the health consequences of abnormal metabolic regulation are enormous. These include inborn errors of metabolism that affect specific enzymes of a biochemical reaction and defects in hormone systems, such as diabetes, that coordinate nutrient utilization (59). In addition, changes in modern diet and lifestyle have drastically increased unwanted consequences in body metabolism even in nondiseased condition, as seen in the dramatic increase in obesity and cardiovascular disorders (25, 84). Although some of the metabolic defects can be controlled by appropriate dietary modification, e.g., avoiding food containing high levels of phenylalanine can prevent fatal consequence of phenylketonurea, molecular mechanisms underlying the physiological response to different nutrient conditions are largely unknown. The controversies surrounding the relative efficacy of different diets and diet supplements to battle a wide spectrum of health conditions attest to this lack of knowledge (69). On the other end of the health spectrum, evidence is accumulating that restricting caloric intake can slow down the effects of aging and increase life span in all model organisms tested to date (41, 63). This has been attributed to several factors, including oxidative stress, genome integrity, and endocrine signaling. To understand how diet affects progression of diseases and aging, one of the goals would be to identify the metabolic pathways affected by altered nutrient conditions and the underlying genetic and signaling control mechanisms that regulate them.

The major reactions of the biochemical pathways leading to the metabolism of essential nutrients have been worked out (46). These have come mostly through studying the substrates and products of the reactions and the enzymes catalyzing them. Regulations of the reactions have furthermore focused on the activity and specificity of the enzymes in terms of allosteric control and posttranslational modifications. Thus one way to study changes in flux through a particular metabolic pathway in response to altered diet would be to measure substrate and product concentrations of a particular reaction. This, however, is not yet feasible for multicellular organisms on a large scale. A complementary approach is through analysis of genes that encode these enzymes using recently developed genomic technologies. Although altering enzyme levels may not necessarily result in altering flux through that pathway, significant alterations in the expression of the enzyme-encoding genes may reflect responsive flux changes. In addition, this approach may identify molecules that regulate specific metabolic pathways, such as transcription factors or components of signal transduction cascades. We have already utilized such a strategy to study nutrient-regulated gene regulation in Drosophila (85). Based on these considerations, we have analyzed nutrient-dependent gene expression changes using microarrays in mouse liver.

In mammals the liver plays a key role in coordinating body metabolism in response to dietary conditions. Much of the regulatory effects occur initially in the liver, which then modulates the activities of other organs regarding nutrient utilization and metabolism. We have carried out a comprehensive microarray gene profiling analysis in mouse liver under fasting and under sugar conditions. Numerous studies have utilized microarrays to study nutrient regulation of gene expression, but these differ with respect to the number of genes, type of microarrays, organs tested, and the specific nutrient conditions. Our study provides the most comprehensive gene profiling to date on the patterns of gene expression under fasting and sugar conditions in the same experimental setup, representing some 20,000 different mouse genes. Our results pinpoint several metabolic and signaling pathways that may be part of a global regulatory response to food restriction. These pathways may be important for providing insights into understanding nutrient-dependent disorders. Furthermore, our analysis reveals a connection between starvation response and those which may slow aging. Our results suggest that many of the physiological pathways operating upon starvation underlie those operating during life-span extension brought on by caloric restriction.

MATERIALS AND METHODS

Microarray Production

A total of 20,160 PCR products were amplified from a collection of cDNAs (Array Tag, set A and B) that was made by LION Bioscience, Heidelberg, Germany. These cDNAs (150 to 700 bp) were derived from 3′ end of expressed sequence tags (ESTs) and cloned and sequenced by the company. We amplified the inserts using primers designed from the vector sequence flanking the insert (primer sequence provided by LION Bioscience). The PCR products were cleaned using Millipore 96-well filtration plates (HPPV, 0.45 μm, clear), then transferred into 384-well plates, dried, and resuspended in 1.5 M betaine/1× phosphate printing buffer (150 mM sodium phosphate, pH 8.5). The amplified cDNA products (200–500 ng/μl) were spotted in duplicates in two separate subarrays using a GeneMachines OmniGrid 100 (San Carlos, CA) and TeleChem SMP3 pins (Sunnyvale, CA) on Corning CMT-GAPS II slides (Corning, NY).

Nutrient Conditions

Mice used in this study were 129/SV males, 8–15 wk of age, which were housed individually and received water ad libitum. Twelve animals were taken for the 24 h experiment, in two batches of six animals each (3 fasted, 3 normal fed controls). Total RNA and poly-A RNA were prepared for each sample probe separately, and equal amounts of poly-A RNA from each batch and each condition were pooled, i.e., three animals per pool. A fifth pool was made in which all starved animals were included, as well as a sixth pool that contained all normal controls. Pool 1/2 (normal fed) was hybridized the same manner against pool 3/4 (24 h starved), and pool 5 (starved) was hybridized against pool 6 (normal fed). For experiments on 48 h nutrient condition, 10 normal fed and 10 starved animals were used (distributed in groups of three, three, and four per condition). Pooling was done as described above for each group, and eight arrays were hybridized per group (with dye swap). In the third set up, 10 animals received a sugar solution (40% sucrose and 10% glucose) and water ad libitum, and these were compared with 10 normal fed animals. Pooling and hybridizations were done as described for 48-h starved experiment.

RNA Preparation, Labeling and Hybridization

Mice were killed by CO₂ asphyxiation. The livers were snap frozen in liquid nitrogen and stored at −80°C. Total RNA was prepared using the NucleoSpin RNA L kit (Macherey-Nagel, Dueren, Germany), with an additional step in which the final eluate was used for a subsequent elution step. Poly-A RNA was extracted with the Ambion Purist kit (Austin, TX) according to the manufacturer’s protocols. Equal amounts of poly-A RNA were pooled to generate normal or starved or sugar pools of all animals. These were hybridized against each other on 24 arrays for the 24-h and 48-h starvation experiments each and on 12 arrays for the 48-h sugar experiment. Half of the arrays were hybridized as dye swaps. Labeled cDNA was synthesized from 1.5 μg poly-A RNA using the Amersham direct cDNA labeling kit (Amersham Europe, Freiburg, Germany). Removal of the unincorporated dyes and concentration of the target were done with Microcon 30 spin columns (Millipore, Bedford, MA) according to the manufacturer’s instructions. The concentrated probes were hybridized to the microarray in 1× DIG Easy Hyb buffer (Hoffman-La Roche, Basel, CH) overnight at 42°C.

Microarray Scanning

Arrays were scanned using the Axon model 4000B dual-laser scanner and the corresponding GenePix 4 software (Axon, Union City, CA). Both channels (532 nm for Cy3 and 635 nm Cy5) were scanned in parallel and stored as 16-bit TIFF files. The absolute intensity values span the range from 0 to 65,535. The scans were performed with a resolution of 10 μm. From each spot with a mean diameter of 100 μm, 100 data pixels were recorded. Individual local background area around the spots were defined, which included ∼400 pixels and excludes neighboring spots. For each channel, the raw data was calculated as the median intensity of all foreground pixels with respect to all background pixels.

Each array was scanned three times (low, medium, and high scan) with different signal amplification factors (voltage settings of the photomultiplier tubes), but with the same laser power. The channels for Cy3 and Cy5 were balanced in each scan for approximately the same intensity profile. In the low scan no spot was saturated; in the high scan the signal amplification for Cy5 was set to ∼80% of maximum and the Cy3 amplification was adjusted to this. The settings used in the medium scan lie between the low and the high scan. This method for scanning has several advantages. In the low scan where no spot is in saturation, it is possible to calculate the real ratio for genes with high expression levels; however, those with a low expression level are most likely not recognized. In order not to lose these, a high scan is made; in this case, the information on the saturated spots is lost, so the two scans complement each other. The medium scan produces additional values for subsequent calculations. By scanning the arrays three times, errors which occur while recording and which might increase the error factor in the normalization are averaged.

Data Processing and Normalization

The data processing was automated in a Visual Basic/Excel (Microsoft) program, which integrates the following steps. Raw data are derived from the result files generated by GenePix scanning and picture analysis software. The spot diameter is to lie in the range from 80 to 120 μm, otherwise these spots are not included in analysis. From the pixel-to-pixel ratios between the foreground values of both colors, the standard deviation (SD) is calculated. Those spots that show a ratio SD of higher than 3 are excluded from analysis due to inconsistency. The foreground signal of each spot is corrected by subtracting the corresponding local background. Resulting values that are lower than the background are replaced by the background value. This defines the minimum measurement threshold as the background and avoids calculations of completely wrong ratios, e.g., if the background corrected value for one channel is close to zero. Spots are marked as saturated in one channel if more than 10% of the foreground pixels are at the highest absolute value. From the unsaturated data points of all three scans, linear regression between the low and medium scans and between low and high scans are calculated for both colors. Taking these regressions, the values of saturated spots are replaced by estimated values that reflect the real intensity.

The generated data set of each scan is normalized afterward based on the intensity-dependent methods as described by Yang et al. (80). The ratios are calculated as log₂ transformed values: M = log₂(Cy5/Cy3). Number describing the spot intensity, A = log₂ , is calculated according to Yang et al. (80). A within-pin-tip group Loess fit to the MA plot was made. The Loess scatter plot smoother performs robust local fits using a tricube function to weight the values relative to the median point in the intensity interval. An interval size of 20% was chosen, which corresponds to 180 data points in a pin-tip group containing 30 by 30 spots. After combining the normalized data from all blocks, the data sets of all scans are additionally normalized using a global approach. The normalization was performed so that the sum of all log transformed ratios (M) is 0.

All statistical normalizations are based on the assumption that 1) the majority of genes are not changed in their expression and 2) that the overall up- and downregulations statistically compensate each other in sum. Taking this into account, the standard deviation in the ratios of the stable genes (80% of the most unchanged genes) could be seen as the measurement noise of the array experiments. To be able to compare the different scans and different normalized microarrays, the SD of the stable genes ratios was adjusted to a medium estimated value of 0.2, which is dependent on the used array system. This value does not influence the subsequent statistical analysis because the normal distribution of the data is not changed. For each microarray the normalized and adjusted log ratios of the three scans are averaged. On the used arrays, two identical subarrays are present; thus the data were divided block-wise into two sets. Each set is taken afterward as a distinct element in the statistical test. The data are stored in a relational database using the FileMaker Pro software (FileMaker, Santa Clara, CA).

Statistical Analysis

RNA was hybridized to minimum 12 (for 48 h sugar) replicate microarrays, giving rise to 24 data points for each gene. The criteria for a gene to be considered for statistical analysis in a t-test was that at least 16 of 24 are present (i.e., the data was derived from at least 8 different arrays). In subsequent examinations, only those genes whose ratios placed it in the 99.5% confidence interval were included. RT-PCR analysis was performed for nine genes, and all were consistent with the microarray data. The primer sequences used are listed below; β-actin was used as control. Cog2 was used as additional control, as this showed no regulation from the microarray analysis. These represented both upregulated (Cyp4a14, Ela2, Fkbp5, Gadd45b, Got1, Igfbp1, and Lepr) and downregulated (Car3 and Temt) genes upon starvation, and all but one of the signaling or metabolic pathways are represented by at least one gene. The gene symbols and sequence forward (F) and reverse (R) primers are as follows: β-actin, F-AAGGCCAACCGTGAAAAGATGA/R-TGTCAGCAATGCCTGGGTACAT; Car3, F- TCTGGCCAGTTAGAAAGCCTGTG/R-GTCCGCATACTCCTCCATACCC; Cog2: TTCAGAGTGGACATGGGGACAA/R-CACCAGCCACACTTGTCGATTT; Cyp4a14, F-GCCTGTTCACCCCTCATAACCA/R-CGCCAACCTGCATTTCTACACA; Ela2, F-GTCTCCCTGCAGGTCCTTTCCT/R-GTTGGAGACCCTGGTGAAGACG; Fkbp5, F-GTGTCCATGCATCAAGCCAAAG/R-ATACCAGTCTCCTTGGCCCACA; Gadd45b, F-GAAGGCCTCCGACACTTCTGGT/R-GAGTGGGTCTCAGCGTTCCTCT; Got1, F-CTGACCGTGGTCGGAAAAGAGT/R-TCCAACAGGCTCTAACCCCAGA; Igfbp1, F-ACATCCCTGGATGGAGAAGCTG/R-AGCTTTCCACGTTCAGCTTTGG; Lepr, F-GCCAGTCTTTCCGGAGAATAACC/R-CTGCTGCTCAGGGGATAAGCAC; Temt, F-CTGACTACACCCCGCAGAACCT/R-GCTGTGGAGCAAGGCTTTACAA.

RESULTS

Experimental and Microarray Setup

We performed experiments to identify genes regulated under 24 and 48 h fasting, as well as 48 h only on sugar (Table 1). The sugar treatment was used as a first step to see whether the starvation effect was due to a simple lack of energy source or whether some other nutrient signals, for example, amino acids, could be responsible. For 24 h fast, a total of 12 mice were used (6 normal, 6 starved) and hybridized to 24 arrays. For 48 h fast, a total of 20 mice were used (10 for normal and 10 for starved) and hybridized to 24 arrays. For 48 h sugar fed, a total of 20 animals were used (10 normal and 10 sugar) and hybridized to 12 arrays. Each microarray contains ∼20,000 PCR products representing different genes printed in duplicates (see materials and methods). For each experimental condition, dye swaps were performed (Table 1). Scanning of each array was done three times at three different intensities. Normalization was carried out using Loess (80). For every experiment, each sample class was hybridized to a minimum of 12 replicate microarrays; since there are two subarrays, there are a minimum of 24 data points for each experimental setup. Only those genes with 99.5% confidence level by t-test were used (see materials and methods for details). These data provide a large-scale, statistically reliable view of the changes in gene expression in the mouse liver during fasting and sugar-fed conditions.

The data on which this paper is based have also been deposited with the National Center for Biotechnology Information (NCBI) Gene Expression Omnibus (GEO; http://www.ncbi.nlm.nih.gov/geo/). The sample series GSE853 contains data of the 24 h starvation, GSE852 contains data of the 48 h starvation, and GSE851 contains data of the 48 h sugar experiment.

Overview

As a first step, we compared the number of genes regulated in the three conditions (24 h starvation, 48 h starvation, and 48 h sugar). As expected, the number of genes increased in going from 24 to 48 h of starvation. Furthermore, the number of genes regulated under sugar at 48 h is much lower than starved (Fig. 1). This is not surprising, since sugar feeding is a much less radical alteration in diet than total fasting. Furthermore, most of the regulated genes in starvation no longer become regulated to as high a degree when sugar is present. To discern which of the gene expression changes might have biologically meaningful consequences, we reasoned that significant alterations in flux through a given metabolic pathway may come about through large changes in single key components and/or a series of smaller changes in different components comprising the pathway. To structure the data, we did not use the common approach of clustering based on expression changes. Instead, we grouped the genes in specific metabolic or signaling pathways to obtain a more functional sense of the data.

Fat Metabolism

During fasting, the body breaks down fats to generate energy. As expected, many genes encoding enzymes used in β-oxidation of fatty acids are highly upregulated upon starvation (Table 2). One of the highest upregulated genes in our list encodes a peroxisomal acyl-CoA thioester hydrolase (Pte-2a). This result is consistent with previous results from Northern analysis showing increased expression of Pte-2a upon fasting (28). The gene encoding carnitine acetyltransferase (Crat), which is required for transport of fatty acids into the mitochondria for β-oxidation (30, 34), is also upregulated. There is also upregulation of long-chain acyl-CoA dehydrogenase gene (Acadl) and peroxisomal enoyl-CoA hydratase gene (Ech1) (18, 38), which are involved in β-oxidation.

Concomitant with upregulation of genes involved in fat breakdown, genes encoding enzymes involved in fat synthesis are downregulated. These include stearoyl-CoA desaturase (Scd1), a key lipogenic enzyme for synthesis of monosaturated oleic acid (52). Knockout of Scd1 gene protected mice against adiposity, demonstrating its importance in fat synthesis (51). A special regulatory effect is observed for the Scd1 gene, namely, that it is decreased upon starvation and increased in sugar. There are very few genes that show opposite regulation to this extent in the two nutrient conditions, strengthening the view that Scd1 expression is especially sensitive to sugar and fat metabolism. Other downregulated genes include fatty synthase (Fasn), peroxisomal nudix hydrolase (Nudt7), and butyryl-CoA synthetase 1 (Bucs1) (16, 22).

Recently, the gene for Hyplip1 locus, which causes familial combined hyperlipidemia, was identified as thioredoxin interacting protein (Txnip) (4, 76), and this gene is upregulated upon starvation. Interestingly, the Txnip gene is also dramatically upregulated in glucose-treated pancreas (60). We also see upregulation of leptin receptor and Fsp27, suggesting their function in fat breakdown in the liver (12, 48). Genes involved in lipid transport are also affected (57). Apolipoprotein A-IV (Apoa4) is upregulated, whereas apolipoprotein 2 gene is downregulated, suggesting that the former has a role in fat breakdown, while the latter has a role in fat synthesis. There is also downregulation of the different serum amyloid A genes (Saa), whose products bind and transport HDLs, as well as numerous esterases (Es).

Pxr and Car Nuclear Receptors

The breakdown products of dietary lipids, as well as drugs and other xenobiotics, must be detoxified and eliminated. One signaling cascade that performs these functions comprises nuclear receptors that are activated by lipid ligands and regulate target genes that function in lipid metabolism (9). By a similar token, catabolism of endogenous fats upon starvation appears to operate through the same signaling cascade. For example, the high upregulation of Pte-2a mentioned above is dependent on Pparα, a major fatty acid sensor (28). Although the Ppars are not transcriptionally regulated themselves, the Pxr and Car genes, which act as xenobiotic sensors, show increased expression upon starvation (Table 3). These are among the highest regulated transcription factors. Targets of lipid-activated nuclear receptor include Cyp enzymes, cytosolic binding proteins, and ABC transporters (9), and we also observe strong regulation of these genes (Tables 3 and 4).

Cytochrome P-450s, Cytosolic Binding Proteins, and ABC transporters

A critical family of proteins that regulate lipid metabolism and detoxification is cytochrome P-450s. The expression profiles of the different members of this family represent one of the most striking in our analysis (Table 4). The highest upregulated gene among the 20,000 genes represented, in both 24 and 48 h starvation, is cytochrome P-450 4A14 (Cyp4a14). Disruption of this gene in mice has been shown to cause hypertension (26). Another highly upregulated gene is Cyp17a1, which is involved in dehydroepiandrosterone (DHEA) synthesis from cholesterol (see also Table 7). Among the downregulated genes, cytochrome P-450 2C70 (Cyp2c70) is the second-most downregulated gene in our microarray; little is known about the function of this gene, but its strong downregulation upon starvation suggests a role in lipid synthesis. There are two other cytochrome P-450s that are highly downregulated. Cyp4f14 may play a role in the inactivation of eicosanoids (37), whereas Cyp7b1 is involved in bile acid synthesis (54).

We have also noted a potentially interesting regulation of cytochrome P-450s by aminolevulinic acid. Aminolevulinic acid synthetase (Alas) is the first enzyme in the synthesis of heme, a component of cytochrome P-450 (33). Of the aminolevulinic acid produced in the rat liver, as much as 65% is used for formation of cytochrome P-450s. Thus an increase in Alas would result in increased heme levels required for increased antioxidant and detoxification function of cytochrome P-450s. It is interesting that while the liver-expressed Alas1 is upregulated, Alas2, which in literature is known to be red blood cell specific, is downregulated (Table 4). This could mean that as red blood cells become degraded upon nutrient shortage, the freed amino acids and heme become refunneled for cytochrome P-450 production.

Among the genes encoding the family of 14–15 kb intracellular fatty acid binding proteins, the epidermal-specific lipid binding protein (E-Fabp/Fabp5) (24, 77) is the most strongly downregulated (Table 3). It is expressed in adipocytes and skin, but its role in the liver has not been characterized. There is also small but significant regulation of other fatty acid binding proteins and ABC transporters, including the downregulation of Abcb11, a target of farnesoid receptor Fxr (15). Some of the others may be targets of lipid-activated nuclear receptors Pxr and Car.

Urea Cycle and Amino Acid Metabolism

In addition to fat breakdown, prolonged fasting results in the breakdown of endogenous proteins and amino acids to generate energy and to reallocate available metabolic resources. In keeping with this, we observe gene expression changes that may reflect an increase in flux through amino acid metabolism and urea cycle (Table 5). For example, the lysosomal cysteine protease cathepsin L (Ctsl) gene expression is unchanged at 24 h starvation but becomes upregulated after 48 h starvation. It has already been reported that the activity and protein level of Ctsl is increased in starved rats (29) and that a knockout of this gene results in heart defects (66). This upregulation of Ctsl is completely suppressed by sugar, suggesting that with sufficient energy source, Ctsl targeted proteins need not to be catabolized. We also observe an upregulation of saccharopine dehydrogenase (Aass), which is involved in lysine catabolism: this upregulation is also suppressed in sugar condition. Consistent with this finding, treatment with glucagon has demonstrated increased flux through lysine catabolism pathway (58). There is also downregulation of the gene encoding glutamine synthetase (Glu1), a key enzyme in amino acid metabolism. In Escherichia coli, the enzyme glutamine synthetase is rapidly degraded upon nitrogen starvation (64).

More specifically, we observe a high upregulation of aspartate aminotransferase encoding gene Got1 (also known as glutamic-oxaloacetate deaminase), together with three other genes of the urea cycle, namely, arginase (Arg1), argininosuccinate lyase (Asl), and carbamoyl phosphate synthetase (Fig. 2, Table 5). This most likely reflects the need to utilize amino acids as energy source. Remarkably, the four genes upregulated in starvation are all downregulated in sugar condition. This suggests that in the presence of sugar, the body tries to spare protein degradation, resulting in decreased flux through the urea cycle. This pattern is reminiscent of Scd1 regulation during fat metabolism (see above) and further strengthens the view that flux through the urea cycle is particularly sensitive to nutrient conditions.

We also observe a large downregulation of carbonic anhydrase III gene (Car3). It is in fact the highest downregulated gene in our microarray. The carbonic anhydrases are involved in bicarbonate buffering function. This is necessary to maintain acid-base homeostasis since increased amino acid catabolism would create a pH imbalance. On the other hand, it has been reported that Car3 is found at a much higher level in male than in female mice, in which case the downregulation may reflect a need to suppress synthesis of certain sex hormones.

S-Adenosylmethionine Cycle and Methyl Transferases

Several key enzymes of the S-adenosylmethionine (SAM) cycle are upregulated upon starvation (Fig. 3, Table 6). SAM plays a critical role in methyl transfer reactions, including one carbon metabolism. The gene for methionine-adenosyl transferase (Mat1) is the highest upregulated; this enzyme catalyzes the only known biosynthetic pathway for SAM from methionine and is required for proper liver function (42, 43). The regulation of the gene encoding betaine-homocysteine methyltransferase (Bhmt) seems to be especially sensitive to different nutrient conditions, since there is a large opposite regulation between starvation (upregulated) and sugar condition (downregulated). This is very similar to the behavior of Scd1 during fat metabolism. This regulatory pattern of Bhmt may be important for setting homocysteine levels under various nutrient and health conditions (20). For example, high homocysteine levels have been associated with heart defects (19, 57). In this context, decreased Bhmt (as observed in sugar condition) would lead to higher homocysteine levels, whereas increased Bhmt (as observed during fasting) would lower homocysteine levels.

Several methyltransferases that utilize SAM are also regulated upon starvation (Table 6). Especially interesting is the gene encoding catechol-O-methyltransferease (Comt), which catabolizes several catecholamines in the liver, including epinephrine and norepinephrine. Comt has been implicated in Alzheimer disease and aging, as well as in responding to a pain stressor (86). In addition, Comt inhibitors have been used as treatment for Parkinson and other age-related diseases (78, 83) There is also a strong downregulation of Temt (thioether S-methyltransferase), an important enzyme involved in detoxification of selenium- and sulfur-containing compounds (47). An evolutionarily related gene for the enzyme nicotinamide N-methyltransferase (Nnmt), which methylates nicotinamide, is also regulated during starvation and sugar.

Cholesterol and DHEA Regulation

DHEA is a cholesterol-derived hormone, which together with its sulfated ester, DHEA-S, is present at the highest level of any circulating hormone in humans (1, 81). Our analysis reveals a striking regulatory pattern concerning DHEA metabolism (Fig. 4, Table 7). First, genes involved in synthesis of cholesterol from acetyl-CoA are downregulated (including the cytoplasmic form of HMG-CoA synthase, which produces HMG-CoA). Second, the key gene in DHEA biosynthesis from cholesterol, Cyp17a1 (steroid-17α-monooxygenase), is highly upregulated. Third, genes involved in converting DHEA to other compounds, such as bile acid and other steroids, are downregulated (Table 7). Fourth, there is a downregulation of Cyp7b1, which is involved in the breakdown of cholesterol to bile acid, and the corticosteroid binding globulin protein (CBG). In sum, the goal of these regulatory changes during fasting appears to be to maintain as high a level of DHEA as possible, both by increasing DHEA production from the available cholesterol pool and by decreasing DHEA conversion to other compounds.

As an aside, we note that in contrast to the cytoplasmic form, the mitochondrial form of HMG-CoA synthase is increased in starvation. The latter form is used in ketogenesis to make ketone bodies from acetyl-CoA derived from fatty acid catabolism. This makes physiological sense, since ketone bodies provide fuel for the brain during starvation. It has further been reported that the mitochondrial HMG-CoA synthase is directly regulated by forkhead transcription factor (FKHRL1) via the insulin signaling pathway (50).

IGFBP and IGF Regulation

One of the most striking regulations is observed for insulin-like growth factor binding protein 1 (Igfbp1) (Fig. 5, Table 8). It is unchanged at 24 h starvation but becomes highly upregulated after 48 h. This upregulation is almost completely suppressed by sugar feeding. Igfbps function by binding insulin-like growth factors (Igfs), thereby interfering with Igf activity on target organs (35). Igfbp1 is the major form that controls body growth. Thus it makes physiological sense that starvation effects upregulation of Igfbp1, thereby signaling stoppage of cellular growth upon nutrient limitation. There is also a downregulation of Igf1, the cognate binding factor of Igfbp1. The results of other Igf/Igfbp members are shown in Table 8. The Igfbp1 signaling pathway is used not only during nutrient deprivation but in other processes that require stoppage of growth, including liver regeneration (67). This is probably because global body growth must be halted during the biosynthetic process of regenerating an organ. We have also seen a marked decrease in growth hormone (GH) gene expression in the brain upon starvation (unpublished data). This further reflects a coordinated alteration in growth factor signaling program (Fig. 5).

DNA Repair and Apoptosis

In addition to changes in metabolic and hormonal systems in the face of nutrient shortage, animals must also stop growth, since attempted proliferation in the absence of necessary biosynthetic resources would be detrimental to survival. Various genes that may function to maintain genome integrity or coordinate cell proliferation and apoptosis are listed in Table 9. A highly upregulated gene is Gadd45b, which functions in growth arrest and is also inducible by DNA damage, genotoxic stress, and TGFβ signals (73, 82). As with Igfbp1, Gadd45s are also upregulated during liver regeneration (67). The LIM-only transcriptional regulator Lmo4 is implicated in proliferation control and is also upregulated (highest upregulated transcription factor in our microarray), as is the radiation-induced cysteine-rich LIM protein gene Rad51L1 (61, 75). Other upregulated genes implicated in apoptosis include a Bcl-2 family member, Bnip3 (5, 6, 53, 74), and Creg (cellular repressor of E1A suppressed genes), which encodes a glycoprotein involved in growth inhibition and binds to insulin-like growth factor receptor Igf2r (Table 8) (14). There is also strong downregulation SMP-30/regucalcin (Rgn), which has been implicated in calcium homeostasis and growth control (21, 49).

DISCUSSION

Connection Between Physiological Pathways Underlying Starvation Response and Those Regulating Life Span

In a large-scale gene expression profile screen in mouse liver, we have identified several metabolic and signaling pathways whose components are altered transcriptionally in starved and sugar-fed conditions. Although the microarray approach we used cannot provide a direct link between changed RNA levels and influence on protein activity or metabolic flux, we observe a strong correlation between these pathways and those which have been shown to be important for longevity. We outline below the similarities between starvation response and life-span-prolonging processes.

Lipid catabolism and activation of the oxidative stress response.

The need for the body to break down lipids can come externally through dietary intake or internally through breakdown of endogenous lipids during fasting. We have observed a large number of genes regulated upon starvation which are part of this lipid signaling cascade (9). The relevance for this cascade for aging is clear: it regulates a battery of antioxidant factors, including nuclear receptors and cytochrome P-450s, for detoxification and protection from oxidative and xenobiotic stress. Since oxidative stress is one of the major sources thought to affect the aging process, factors that increase protection from oxidative stress should slow down aging. In C. elegans for example, the daf9 and daf12 genes, which affect life span, encode a cytochrome P-450 and a nuclear receptor with similarity to Pxr, respectively (3, 23). In Drosophila, it has recently been shown that the steroid hormone ecdysone, acting through its nuclear receptor, controls aging (62). Thus the lipid-activated nuclear receptor pathways would bring about increased resistance to oxidative stress, thereby increasing survival and longevity.

Regulation of central metabolic pathways: SAM and urea cycles.

SAM, the central factor in methylation reactions in the cell, has long been implicated in cellular aging through regulation of various methyltransferases that participate in protein repair (10, 11, 64). It is interesting that the overall regulatory alterations in SAM cycle upon starvation would have the effect of lowering homocysteine levels, since high homocysteine levels have been associated with increased chance for coronary diseases (19, 57). It has also been recently shown that in long-lived Ames dwarf mice the activities of Mat1 and Gnmt are elevated (72); since both of these genes are upregulated upon starvation, it is consistent with the notion that increased flux through the SAM cycle is associated with life-span extension and starvation response. Furthermore, the methyltransferase Comt, which is downregulated in starvation, has been implicated in Alzheimer disease and its inhibitor is used to treat Parkinson disease (78).

In contrast to the association of methyltransferase reactions with protein damage repair and aging, little has been documented on the potential role of urea cycle and aging. There is, however, one intriguing observation. It has recently been shown in Drosophila that feeding sodium 4-phenylbutyrate (PBA) can extend life span (32), and PBA is a FDA-approved drug for treatment of urea cycle disorder (7). Another potential connection comes from a recent study showing that a gene involved in autophagy can extend life span in C. elegans (45). Since autophagy entails degradation of internal proteins in the face of nutrient deprivation, it will likely increase flux through the urea cycle.

Regulation of hormonal pathways: DHEA and insulin/Igf signaling.

The level of DHEA and its sulfated derivative steadily decreases with age, a fact which underpins many discussions on its role as an anti-aging substance. The role of DHEA is controversial, but it has been implicated in a variety of processes associated with youth and virility (1, 44, 81). Our analysis indicates that upon starvation, the body tries to maintain as high a level of DHEA as possible: the gene for DHEA synthesis from cholesterol is highly upregulated, whereas those which are involved in making other steroid hormones are downregulated. This further suggests a connection between starvation response and an anti-aging process.

There is substantial evidence for insulin/IGF signaling playing a fundamental role in aging. This comes from mammalian studies, as well as worms and flies (17, 36, 41). In these cases, a decrease in insulin signaling is correlated with increased life span. Recent studies further suggest that insulin may increase the chance of Alzheimer disease (70). Consistent with our view that starvation response shares signaling pathways with those underlying anti-aging processes, there is a large decrease in insulin signaling upon starvation as illustrated by the large increase in Igfbp1 expression. The decreased insulin signaling could also affect cytochrome P-450 activity through Alas, since it has been reported that insulin inhibits Alas in a liver-derived cell line (56). Therefore, upon starvation decreased insulin level would de-repress expression of Alas, which is what is observed, leading to increased cytochrome P-450 synthesis, thereby connecting the cytochrome P-450 pathway with the insulin pathway.

Genome instability and apoptosis.

Genes involved in maintaining genome integrity, such as DNA repair, play an important role in aging (31). The NAD-dependent histone deacetylase Sir2 in yeast and the histone deacetylase Rpd3 in Drosophila have been shown to regulate aging (39, 55). Furthermore, yeast life span can be increased by changing the level of NAD through increased Nnmt (nicotinamide N-methyltransferase) activity (2), and we also see an increase in Nnmt upon starvation. It has also been suggested that certain compounds that extend life span in yeast through Sir2 operate by suppressing p53 and delaying apoptosis (27). In this respect, we note that Rad51L1, which is upregulated in starvation, is a vital, radiation-induced gene which may function through interaction with p53 (61). In addition, Starai et al. (65) have pointed out a link between Sir2 function and lipid metabolism through the activation of acetyl-CoA synthetase. Thus many of the metabolic response to starvation may be mediated by factors that influence aging through their function in genome stability and cell survival.

Regulatory Overlap Between Starvation Response and Caloric Restriction

The one documented regimen for prolonging life span in various organisms has been caloric restriction (41, 63). Since starvation can be viewed as the most extreme form of caloric restriction, it is not surprising that many of the connections made here between starvation and longevity have been previously observed for caloric restriction and longevity. It is therefore also not surprising that life-span extension is associated with increased survival under starvation stress in different organisms (40). Thus one can ask whether the regulatory mechanisms that operate during starvation also operate during caloric restriction at the gene expression level. We indeed find informative overlaps between genes regulated for liver in caloric-restricted mice and our current analysis. For example, carbamoyl phosphate synthetase of the urea cycle is upregulated upon starvation and caloric restriction, whereas glutamine synthetase is downregulated in both cases (13, 71, 79). Furthermore, in a microarray analysis of calorically restricted mouse liver (8), 33 genes were identified as having caloric-restriction-specific changes. Of the 28 genes that are represented in our microarray, 19 of these are regulated in the same direction, 5 in the opposite direction, and 4 are unchanged. Of the 24 regulated genes, 9 can be found in the list of genes representing the signaling or metabolic pathways highlighted in our starvation condition, and 8 of the 9 genes are regulated in the same direction as in caloric restriction (Got1, Fasn, Gamt, Cypa13, Cyp4a14, Ese1, Cbg, and Rgn; the one exception is Apoa4). Furthermore, the eight genes are distributed over the different pathways and are not restricted to any specific pathway. This comparison reveals a significant overlap of the expression profile between starvation and caloric-restricted conditions, although it should be emphasized that the comparison is based on a single study, and further independent data will be required. On the other hand, although caloric restriction can extend life span, prolonged starvation is life threatening. What could then be the common theme linking starvation response to caloric restriction? We favor the view that it is the degree of response to stress brought upon by the two conditions. Caloric restriction might invoke a low-level stress response, which in the long run will have a protective function and increase survival; starvation would invoke a similar response, but at a much higher level and which cannot be sustained for long periods. The genetic regulatory patterns that we have found in starvation could therefore lead to a better understanding of anti-aging mechanisms as well as global biological functions such as interactions between antioxidative reactions, xenobiotic protection and maintenance of genome integrity.

GRANTS

This work was supported by the Forschungszentrum Karlsruhe and Bundesministerium fuer Bildung und Forschung.