Supplementary Information Drugs That Reverse Disease Transcriptomic Signatures Are More Effective in a Mouse Model of Dyslipidemia

High-throughput omics have proven invaluable in studying human disease, and yet day-to-day clinical practice still relies on physiological, non-omic markers. The metabolic syndrome, for example, is diagnosed and monitored by blood and urine indices such as blood cholesterol levels. Nevertheless, the association between the molecular and the physiological manifestations of the disease, especially in response to treatment, has not been investigated in a systematic manner. To this end, we studied a mouse model of diet-induced dyslipidemia and atherosclerosis that was subject to various drug treatments relevant to the disease in question. Both physiological data and gene expression data (from the liver and white adipose) were analyzed and compared. We find that treatments that restore gene expression patterns to their norm are associated with the successful restoration of physiological markers to their baselines. This holds in a tissue-specific manner—treatments that reverse the transcriptomic signatures of the disease in a particular tissue are associated with positive physiological effects in that tissue. Further, treatments that introduce large non-restorative gene expression alterations are associated with unfavorable physiological outcomes. These results provide a sound basis to in silico methods that rely on omic metrics for drug repurposing and drug discovery by searching for compounds that reverse a disease's omic signatures. Moreover, they highlight the need to develop drugs that restore the global cellular state to its healthy norm rather than rectify particular disease phenotypes.


Reversal of disease phenotypes in the gene expression spaces
It is reasonable to assume that effective clinical interventions should reverse diseaseinduced pattern to the gene expression of affected organs. We therefore tested whether the treated animals tended to lie closer in the gene expression space to the healthy (LFD) group than did the untreated HFD group (see also Supplementary Figure 1). Formally, for each treatment group we computed the TDIs of animals in that group, and ran a two-sample t-test between them and the TDIs of the untreated HFD (16 weeks) group. The following table presents the one-sided p-values with which the null hypothesis of equal means for the TDIs of both groups can be rejected in favor of the alternative that the treatment led to a smaller mean TDI. P-values were adjusted to multiple hypotheses testing by the Benjamini-Hochberg (BH) method.  Figure 2a-b suggests that there exists considerable variability in the transcriptomic effects of some of the treatments. For example, animals treated with salicylate do not tend to cluster together in both adipose and the liver, whereas animals treated with rosiglitazone tend to cluster together in the gene expression space. In order to better study intra-group variability, hierarchical clustering of the animals in the gene expression space was conducted. Results are shown in Supplementary Figures 3-4. Dendrograms were created with the Euclidean metric, conforming to the rest of the study, and with average linkage. Unlike the PCA plots, which show only the first two principal components, and thus do not capture the entire variability in the data, the dendrograms were computed based on all the dimensions of the gene expression spaces. Nevertheless, the trends that they show are similar to those that can be observed in the PCA plots.

Experimental group Liver BH-adjusted p-value Adipose BH-adjusted p-value
First, in both tissues, the LFD animals cluster together, and most of the dietary lifestyle intervention (DLI) animals cluster with them. The dietary intervention is so successful at reversing disease gene expression patterns that these animals become practically indistinguishable from the animals that were fed LFD throughout the experiment. In the case of adipose tissue, the T0901317 cluster with LFD and DLI animals, yet still forms a distinct subgroup, in contrast to LFD and DLI which are "mixed together". Yet, this occurs only in adipose; in the liver the T0901317 group is distinct from the LFD+DLI cluster.
Second, there exists large intra-group variability due to which many of the animals do not cluster with other animals from the same treatment group. Two notable exceptions occur: in the adipose tissue, rosiglitazone and pioglitazone cluster together and distinctly from other drugs. The same happens with fenofibrate and T0901317 in the liver. This seems to occur because each of these four compounds activates a key regulator that is expressed in the tissue in which it exerts considerable transcriptomic changes, whereas the other drugs either work through other mechanisms that have a subtler effect or exert their primary effect in tissues that were not examined in the current study, such as pancreatic β-cells (see Supplementary Table 2).

GSEA analysis
Gene set enrichment analysis (GSEA, (Subramanian et al, 2005)) was conducted to detect pathways that are enriched in genes that are either upregulated or downregulated in each treatment group compared with the HFD-16weeks group. Comparisons of the LFD and HFD-9weeks groups with the HFD-16weeks groups were made as well for the sake of completeness. The analysis was limited to gene sets from the collection of canonical KEGG pathways in the Molecular Signatures Database (MSigDB) v4.0 (accession: CP:KEGG; 186 gene sets in total) so as to retain statistical power in the face of multiple comparison. On the other hand, we emphasize that input to GSEA consisted of all the genes whose expression was measured, and not only the subset of top differentially-expressed genes that was used to define the gene expression space for the purpose of TDI computations (Methods).
Gene sets were downloaded from MSigDB (www.broadinstitute.org/gsea/msigdb; accessed July 2014) and translated from human gene identifiers to mouse gene identifiers using homology data from the Jackson laboratory (www.informatics.jax.org; accessed July 2014). GSEA software available from the Broad Institute (v2.1.0; www.broadinstitute.org/gsea/index.jsp; accessed July 2014) was run with default parameters. We note that phenotype permutation was used to assess the statistical significance of the enrichment scores. Phenotype permutation is more stringent and biologically reasonable than gene set permutation (Subramanian et al, 2005), and was therefore preferred despite the limits it poses on statistical power in experiments with small number of samples in each group (our dataset typically has 8 animals per group).
Overall, the results agreed with prior expectations. Major hepatic and adipotic pathways were indeed modulated in the study animals by the drugs that are known to target them. Fenofibrate upregulated peroxisome proliferator-activated receptors (PPAR) signaling in the liver; fenofibrate, atorvastatin and T0901317 modulated hepatic fatty acid metabolism. Pioglitazone and rosiglitazone activated PPAR signaling and genes associated with fatty acid metabolism in white adipose. An exception to that was metformin, which did not alter any hepatic pathway in a statistically-significant way. This does not seem to stem from under-dosage because the dosage used was comparable to the one given in previous studies (250 mg/kg, 0.25% w/w), alleviated some of the clinical phenotypes of the disease (Radonjic et al, 2013), and significantly decreased the hepatic TDI compared with untreated HFD group (Supplementary Results 1). The indiscernibility of metformin's effects in GSEA analysis may thus stem from the lack of statistical power.
Interestingly, pro-inflammatory pathways were downregulated in adipose gene expression in the LFD and DLI groups compared with the HFD group, which accords with the importance of adipose-related inflammatory processes in HFD-induced pathologies (Wellen & Hotamisligil, 2003;Berg & Scherer, 2005). Nonetheless, inflammatory pathways were upregulated in the liver by T0901317, which is also apparent in direct inspection of the expression of known pro-inflammatory genes (Supplementary Results 5). The hepatic inflammatory response is probably associated with the deleterious physiological outcomes observed in T0901317 mice, most notably abnormal hepatomegaly (see main text in the results subsection "Non-restorative alterations to the gene expression are associated with unfavorable outcomes").
GSEA results also accord with observations that were reached in this study through other means: first, the dietary regimen seems particularly effective in inducing the opposite transcriptomic patterns than HFD. In both tissues there are multiple pathways which are altered between the LFD and the untreated HFD group due to HFD-feeding, and are altered in the opposite direction by DLI; the same reversal occurs in none or only in a handful of the drugs in each case, and only those that were shown to exert the most positive effect in the study animals. Second, taking the number of significantly altered gene sets as a proxy for the magnitude of the drug effect, we find that the drugs that had the most significant effects are the same ones that exert the largest effects as seen in the gene expression space (Supplementary Figure 12).
One result that we did not anticipate was the frequent occurrence of the KEGG_RIBOSOME gene set among the significantly altered ones in the liver. This gene set is upregulated in the liver by pioglitazone, rosiglitazone, fenofibrate, and T0901317 compared with the HFD group. Moreover, it is downregulated in the LFD and DLI group compared with the HFD group, suggesting that this pathway's downregulation is a phenotype associated with HFD and rectified by DLI. The gene set is also upregulated in the animals fed HFD for 9 weeks compared with the ones fed HFD for 16 weeks, which may be interpreted as a sign that this pathway's downregulation should be associated with a late phase in the disease progression and as a marker for a severe form of the disease state. Indeed, it has been recently shown that HFDfeeding repressed liver ribosomal RNA transcription in both wildtype (C57BL6) mice fed HFD and in an obese mouse model (ob/ob) fed normal diet compared with wildtype mice fed a normal diet (Oie et al, 2014).
Another noteworthy effect occurs in the liver, where T0901317 downregulated the gene set KEGG_COMPLEMENT_AND_COAGULATION_CASCADES compared with the HFD group. A similar result was previously reported in a zebrafish study of T0901317's hepatic effects (Sukardi et al, 2012), suggesting that it is not accidental but rather concerns a conserved biological mechanism in the two species. In our data, rosiglitazone (yet not pioglitazone) had the opposite effect and significantly upregulated this gene set in the liver. The gene set was significantly downregulated in the adipose tissue of the LFD group, yet not in the livers of the LFD animals.

GSEA-based methods
We define the Transcritome Deviation Index as the Euclidean distance in the gene expression space between an animal and the mean of the healthy (low-fat diet) animals. Note that the gene expression space is defined through genes that are differentially expressed between the HFD and LFD groups, and are thus associated with the disease phenotypes. A different approach was taken by several previous studies (Lamb et al, 2006;Iorio et al, 2010;Sirota et al, 2011;Pacini et al, 2013) that sought inverse correlations between drug and disease profiles derived from gene expression data, and applied Gene Set Enrichment Analysis (GSEA) (Subramanian et al, 2005) towards that purpose. Briefly, these studies computed a score that is based on the Kolmogorov-Smirnov statistic and quantifies the extent by which genes that are up-regulated in the disease profile tend to be up-regulated in the treatment profile and, similarly, genes that are down-regulated in the disease profile tend to be down-regulated in the treatment profile. Negative scores occur when genes that are up-regulated by the disease tend to be down-regulated by the treatment, and vice-versa, and suggest that the drug might be effective in treating the disease. Following (Sirota et al, 2011), these score are denoted DDS (which stands for drug-disease-score, although here they are applied in the case of the nonpharmacological dietary intervention, and in an individual manner, see below).
TDIs and DDSs are thus two ways to quantify the success of a treatment to reverse gene expression patterns induced by the disease. Treatments that successfully act towards this goal should have both small (close to 0) TDIs and small ("very negative") DDSs compared with unsuccessful treatments. Therefore, one expects TDIs and DDSs to be directly correlated; we verified that a strong correlation indeed exists.
DDSs were computed for each animal in the dataset studied here. There are minute differences between the ways the scores are computed in the various studies that used a GSEAbased approach; we followed (Sirota et al, 2011). As in the TDIs, a) DDSs were computed separately for the adipose and for the liver gene expression, and b) DDSs were computed for each individual animal, rather than for an entire treatment group. Thus, DDSs offer an alternative quantification for the tissue-specific reversal of the HFD gene expression patterns in a certain animal. As expected, TDIs and DDSs are highly correlated (Supplementary Figure 5; Pearson rho = 0.97, 0.96, p-values < 1.4e-38, 3.4e-62 for the adipose and liver tissues, respectively). On the other hand, the definition of TDIs allows a simple decomposition of the TDI into two orthogonal components: one that corresponds to disease reversal, and one which is associated with adverse outcomes (Supplementary Figure 13a). It is not as straightforward to do the same for GSEA-based scores, and therefore the definition of TDIs that is presented in the main text was chosen for the current study.

Genes mapped to multiple probes
A subtle choice in the definition of TDIs concerns the way probes that are mapped to multiple genes are handled. We opted for the most data-driven approach, and treated each probe as a separate feature, thus accommodating the possibility that a particular probe might be much more correlated with disease phenotypes than other probes mapped to the same gene. We verified, however, that our results do not depend on this choice. Similar results are obtained when all the probes that are associated with a particular gene are collapsed into a single feature. 5. Up-regulation of pro-inflammatory genes in the T0901317 treatment group T0901317 and fenofibrate are associated with unfavorable physiological outcomes that are indicative of liver pathologies, and particularly with notable hepatomegaly (see main text and Supplementary Figure 14), as well as with large non-restorative alterations in the liver gene expression (Supplementary Figures 12a, 13b). We hypothesized that these unfavorable phenotypes are accompanied by hepatic inflammation (Reddy & Sambasiva Rao, 2006). Therefore, we tested whether 13 known inflammatory genes were up-regulated in the livers of mice treated with one of these drugs compared with untreated HFD mice (one-sided t-test; pvalues were adjusted to multiple hypotheses by the Benjamini-Hochberg method; significance level was set at 5%). No significantly up-regulated genes were observed in the fenofibrate group. However, 6 out of the 13 tested genes were significantly up-regulated in the T0901317 (we stress that the comparison is with the untreated HFD group and not with the LFD group): MCP-1, CD86, EMR-1, ICAM-1, VCAM-1, and IL-1β. In addition, TNF-α was up-regulated, but not in a statistically-significant manner (adjusted p-value = 0.11, unadjusted p-value = 0.036). The other 6 genes that were tested are: SELE, SELP, NOS-1, NOS-2, IL-6, IL-18. Dendrogram was built with Euclidean distances and average linkage. Each leaf in the dendrogram corresponds to one animal, and leaf labels denote the treatment group to which the animal belonged. Two clusters are highlighted. The first (blue, upper part of the dendrogram) contains all the LFD animals, most of the dietary intervention animals, and all the animals treated with T0901317. The second (pink, bottom part of the dendrogram) contains almost all the animals treated with the thiazolidinediones (TZDs) rosiglitazone and pioglitazone. These two drugs activate a master transcription factor that is highly expressed in adipose. See Supplementary Results 2 and main text for details.

Supplementary Figure 5: Correlation between GSEA-based scores and TDIs
Scatter plots of DDSs, which are GSEA-based scores (see Supplementary Results 4.1) that measures a drug's ability to reverse the transcriptomic patters of the disease, and the Transcriptome Deviation Indices (TDIs), which quantify the same ability via other means. We find that the two are highly correlated both when computed for (a) liver and (b) white adipose gene expression. X and y axes are TDIs and DDSs, respectively. Each dot represents one animal, color-coded according to its treatment group as in the rest of the study. Pearson correlation coefficients and their corresponding p-values are given for each tissue. First two principal components of the liver metabolome space. Each dot represents one animal; color codes denote the different experimental groups. The dashed arrow connects the HFD centroid (yellow square) to the LFD centroid (yellow triangle), and denotes the direction of a reversal of the gene expression or physiological state back to the norm.

Supplementary Figure 8: White adipose TDI correlation with individual PDIs
Each panel presents the ranked PDI values of a particular physiological marker (y-axis) as a function of the ranked adipose TDI (x-axis

Supplementary Figure 10: Liver MDI correlation with individual PDIs
Deviations from the baseline liver metabolome (MDI) are correlated with deviations from the normal physiology (PDIs) in markers that are known to be associated with liver functions. Bar lengths represent the Spearman correlations between the hepatic MDI and PDIs of the measured 26 physiological markers. The liver has a central role in lipid metabolism, reflected in the relatively high correlations of its MDI and the physiological markers at the bottom part of the figure. WAT stands for white adipose tissue, ratio visc/sub WAT for ratio of visceral to subcutaneous WAT. Asterisks mark statistically-significant correlations (using the Benjamini-Hochberg correction for multiple hypotheses testing with a 5% FDR level). One marker (plasma MCP-1) had a negative correlation of -0.11 with the liver MDI, which is not shown in this figure.

Supplementary Figure 11: Liver MDI correlation with individual PDIs (scatter plots)
Each panel presents the ranked PDI values of a particular physiological marker (y-axis) as a function of the ranked liver MDI (x-axis). Each dot represents one animal, color-codes denote the different experimental groups. The dashed lines are linear regression lines. Refer to Supplementary Table 1 for complete details concerning the physiological markers. The Spearman correlations values and their respective p-values are given in Supplementary Table 8. 12. Supplementary Figure 12: Drugs that induce major non-restorative gene expression alterations This figure reproduces Figure 1a-b from the main text, highlighting the four "outlier" drugs, Fenofibrate (pink) and T0901317 (purple) in the liver, and the two thiazolidinediones rosiglitazone (dark cyan) and pioglitazone (light cyan). These drugs induce major gene expression changes that are not congruent with reversal of the disease transcriptomic patterns. The direction of reversal is denoted by the dashed arrow that connects the HFD centroid (yellow square, circled in red) to the LFD centroid (yellow triangle, circled in blue). Figure 13: Non-restorative gene expression alterations are associated with unfavorable physiological outcomes (a) A schematic illustration demonstrating the definition of non-restorative gene expression alterations. The gene expression space of a particular tissue is shown. Blue, red, and green markers represent LFD, untreated HFD, and treated HFD subjects, respectively. The dashed axis goes from the HFD mean to the LFD mean (yellow square and triangle, respectively). The treatment effects on each subject can be decomposed into two components: (1) reversal of the disease-induced gene expression patterns, which operates along the direction of the axis that goes from the HFD mean to the LFD mean, and (2) additional alterations which are orthogonal to that axis and hence incongruent with the healthy (LFD) state (Methods). We term the latter "non-restorative alterations" and hypothesize that they are associated with unfavorable physiological outcomes. (b-c) The distributions of the magnitudes of non-restorative alterations to the (b) liver and (c) white adipose gene expression induced by the various drugs. In (b) experimental groups are ordered from left to right in the same order that they appear in the legend; in (c) the groups are ordered from left to right as follows: DLI, pioglitazone, rosiglitazone, salicylate, T0901317. Evidently, four drugs induce the largest non-restorative alterations: fenofibrate and T0901317 in the liver and rosiglitazone and pioglitazone in white adipose (compare Figure 1a-b and Supplementary Figure 12). (d) a schematic illustration of the method employed to detect unfavorable outcomes in the physiological data available in the studied animals. Intuitively, a marker was considered as manifesting an unfavorable outcome of a certain treatment if its values in the treated animals were even farther from the LFD baseline than its values in the untreated HFD animals. This is exemplified in the illustration: while the untreated HFD animals (red bar) have higher blood triglycerides levels than the LFD animals (blue bar), the animals treated with a certain bar (green bar) have even higher blood triglycerides than the untreated HFD animals. Hence, the marker represents an unfavorable outcome of the treatment in this case. See the main text for a complete definition that also quantifies the statistical significance of the observation.

Supplementary
14. Supplementary Figure 14: Unfavorable drug outcomes in physiological marker data This figure highlights the statistically-significant unfavorable physiological outcomes that were ascribed to particular drugs. Each panel presents the distribution of one physiological marker, with each bar representing one experimental group, color-coded as in the rest of the paper. Refer to Supplementary Table 1 for details concerning the measured markers, their units etc. A marker was considered as an unfavorable outcome of a certain treatment if its values in the treated animals were even farther from the baseline than its values in the untreated HFD animals in a statistically-significant manner (see main text for details). The figure shows all the statistically-significant associations of an unfavorable physiological outcome and a drug found in the data (i.e., the panels correspond to all the markers, interpreted as unfavorable outcomes, for which a statistically-significant association with at least one drug is detected; boxes shown in the panel correspond to the all the drugs which were associated with this unfavorable outcome). Note that an exception was made in the case of fenofibrate, which was not associated with elevated plasma triglycerides and atherosclerotic lesion area as unfavorable outcomes in a statistically significant manner; it is shown in those panels only for completeness of the presentation. One LFD outlier in the right panel of the middle row had a value of 1.6, but was clipped to a value of 3 (dashed line) for the sake of visualization.

Supplementary Tables
1. Supplementary Table 1: List of physiological markers measured in the study animals WAT = White Adipose Tissue. The termination column indicates the time point at which these markers were measured: 9 weeks for the HFD-9wks group, and 15 or 16 weeks for the other groups.

Physiological marker Units
Fasted / non-fasted Termination creatinine ratio) (*) The marker was log-transformed because it was highly skewed, and followed an approximately normal distribution much more closely after taking the log. Note that (Radonjic et al, 2013) did not carry a similar transformation.
(***) computed as 1/[log( ) + log( )], fasting insulin in uU/ml and fasting glucose in mg/dl (Katz et al, 2000). Table 2: Drug mechanism of action The mechanism of action of drugs studied in this paper. Data is based on Drugbank (Law et al, 2013) (accessed July 2014) except where otherwise noted.

Drug
Mechanism of action metformin Metformin's mechanism of action is two-fold, inhibiting liver glucose production, and additionally augmenting peripheral glucose uptake, mainly in muscles. These effects are believed to be partly mediated by activation of liver kinase B1 (LKB-1) (Shaw et al, 2005), which in turn regulates 5' adenosine monophosphatase-activated protein kinase (AMPK), a key sensor of cellular metabolism and energetics. Nonetheless, metformin has been reported to improve glucose tolerance in liver AMPKdeficient mice (Foretz et al, 2010), which suggests that part of its effects occurs through AMPK-independent pathways (Rena et al, 2013).
glibenclamide Glibenclamide is a second generation sulfonylurea, which stimulate insulin secretion by pancreatic β cells. Sulfonylureas bind to a sulfonylurea receptor that is associated with inward rectifying adenosine triphosphate (ATP)-sensitive potassium channels in β cells. Binding of a sulfonylurea inhibits the efflux of potassium ions through the channels and results in depolarization that opens voltage-gated calcium channels. This leads to calcium influx and to the release of preformed insulin (Nolte Kennedy, 2012).
rosiglitazone Pioglitazone and rosiglitazone are thiazolidinediones (TZDs), which exert their antidiabetic effects through activation of the gamma isoform of the peroxisome proliferator-activated receptor (PPARγ), a transcription factor that is highly expressed in adipose tissue, and is known to be a key regulator of adipogenesis and insulin sensitivity (Escher et al, 2001;Larsen et al, 2003;Evans et al, 2004;Vasudevan & Balasubramanyam, 2004;Poulsen et al, 2012;Ahmadian et al, 2013). Activation of PPARγ receptors regulates the transcription of insulin-responsive genes involved in the control of glucose production, transport and utilization. Thus, TZDs improve glycemic control in type 2 diabetic patients through insulin sensitization, rather than increased insulin secretion by pancreatic β cells (Soccio et al, 2014).
pioglitazone fenofibrate The chief mode of action of fenofibrate is binding to PPARα, a transcription factor that is highly expressed in the liver (as well is in brown, but not white, adipose cells) (Escher et al, 2001;Evans et al, 2004;Oosterveer et al, 2009;Poulsen et al, 2012). Upon its activation PPARα heterodimerizes with retinoid X receptor (RXR); the heterodimers recognize specific PPARα response elements and modulate the expression of genes responsible for fatty acids and cholesterol metabolism (Staels et al, 1998). The decrease in plasma triglycerides induced by fibrates has been attributed to an inhibition of the synthesis and secretion of VLDL by the liver and increased degradation of triglyceride-rich lipoproteins through the expression of lipoprotein lipase and a decreased expression of apolipoprotein CIII (Forcheron et al, 2002).

T0901317
T0901317 is a synthetic Liver X Receptor (LXR) agonist. LXR has two isoforms, one of them (LXRβ) is ubiquitously expressed, whereas the other (LXRα) is restricted to particular tissues, including the liver. LXRs regulates lipid and cholesterol metabolism and also have antiinflammatory properties (Schultz et al, 2000;Steffensen & Gustafsson, 2004;Ulven et al, 2005;Zhao & Dahlman-Wright, 2010). T0901317 has been found unsuitable for clinical use due to its pleotropic effects, but LXRs continue to be studied as an attractive drug targets (Jakobsson et al, 2012;Hong & Tontonoz, 2014).
atorvastatin Atorvastatin lowers LDL cholesterol by inhibiting hydroxymethylglutarylcoenzyme A (HMG-CoA) reductase, which catalyzes the conversion of HMG-CoA to mevalonate in the cholesterol biosynthesis pathway. Ample evidence support that statins' protective cardiovascular effects is not restricted to cholesterol metabolism but may also be related to their antiinflammatory properties (Tousoulis et al, 2014).
salicylate Salicylates are anti-inflammatory compounds that inhibit the activity of both types of cyclooxygenase (COX-1 and COX-2) and thus suppress platelet thromboxane synthesis. The artificial derivative acetylsalicylic acid, better known as aspirin, is broadly used to prevent atherosclerotic complications, most importantly myocardial infarction and ischemic stroke (Awtry & Loscalzo, 2000;Campbell et al, 2007;American Diabetes Association, 2013). Aspirin effectively inhibits platelet aggregation, yet this effect is partly mediated through its acetyl group (Furst et al, 2012;Steinberg et al, 2013). There may also be other mechanisms through which salicylates exert their favorable effects, such as inhibition of the pro-inflammatory κ-light-chain-enhancer of activated B cells (NF-κB) signaling pathway ( (Kopp & Ghosh, 1994) but see also (Frantz et al, 1995;Steinberg et al, 2013) ), and activation of AMPK (Hawley et al, 2012;Steinberg et al, 2013) that is also a target of metformin (see above).
rofecoxib Rofecoxib is a selective cyclooxygenase-2 (COX-2) inhibitor, which has been withdrawn in 2004 worldwide by Merck & Co, due to an increased risk of cardiovascular events (Praticò & Dogné, 2005). Table 3: GSEA analysis of the liver transcriptome Gene set enrichment analysis (GSEA, (Subramanian et al, 2005)) of the effects of the pharmacological and dietary interventions was conducted on the hepatic transcriptome. The analysis sought KEGG gene sets that were enriched with either upregulated or downregulated genes when comparing the treated animals with the untreated HFD-16weeks group.