplification of any potential contaminating genomic DNA. Nevertheless, in addition to “no RNA” negative control reactions for each primer, for all RNAs we performed “no RT” PCR reactions (i.e., without reverse transcriptase) in order to ensure that there is no contaminating genomic DNA that would result in amplification products in the Q-RT-PCRs. First strand cDNA synthesis was performed with 10 µg of total RNA, random hexamers, and SuperScriptIII Reverse Transcriptase (Invitrogen).

Each quantitative PCR reaction was carried out in iCycler IQ Real-Time Detection Systems (Bio-Rad), in a total volume of 25 µl, which contained 1× iQ SyBr Green Supermix (BioRad) and 200 nM forward and reverse primers. PCR amplification for each cDNA sample was performed in at least triplicate wells. The quantitative PCR conditions were as follows: 3 min at 95°C, followed by a total of 45 three-temperature cycles (30 sec at 95°C, 30 sec at between 60.5°C and 65.0°C, and 45 sec at 72°C). Relative gene expression data analysis was carried out with the standard curve method [69], [70], and statistical significance of fold-changes was calculated using a t-test, requiring a P-value of at least 0.05 for significance [71], [72].