Transcriptional profiling identifies strain-specific effects of caloric restriction and opposite responses in human and mouse white adipose tissue

Tissue and strain impact the number of differentially expressed genes identified in comparisons between CR and control mice

Gene expression data was obtained from a previously published study [37] evaluating the effects of a short-term (14 week) CR diet in males starting at 8 weeks of age. For each of 7 strains, CR-fed mice were provided a diet at least 23% less per week by weight compared to CTL mice (range: 23 – 42%; see Table S1 and Methods). Genes differentially expressed between CR-fed and ad lib-fed mice were identified with respect to the 7 mouse strains and 4 tissue types. The number of differentially expressed genes (DEGs) varied among strains and tissues, with no DEGs identified in some strains/tissues but as many as 984 DEGs identified with respect to eWAT from F1 mice (FDR < 0.10, FC > 1.50 or FC < 0.67) (Figures S1 and S2). Effects of CR were strongest in eWAT with the largest number of DEGs identified in eWAT for 5 of 7 strains (exceptions: B6 and BALB/c mice). In eWAT, the largest number of DEGs was identified in F1 mice (984), whereas in muscle and heart most DEGs were identified in C57BL/6 (170 and 557, respectively). No DEGs were identified with respect to cortex except 3 DEGs were increased by CR in CBA mice. Among the 7 strains, the fewest DEGs was usually identified in BALB/c mice and altogether only 13 DEGs were identified in BALB/c mice among the 4 tissues. The magnitude of CR’s effects thus varied considerably among the 28 strain/tissue combinations evaluated.

Genome-wide transcriptional responses to CR in C57BL/6 mice are weakly correlated with those in other mouse strains (muscle, heart and cortex)

Gene expression variation among samples from the same tissue could be explained by strain or diet, and the relative role of these two factors was examined using linear models and likelihood ratio tests. Strain tended to explain a larger fraction of variation in gene expression as compared to diet in all 4 tissues (Figure 1A). Overall, strain was the dominant factor explaining gene expression variation for 64% (eWAT), 73% (muscle), 89% (heart) and 94% (cortex) of protein-coding genes (Figure 1B). The comparatively stronger effect of diet in eWAT could be seen from analysis of CR response vectors in principal component space (Figures 1C – 1F). eWAT response vectors were consistent among strains in terms of length and direction (Figure 1C). However, for muscle, heart and cortex, CR response vectors differed in both length and direction among the 7 strains (Figures 1D – 1F). These patterns were reinforced by visual comparisons of CR response patterns in self-organizing maps (SOMs). In eWAT, CR responses were stronger and more consistent among strains, whereas in muscle, heart and cortex, CR responses were weaker and varied more among strains (Figure 1G).

The genome-wide correlation between FC estimates among strains was strongest in eWAT but weaker in muscle, heart and cortex (Figures 2A – 2D). In heart and cortex, genome-wide expression responses to CR were negatively correlated for some strain pairs, indicating that many genes were oppositely affected by CR depending on strain (Figures 2C and 2D). Several negative correlation estimates were obtained in pairings that involved the C57BL/6 strain (Figures 2C, 2D and 2F). In heart, for example, the correlation between CR responses in C57BL/6 and F1 mice was -0.29, and likewise, the correlation in C57BL/6 and CBA mice was -0.25 (Figure 2F). When strains were clustered based upon genome-wide correlations among expression responses, the C57BL/6 response was only weakly correlated with that of other strains for muscle, heart and cortex (Figure 2E). For these 3 tissues, expression response correlations between C57BL/6 mice and other strains did not exceed 0.24 (Figure 2F).

Side-by-side heatmap comparison of clustered CR responses confirmed better agreement among strains in eWAT, although groups of genes from each tissue could be identified as having strain-specific CR responses (Figure 3A). We analyzed expression of genes previously identified by meta-analysis studies to be similarly altered by CR in multiple strains and tissues [47,48], but found that many strains/tissues did not demonstrate expected trends for such genes (Figures 3B and 3C). For instance, genes expected to be increased by CR were decreased in heart from F1 mice, while genes expected to be decreased by CR were increased in muscle from 129 mice (Figures 3B and 3C). Finally, a summary visual presentation of CR responses was generated using Chernoff faces [49], with facial features scaled and colored according to principal components extracted from a matrix with FC estimates from all expressed protein-coding genes for the 28 strain/tissue combinations (Figure 3D). This multivariate presentation identified strain/tissue combinations with the most aberrant genome-wide CR responses (e.g., B6-muscle, 129-muscle, B6-heart, F1-cortex and C3H-cortex; Figure 3D).

The number of genes with similar CR responses across strains is larger than expected and primarily involves genes with metabolic functions

Global analysis of gene expression provided evidence for both shared and strain-specific CR responses (Figures 1 – 3). Shared responses to CR among strains or tissues are of special interest, since these may provide the most useful biomarkers or may be conserved among species (and thus replicable in humans) [47,48]. To evaluate shared CR responses, we identified genes similarly altered by CR across strains based upon a nominal p-value threshold of 0.05 per strain (Figure S3). This showed that for all tissues the number of similar CR responses in 6 or all 7 strains was significantly larger than expected under a null hypothesis in which CR responses are randomly associated among strains. A total of 316, 33 and 11 genes were similarly altered by CR in all 7 strains for eWAT, muscle and heart, respectively (P < 0.05 for each strain; Figures S3A – S3C). No genes were similarly altered by CR in cortex from all 7 strains, although a larger-than-expected number were similarly altered by CR in 2 – 6 out of the 7 strains (Figure S3J and S3K). Gene ontology analyses were performed to identify functional themes among those genes most consistently altered by CR across multiple strains (Figure S4). In eWAT, genes increased by CR in all 7 strains were associated with various metabolic functions, while genes decreased by CR in all strains included other metabolic pathways such as gluconeogenesis and ketone metabolism as well as angiogenesis, and response to insulin (Figures S4A and S4B). In other tissues, genes most consistently altered by CR among strains were predominantly associated with metabolism, translation, and/or RNA splicing/metabolism (Figures S4C – S4H).

C57BL/6 mice have the largest number of strain-specific CR responses including decreased expression of IGF-1 pathway genes in heart

To characterize strain-specific CR responses, linear models were used to identify genes demonstrating strain-by-diet interaction effects, in which responses to CR differed in a given strain as compared to the 6 other strains. In all tissues, the largest number strain-by-diet interaction effects was identified in C57BL/6 mice (Figure S5). Applying an FDR < 0.10 threshold, a total of 256, 833 and 1805 genes exhibited significant interactions in C57BL/6 eWAT, muscle and heart, respectively (Figure S5A – S5D). At this FDR threshold, only 1 gene demonstrated strain-by-diet interaction in cortex of C57BL/6 mice (Sorcs3), but 1074 genes were identified at a weaker P < 0.05 threshold, which was a larger number than any of the other 6 strains (Figure S5D).

The largest number of C57BL/6-specific CR responses were identified in heart and muscle, consistent with global analyses of gene expression (Figures 1 – 3). In heart, unique CR responses of C57BL/6 mice included increased expression of genes associated with inactivity or denervation (Figure S5G) and decreased expression of genes connected to IGF-1 signaling (e.g., Igfbp5, Igfbp4, Igf1; Figure S5H). Loss of Igf1 expression in C57BL/6 mice was likely an acute effect or immediate response to CR, since decreased Igf1 can also be demonstrated from a prior microarray study (GSE68646) of cardiac tissue in young (10-12 week) C57BL/6 males subjected to only 1 week of 30% CR (Figure S6) [50]. Top-ranked increased genes in C57BL/6 heart were Cyfip2, Ptgds and Scn5a, while top-ranked decreased genes were Bmp10, Ucp1 and Myl7 (Figures S7A – S7D). In muscle, C57BL/6-unique responses to CR included increased expression of genes related to development (Figure S5I) and decreased expression of genes with metabolic functions (Figure S5J). The top-ranked increased genes in muscle were Nnat, Lep and Tnni1, and similarly the top-ranked decreased genes were Otub2, Aldh1a2 and Ndufb2 (Figures S7E – S7H)).

Shared and strain-specific CR responses among genes belonging to longevity-regulating pathways

Pathways mediating longevity responses to CR include those linked to insulin-like growth factor I (IGF-1), sirtuins, and/or target of rapamycin (TOR) [51]. The KEGG longevity-regulating pathways provide one conceptual model for how such pathways interact and defines connections between genes and pathway components (KEGG identifiers hsa04211 and hsa04213) [52]. We identified genes associated with the KEGG longevity-regulating pathways most frequently altered by CR across the 7 strains and 4 tissues (Figure 4). Genes associated with these pathways were often consistently altered across strains, particularly with respect to eWAT (e.g., Adcy5, Cat, Sesn2, Sod2, Eif4e2; Figure 4A and 4B). Expression of catalase (Cat) was also consistently down-regulated by CR across strains in muscle and heart (Figures 4A, 4C, 4D and 4G). On the other hand, there were many instances in which longevity-regulating genes were altered in some strains but not others (Figure 4A). Genes with divergent response patterns among strains were in some cases linked to the insulin/IGF-1 system, such as Insr (eWAT), Irs2 (muscle) and Igf1 (heart) (Figures 4H – 4J).

Olfactory receptor gene expression is uniquely elevated by CR in F1 heart and C3H muscle

Global analysis of CR responses provided evidence for a group of genes uniquely up-regulated by CR in C3H muscle and F1 heart (Figures 1G and 3A). These strain/tissue combinations also exhibited the second-largest number of genes with strain-by-diet interaction effects (Figures S5B and S5C). Closer inspection revealed that nearly all genes most specifically up-regulated by CR in these cases encoded olfactory receptors. In F1 heart, for example, 21 of the 22 genes most specifically up-regulated by CR encoded olfactory receptors (Figure 5A). Those genes demonstrating the strongest patterns, such as Olfr1184, Olfr910 and Olfr488, were uniquely elevated by CR in F1 heart but were also expressed at higher levels in F1 mice regardless of diet (Figures 5B – 5D). Overall, 90% of genes encoding olfactory receptors were elevated by CR in F1 heart tissue with most genes not similarly altered in other strains (Figure 5G). This pattern was unique to F1 heart tissue and not observed in eWAT, muscle or cortex (Figures 5E, 5F and 5H). The set of olfactory receptor genes elevated in F1 heart were similarly elevated in C3H muscle but the same trends were not observed in other strain/tissue combinations (Figure 5I).

Transcriptional responses to CR in mouse eWAT are negatively correlated with those in human scWAT

The laboratory mouse is a cornerstone for experimental aging research although humans and mice may differ in their CR diet responses [53]. In humans, there is no WAT depot that is strictly analogous to mouse eWAT [54], but several studies have evaluated gene expression responses to CR in scWAT [38–46]. We therefore compared CR responses in eWAT from mice to those observed for orthologous genes in scWAT samples from 28 experiments in which human subjects followed a CR diet (Table S2).

In 25 of 28 experiments, gene expression responses in most or all mouse strains were negatively correlated with human CR responses (Figure 6A). The average correlation across 28 experiments was negative for each strain (-0.115 ≤ r_s ≤ -0.037) but was least negative with respect to C57BL/6 mice (r_s = -0.037) (Figure 6B). Consistent with this, CR responses in each strain were negatively correlated with the average meta-response observed across the 28 human experiments (Figure 6C). The 100 genes most strongly increased by CR in each mouse strain were significantly enriched with respect to CR-decreased genes from humans (Figure 6D; P < 0.01 for all strains), although there were no significant trends with respect to the 100 genes most strongly decreased by CR in each mouse strain (Figure 6E; P ≥ 0.248). Interestingly, genes increased by CR in eWAT from multiple mouse strains were more strongly decreased by CR in human scWAT (Figure 6F). Genes decreased by CR in all 7 mouse strains were conversely increased by CR in humans, although this trend was non-significant (Figure 6G). Among genes most strongly altered by CR in humans, some exhibited consistent trends in mouse (Tce3, Sncg, Gpx1) but others were oppositely regulated by CR (Kmt2a, Hmbs, Dhcr7) (Figures 6H and 6I). None of the genes most consistently increased by CR across strains were similarly altered in humans (e.g., Atox1, Chchd4, Mrpl34; Figure 6J), and likewise, most genes consistently decreased by CR across strains were not similarly altered in humans (e.g., Pfkfb3, Ecm1, Tns2; Figure 6K).

To provide finer-scale comparison of mouse and human CR responses, correspondence was evaluated with respect to Gene Ontology (GO) Biological Process (BP) terms most strongly enriched among genes robustly elevated by CR in human scWAT (Figure 7A). This identified processes for which associated genes were increased by CR in humans and decreased by CR in most mouse strains, e.g., blood vessel remodeling (Figure 7B), stem cell population maintenance (Figure 7C), biosynthetic process regulation (Figure 7D) and aging (Figure 7E). Among 70 genes increased by CR in humans and mouse strains (Figure 7F), enriched GO BP terms included oxidation-reduction process and biosynthesis of small molecules, carboxylic acid and cofactors (Figure 7H). Likewise, among 115 genes increased by CR in humans but decreased in mouse strains (Figure 7G), enriched GO BP terms included protein heterotrimerization, endocrine hormone secretion and plasma lipoprotein particle (Figure 7I).

GO BP terms most strongly associated with genes decreased by CR in humans were identified and a term-by-term comparison was made to mouse CR responses (Figure 8A). Genes associated with platelet aggregation were down-regulated by CR with respect to both humans and mice (Figure 8A). However, contrasting patterns were observed with respect to genes associated with response to oxidative stress (Figure 8B), glucose metabolic process (Figure 8C), apoptotic mitochondrial changes (Figure 8D), and cofactor biosynthesis (Figure 8E). Among 74 genes decreased by CR in both humans and mice, there was significant enrichment with respect to GO BP terms linked to regulation of insulin response, protein metabolism and monosaccharide transport (Figure 8H). Likewise, among 233 genes decreased by CR in humans but increased in mice, we noted significant enrichment with respect to energy derivation by oxidation and metabolism of cofactors, oxidoreduction coenzymes, pyridine, nucleobase small molecules, organic acids, ribonucleoside triphosphate, purine triphosphate and purine monophosphate (Figure 8I).

Effects of CR on the expression of genes related to stem cell maintenance, blood vessel remodeling and lipid metabolism in human scWAT are not replicated in 4 mouse WAT depots

We noted an inverse association between CR responses in mouse eWAT and human scWAT (Figure 6). These are the most commonly studied forms of WAT in mice and humans, respectively, although potentially CR responses in other mouse depots would correspond better to those in human scWAT [54]. To evaluate this possibility, we compared CR responses in human scWAT and 4 WAT depots from obese C57BL/6 male mice (eWAT, scWAT, perirenal [prWAT], mesenteric [mesWAT]). In these experiments, samples were obtained from mice fed a high fat diet with 25% CR enforced for 1 – 60 days (see GEO series accession GSE30534). Applying a strict FDR threshold of 0.10, we did not identify any significant CR-regulated genes. However, given a weaker threshold of P < 0.05 (with FC > 1.50 or FC < 0.67), we identified between 4 and 403 genes with CR-regulated expression, depending upon the WAT depot examined and the CR duration (Figures S8 and S9).

Consistent with the above findings, CR responses in C57BL/6 eWAT were negatively correlated with those in human scWAT (Figure S10A and S10B). We expected improved correspondence when comparing scWAT responses in both species, but the average correlation was still negative on average across the 28 human experiments (r_s = -0.016 for the 60 day CR response; Figure S10B). Moreover, genes associated with stem cell maintenance and blood vessel remodeling, although increased by CR in human scWAT (FDR < 0.10), were not correspondingly altered in mouse scWAT (P > 0.05 for all genes; Figures S11A, S11B and S11E). Likewise, genes related to hydrogen ion membrane transport and neutral lipid metabolism decreased by CR in human scWAT (FDR < 0.10) were not correspondingly altered in mouse scWAT (P > 0.05 for all genes; Figures S11C, S11D and S11F). Overall, CR responses in mouse prWAT best correlated with those in human scWAT, with positive correlations on average calculated for 5 of the 7 time points evaluated (Figure S10B). Consistent with this, the 100 genes most strongly increased by CR in C57BL/6J prWAT overlapped significantly with genes increased by CR in human scWAT (Figures S10C and S10D). Conversely, the 100 genes most strongly decreased by CR in C57BL/6J prWAT overlapped significantly with genes decreased by CR in human scWAT (Figures S10E and S10F).

Microarray analysis of gene expression responses to CR in 7 mouse strains

The complete dataset included 448 microarray samples available from the Gene Expression Omnibus (GEO) database under the series accession GSE75574 (7 strains × 4 tissues × 2 diets × 8 mice per treatment group = 448 samples total) [37]. A complete description of animal husbandry and dietary protocols has been reported previously [37]. In brief, male mice were purchased from the Jackson lab at 6 weeks of age and provided AIN93M rodent diet (84 kcal/week). Mice were randomized to either CR or ad lib diet treatments starting at 8 weeks of age with sacrifice and tissue collection at 22 weeks. For the Balbc/J, C3H/HeJ, CBA/J, DBA/2J and B6C3F1/J strains, CR-fed mice received a dietary allowance reduced in weight by 23 – 25% compared to CTL-fed mice (Table S1). For CR-fed C57BL/6J and 129S1/SvImJ mice, the dietary allowance was initially reduced by 25% (weeks 8 – 14) and then decreased further by 42% (weeks 14 – 22) to prompt weight loss similar to that in other strains (Table S1) [37]. Regarding body weights at week 22, effects of CR were similar with 16 – 24% reduction compared to CTL, although Balbc/J was an exception with average weight of CR-fed mice reduced by only 4% (Table S1). At the conclusion of dietary interventions (week 22), mice were fasted overnight prior to sacrifice by cervical dislocation [37]. The 4 tissues collected for RNA analysis were eWAT, skeletal muscle (gastrocnemius), heart and brain neocortex.

Data processing and normalization

Gene expression profiling was performed using the high-density Affymetrix Mouse Gene 1.0 ST array platform for whole transcript expression analysis. Material from one biological sample was hybridized to a single microarray, yielding a total of 448 raw CEL files downloaded under the GSE75574 accession. Raw CEL files were normalized using the robust multichip average (RMA) algorithm for gene-level expression intensity summarization (R package: oligo; function: RMA) [85]. Expression estimates were calculated for 35556 probe identifiers, of which 20353 could be associated with a protein-coding gene (R annotation package: mogene10sttranscriptcluster.db). These 20353 probe identifiers were collectively associated with 18763 unique protein-coding genes. To limit redundancy in downstream analyses, a single probe identifier was chosen to represent each protein-coding gene [86]. For each tissue, the probe identifier with highest average expression across all samples was chosen as the representative in cases where multiple probe identifiers were associated with the same gene symbol. Following these steps, there remained 18763 probe identifiers, each of which was associated with a unique gene symbol.

Microarray quality control

The initial set of 448 array samples was analyzed to identify outliers or poor-quality hybridizations. Microarray samples for each tissue were clustered and each sample was plotted with respect to the first two principal components calculated from genome-wide expression intensity estimates (Figures S12 – S15). Additionally, the probe-level model metrics normalized unscaled standard error (NUSE) and relative log expression (RLE) were calculated for each sample, with low-quality samples suggested by NUSE median estimates differing substantially from 1 or RLE median estimates differing from 0 (Figures S12 – S15) [87]. A total of 12 samples were removed because they were identified as outliers or probe-level metrics suggested poor-quality array hybridizations (excluded eWAT samples: GSM1959516, GSM1959551; excluded muscle samples: GSM1959292, GSM1959322, GSM1959354, GSM1959359; excluded heart samples: GSM1959406, GSM1959427, GSM1959457, GSM1959468; excluded cortex samples: GSM1959158, GSM1959224). Thus, the final dataset used for subsequent analyses included 436 microarray samples among the 4 tissues (448 – 12 = 436 samples).

Differential expression analyses

Differential expression analyses were performed to evaluate the effect of diet (CR versus ad lib) with respect to each strain and tissue (7 strains × 4 tissues = 28 strain-tissue combinations total). For a given sample, a gene was determined to have detectable expression if at least 50% of probe sets associated with that gene were expressed above-background at a significance level of P < 0.05, based upon a signed-rank test comparison between perfect match and mismatch probe intensities [88]. Differential expression testing was performed using only genes with detectable expression in at least 33% of the microarray samples available for a given CR versus ad lib comparison. Among the 28 strain-tissue combinations, this yielded between 14083 and 17883 protein-coding genes for inclusion in differential expression analyses (average of 16474 genes). A total of 15490, 15375, 13588 and 17406 genes met criteria for all comparisons with respect to eWAT, muscle, heart and cortex, respectively. Likewise, 13129 protein-coding genes were consistently expressed in samples from each tissue type and included in analyses in each of the 28 strain-tissue combinations. Differential expression between CR and ad lib samples was evaluated using empirical Bayes linear models and moderated t statistics as implemented in the R limma package [89]. To control the false discovery rate for each CR versus ad lib comparison, raw p-values were adjusted using the Benjamini-Hochberg method [90]. Unless otherwise indicated, genes were considered to be differentially expressed if the FDR was less than 0.10 and estimated fold-change (FC) was greater than 1.50 or less than 0.67. To evaluate the accuracy of our analyses, we compared microarray FC estimates to those previously reported [37] and obtained using RT-PCR for 11, 10 and 7 genes in eWAT, muscle and heart, respectively (C57BL/6J strain). As expected, microarray-based FC estimates were less extreme than those obtained by RT-PCR, likely reflecting reduced dynamic range of microarrays compared to PCR assays [91]. However, for each tissue examined, we noted a strong positive correlation between FC estimates obtained by microarray and RT-PCR (r ≥ 0.86; Figure S16).

Strain-by-diet interaction effects

Linear models with moderated t statistics were also used to identify genes with strain-dependent CR responses (i.e., strain-by-diet interaction effects). For this approach, gene expression was modeled as a function of diet, strain and an interaction term (i.e., Expression = Diet + Strain + Diet*Strain) with analyses replicated for each of the 28 strain-tissue combinations. Diet was defined as a variable with value 1 for CR samples and 0 for ad lib samples. Likewise, strain was defined as a dummy variable with value 1 assigned to samples associated with a given strain and value 0 assigned to samples associated with any of the other 6 strains. The strain-by-diet interaction effect was then evaluated based upon p-values associated with coefficients estimated for the strain-by-diet model term (Diet*Strain). Interaction effects were analyzed only for genes with detectable expression in 33% of all microarray samples for each tissue (eWAT: 16398 genes; Muscle: 16634; Heart: 15416; Cortex: 17793).

Meta-analysis of human scWAT responses to CR

We identified 28 experiments in which microarrays were used to analyze gene expression in periumbilical scWAT from human subjects before and after CR diet interventions (Table S2) [38–46]. In this context, an “experiment” is defined as two paired sets of microarray samples, including a baseline set obtained from scWAT of subjects prior to intervention and a post-intervention set obtained from subjects post-intervention. Following this definition, multiple experiments could be derived from the same study if repeated measures were obtained from subjects followed longitudinally over time (see below).

The 28 experiments include only dietary interventions with some protocols incorporating an additional exercise component, although we excluded any experiment in which a gastric bypass procedure was performed (Table S2). All dietary interventions were designed to induce weight loss with some experiments utilizing a baseline- and subject-specific calorie reduction protocol, and others utilizing a fixed protocol for all subjects with total calorie count ranging from 450 to 2000 kCal per day. Interventions ranged in duration from 4 weeks to 1 year with 10 of 28 experiments performed using males and 15 of 28 experiments performed using females (the sex was unknown for 3 experiments). Experiments varied in size from 3 to 40 subjects (average 10.9) with a total of 503 microarrays analyzed among all 28 experiments. For each experiment, we excluded from analyses the 15% of genes with lowest expression on average among all samples. For remaining genes and each patient, the difference in log2-transformed expression before and after treatment was calculated. For each gene, the average and standard error of these differences was calculated, and the experiment-specific significance of that difference was evaluated using linear models with moderated t-statistics [89].

A random effects meta-analysis model was applied to integrate results across the 28 experiments and to calculate a meta-signature for the effects of CR in human scWAT [92]. In some cases, 2 or more experiments had been obtained from repeated measures experiments and shared the same baseline set of samples. Results from such experiments cannot be regarded as independent. To obtain a filtered set of independent experiments, the 28 experiments were ranked according to the average standard error calculated among all genes with respect to the before versus after comparison. We filtered out any experiments sharing the same set of baseline samples, preferentially retaining those experiments with the lowest average standard error. This yielded a filtered set of 18 independent experiments with expression measurements available for 16504 genes. We excluded 3723 genes not measured in 6 or more of the 18 independent experiments, yielding a final set of 12781 genes for inclusion in the random effects meta-analysis model (R package: meta, R function: metagen). For this approach, the DerSimonian-Laird estimate was used to calculate a meta-FC for each gene with inverse variance weighting (i.e., greater weight assigned to experiments associated with a lower standard error estimate) [92]. To control for multiple hypothesis testing among the 12781 genes, p-values derived from the meta-analysis model were adjusted using the Benjamini-Hochberg method [90].

Microarray analysis of multiple WAT depots in ad lib and CR-fed obese C57BL/6 mice

Our main focus was to compare effects of CR in mouse eWAT and human scWAT, since these were the most well studied in terms of CR gene expression responses in each species, respectively [54]. However, to replicate this comparison using multiple WAT depots from a given strain, gene expression was evaluated in 4 WAT depots (eWAT, scWAT, perirenal WAT [prWAT]; mesenteric WAT [mesWAT]) from obese C57BL/6 male mice provided a high fat diet and subjected to CR for 0, 1, 3, 7, 14, 21, 42 and 60 days (GEO series accession GSE30534).

The initial dataset included 312 samples with genome-wide expression evaluated using the Affymetrix Mouse Genome 430 2.0 in situ oligonucleotide array platform. The 312 samples were normalized using robust multichip average yielding expression intensities for 45101 probe sets [85]. Samples from each WAT depot were clustered and visualized with respect to principal component axes to identify outliers, and were additionally evaluated using quality control metrics to identify low-quality hybridizations (e.g., NUSE/RLE median/IQR, average background, Affymetrix scale factors) (Figure S17) [87]. A total of 6 samples were identified as outliers and/or poor quality hybridizations and removed from analyses (GSM757584, GSM757585, GSM757599, GSM757392, GSM757502 and GSM757465), leaving a total of 306 samples used in further analyses.

The 45101 probe sets were collectively associated with 20736 human genes. If multiple probe sets were associated with the same gene, a single probe set was chosen to represent the gene by selecting the probe set with highest average expression among all samples [86]. We therefore considered a total of 20736 probe sets, with each probe set representing a unique human gene. Differential expression analyses were performed for 28 comparisons (4 WAT depots × 7 time points = 28 CR vs. ad lib comparisons). For a given comparison, only genes with detectable expression in at least 33% of samples were evaluated, yielding between 14711 and 15939 human genes among the 28 comparisons. A signed-rank test comparison between perfect match and mismatch probe intensities was used to determine if expression intensities for each probe set were detectable above background levels [88]. Differential expression analyses were carried out using linear models and moderated t-statistics [89], with raw p-values adjusted using the Benjamini-Hochberg method [90].