Genetic Variation Determines PPARγ Function and Antidiabetic Drug Response In Vivo

SUMMARY

SNPs affecting disease risk often reside in non-coding genomic regions. Here we show that SNPs are highly enriched at mouse strain-selective adipose tissue binding sites for PPARγ, a nuclear receptor for antidiabetic drugs. Many such SNPs alter binding motifs for PPARγ or cooperating factors, and functionally regulate nearby genes whose expression is strain-selective and imbalanced in heterozygous F1 mice. Moreover, genetically-determined binding of PPARγ accounts for mouse strain-specific transcriptional effects of TZD drugs, providing proof-of-concept for personalized medicine related to nuclear receptor genomic occupancy. In human fat, motif-altering SNPs cause differential PPARγ binding, provide a molecular mechanism for some expression quantitative trait loci, and are risk factors for dysmetabolic traits in genome-wide association studies. One PPARγ motif-altering SNP is associated with HDL levels and other metabolic syndrome parameters. Thus, natural genetic variation in PPARγ genomic occupancy determines individual disease risk and drug response.

INTRODUCTION

A major unanswered question is how most genetic variation causes phenotypic differences, as only a small fraction of single nucleotide polymorphisms (SNPs) affect protein sequence (Shastry, 2002). Current genome-wide association studies (GWAS) reveal a large gap between known causal genes and the observed heritability of common diseases and treatment outcomes (Sadee et al., 2014). Another limitation of GWAS is that each locus nominates a large group of SNPs in linkage disequilibrium, such that causal and neutral variants cannot easily be distinguished. Non-coding SNPs in regulatory regions may affect transcription factor (TF) binding and gene expression, thus contributing to complex phenotypes like disease association and response to drugs (Edwards et al., 2013).

There are examples of regulatory variants causing Mendelian syndromes (De Gobbi et al., 2006; Smemo et al., 2012), but such SNPs may be more likely to associate with complex non-Mendelian diseases in GWAS (Sakabe et al., 2012). Overall, putative causal GWAS SNPs cluster more in promoters and enhancers than in exons (Andersson et al., 2014), and a recent effort to computationally identify causal GWAS SNPs for autoimmune diseases found that ~90% were non-coding, with ~60% in distal immune cell enhancers (Farh et al., 2015). A few specific examples have emerged. The causal SNP for an LDL cholesterol and myocardial infarction locus is a regulatory variant altering hepatic SORT1 expression (Musunuru et al., 2010). Regulatory SNPs in distant enhancers for MYC result in associations with multiple cancers (Sur et al., 2013), and an intronic enhancer SNP in TCF7L2 may mediate type 2 diabetes (T2D) risk (Gaulton et al., 2010). For the PPARG T2D locus, the causal SNP was thought to be a coding Pro12Ala polymorphism, yet recent evidence has implicated a tightly linked regulatory SNP (Claussnitzer et al., 2014).

PPARγ provides an excellent system to study effects of regulatory variation on TF binding, gene expression, drug response, and phenotype. PPARγ is a nuclear receptor TF required for adipocyte development (Wang et al., 2013) that activates many adipocyte genes. PPARγ is genetically implicated in metabolic disease, both through the common SNP associated with T2D (Altshuler et al., 2000), and also through rare ligand binding domain mutations causing an autosomal dominant syndrome of lipodystrophic insulin resistance (Barroso et al., 1999). Since variants affecting the PPARγ TF itself have these consequences, then genetic variation in key PPARγ genomic binding sites may similarly have metabolic effects.

PPARγ is also the target of antidiabetic thiazolidinedione (TZD) drugs, which have a unique and powerful insulin-sensitizing effect, yet clinical use has declined due to concerns over side effects and adverse events (Soccio et al., 2014). Individuals differ in drug response, and ~20–30% of diabetic patients fail to respond to TZDs (Sears et al., 2009). Most pharmacogenomic studies focus on coding or non-coding variants affecting the drug target itself, or drug-metabolizing enzymes and transporters (Mizzi et al., 2014). However, regulatory variants may potentially alter downstream transcriptional effects of drugs, either indirectly after signal transduction from a cell surface receptor, or directly in the case of DNA-binding nuclear receptors like PPARγ.

Here, we set out to determine whether non-coding regulatory variation could affect PPARγ genomic occupancy, and whether such SNP-dependent binding could affect gene expression, drug response, and metabolic phenotype. Chromatin immunoprecipitation followed by deep sequencing (ChIP-seq) was used to define genetically-determined variation in PPARγ sites genome-wide in white adipose tissue (WAT). In mice, sites with inbred strain-selective PPARγ genomic occupancy were highly enriched for SNPs, and in heterozygous F1 mice these SNPs had allelic imbalance in PPARγ binding. These SNPs often altered TF motifs, not only motifs for PPARγ but also motifs for other, cooperating TFs. Importantly, some strain-selective binding sites were functional regulating nearby gene expression in WAT, both basal and TZD-stimulated. Similar studies were performed in human WAT, where SNPs also led to imbalanced PPARγ binding. Remarkably, these human SNPs were enriched in WAT expression quantitative trait loci (eQTLs) as well as in regions linked to metabolic disease in GWAS. Thus, variable PPARγ occupancy due to SNPs determines nearby gene activation by PPARγ and its ligands, and these effects may underlie genetic differences in metabolic phenotypes and drug responses.

RESULTS

Genomic binding of PPARγ in mouse fat is strain-selective and driven by SNPs

ChIP-seq was performed in WAT from two inbred mouse strains, C57Bl/6J (B6) and 129S1/SvImJ (129), which differ in susceptibility to obesity and insulin resistance (Almind and Kahn, 2004). Since highly polymorphic 129 sequencing reads may not align to the B6 reference genome, the SNP-sensitive alignment tool GSNAP (Wu and Nacu, 2010) was used to eliminate alignment bias and identify truly strain-selective binding (Figures S1A–B). Three independent ChiP-seq experiments were performed (Table S1), allowing identification of strain-selective sites at high confidence (3-fold strain difference in reads at least 2 of 3 experiments) and highest confidence (all 3, Figure 1A). Average peak heights in F1 intercross progeny were intermediate between parents, while nonselective sites were equal in all three (Figure 1B), indicating that strain-selectivity of adipose genomic PPARγ occupancy was genetically determined. Of note, while WAT PPARγ cistromes include contributions of other cell types that express PPARγ, such as resident macrophages, the great majority of sites overlapped with those in 3T3-L1 adipocytes but not with previously reported macrophage-selective sites (Lefterova et al., 2010), thus likely represent adipocyte binding sites.

Figure 1. SNPs genetically determine mouse strain-selective PPARγ sites.

(A) ChIP-seq in mouse WAT identified ~35,000 PPARγ sites, and the heat map shows 2,226 B6 or 129 strain-selective sites in 3 independent experiments. High confidence sites were 3-fold strain-selective in 2, while highest confidence were in all 3. (B) For the five PPARγ site classes, average binding profiles are shown for the two inbred strains (B6 red, 129 blue) and F1 progeny (green). (C) For the five classes, occurrence of one or more B6:129 SNPs in each site’s central 200bp is shown, with enrichment of SNPs in strain-selective sites (*P<0.0001 versus non-selective sites by Chi-squared test). (D) In F1 ChIP-seq, allelic imbalance was assayed in binding sites with SNPs (*P<0.0001 versus non-selective sites by Mann-Whitney test, 2-tailed). See also Figure S1, Table S1.

The B6 and 129 genomes differ by ~5.3 million SNPs (Keane et al., 2011), and strain-selective sites were highly enriched for occurrence of SNPs (Figure 1C). Notably, SNPs falling in B6- or 129-selective sites showed PPARγ binding allelic imbalance in F1 WAT favoring the allele with better parental binding, whereas SNPs in non-selective sites showed equal representation of both alleles (Figure 1D). F1 imbalance shows that cis-acting elements determine PPARγ occupancy, as selectivity is evident even when two alleles are in the same nucleus.

To further validate cis-acting PPARγ site SNPs, ChIP-seq was performed in 3T3-L1 adipocytes, a cell line derived from outbred NIH Swiss albino mice (Todaro and Green, 1963) and thus heterozygous at many loci. In these cells PPARγ bound at ~9,000 sites of B6:129 SNPs, and, importantly, at least 18% were heterozygous (Figure S1C). At heterozygous sites where PPARγ binding was strain-selective in B6 versus 129 WAT, the predicted allelic imbalance in binding was observed in 3T3-L1 adipocytes (Figure S1D). This is similar to F1 mice, and confirms the powerful effect of cis-acting SNPs on PPARγ genomic binding.

SNPs alter TF motifs to cause strain-selective PPARγ binding

PPARγ binds DNA at direct repeat 1 (DR1) motifs both in vitro (Tontonoz et al., 1994) and in 3T3-L1 adipocytes (Lefterova et al., 2008; Nielsen et al., 2008), and this was the top motif found in WAT PPARγ sites (Figure 2A). To test whether DR1-altering SNPs cause strain-selective binding, all polymorphic DR1 motifs in PPARγ binding regions were identified and assigned motif scores, such that the B6:129 score ratio indicated strain difference in consensus motif agreement. SNPs with large effects on PPARγ/DR1 motifs (ratio >16) were highly likely to show selective binding in the strain with the stronger motif, and unlikely to be selective for the opposite strain (Figure 2B). This discrimination was apparent even when SNPs had medium (8–16 fold) or small (2–8 fold) motif effects, but effectively lost when SNPs had minimal effects (<2-fold). While this analysis used thresholds for motif effects and binding difference, the same pattern emerged in a quantitative scatterplot analysis correlating binding ratio versus motif ratio (Figure S2A). This approach relies on natural B6:129 genetic variation at PPARγ motifs, but can be extended to other variants. For instance, 3T3-L1 PPARγ sites have 936 heterozygous SNPs that are non-polymorphic between B6 and 129, and when these SNPs altered PPARγ/DR1 motifs, there was the predicted allelic imbalance in PPARγ occupancy (Figure S2B).

Figure 2. PPARγ binding is altered by SNPs affecting motifs for PPARγ or other TFs.

(A) Consensus motif logos for the four top scoring motif families found in PPARγ binding regions, with Cistrome identifiers and Z-scores. (B) Strain-selectivity of PPARγ binding sites at motif-altering SNPs divided into 4 classes based on size of the allelic effect on consensus motif agreement. (C–E) The same analysis for C/EBP, NFI, or GR motif-altering SNPs. The difference between SNPs with large and minimal motif effects was P<0.0001 in B and C, P=0.0026 in D, P=0.039 in E (Chi-squared test). See also Figure S2–S3, Table S1.

Individual nucleotide substitutions at each position of the DR1 motif (Figure S2C) were interrogated for PPARγ occupancy effects. Overall, this strongly validated the motif ratio approach, though there were informative exceptions where SNPs at some locations had large occupancy effects despite apparently small motif effects (Figure S2D–F). This reinforces a key point about consensus ChIP-seq motifs: they reflect nucleotide frequencies at motifs actually bound by the TF, but may not necessarily represent the strongest binding version of the motif. Also, in addition to the core DR1 motif, three upstream bases also determined PPARγ occupancy (Figure S2G), confirming in vivo the importance of 5’ flanking sequence as reported in prior in vitro studies (Juge-Aubry et al., 1997) and the X-ray crystal structure (Chandra et al., 2008). Finally, the most drastic effects on DR1 motifs and PPARγ binding occurred when multiple SNPs altered the same motif (Figure S2H).

PPARγ often binds DNA in close proximity to C/EBP TFs, and the two facilitate each other’s binding by assisted loading (Madsen et al., 2014). Consistent with other reported PPARγ cistromes (Lefterova et al., 2008; Nielsen et al., 2008), at WAT PPARγ sites the top non-DR1 motif was for C/EBP (Figure 2A). As with DR1, SNPs with large C/EBP motif effects caused the predicted strain-selectivity in PPARγ occupancy (Figure 2C). In addition to PPARγ and C/EBP, a motif for the nuclear factor I (NFI) family was also enriched at PPARγ binding sites, consistent with previous reports (Rajakumari et al., 2013). SNPs with large NFI motif effects also gave strain-selectivity in PPARγ binding (Figure 2D), indicating that an NFI TF can modulate PPARγ genomic binding in vivo. Thus, in addition to the well-known PPARγ-cooperating factor C/EBP, this SNP-based method can also suggest functional relevance for novel candidate TFs.

The next highest motif found in PPARγ sites was a glucocorticoid receptor (GR) motif, and SNPs altering this inverted repeat 3 (IR3) motif affected PPARγ occupancy (Figure 2E). The effect of IR3 motifs was independent, as an IR3 motif SNP never also affected an overlapping DR1. GR plays a major role in adipocyte biology (Steger et al., 2010). GR ChIP-seq performed in WAT from B6 and 129 mice revealed strain-selective GR binding (Figure S3A) at sites highly enriched for SNPs (Figure S3B), and SNPs in GR motifs had predicted effects on GR occupancy (Figure S3C). Moreover, the majority of PPARγ binding sites in WAT are also occupied by GR, and hundreds of sites had high confidence strain-selective binding of both factors (Figure S3D). Many had motif-altering SNPs in PPARγ, GR, or C/EBP motifs, and all three types of SNPs could mediate strain-selective binding of both PPARγ and GR (Figure S3E–F). Therefore, SNPs altering PPARγ motifs not only affect PPARγ occupancy, but also binding of other TFs like GR. Conversely, PPARγ binding can be altered by SNPs in PPARγ motifs as well as motifs for other TFs, showing the powerful effect of motif SNPs on cooperative binding of multiple TFs.

Selective sites with SNPs but not identifiable motifs (Figure S3G) may be due to SNPs affecting degenerate or non-consensus motifs for these TFs or other TF motifs, yet many strain-selective sites are non-polymorphic over 200bp. Such unexplained strain-selective sites, with or without SNPs, may result from long-range interactions with other sites in the same locus, and this is consistent with the observed clustering of both B6 and 129-selective binding sites near other sites selective for the same strain (Figure S3H).

Strain-selective PPARγ binding functions to modulate nearby gene expression

The great majority of PPARγ sites, including the genetically determined sites defined here, reside outside promoters >5kb from transcription start sites (TSSs). The function of sites as enhancers correlates with occupancy of cofactors such as Med1 and p300, as well as with transcription of enhancer RNA (eRNA) identified by global run-on sequencing (GRO-seq) (Hah et al., 2013; Step et al., 2014). GRO-seq was performed in WAT from B6 mice, and eRNA was quantified at PPARγ sites. High eRNA transcription in B6 WAT was present at ~18% of B6-selective sites, a significant three-fold enrichment versus the 129-selective sites with little PPARγ binding in B6 mice, the majority of which had no detectable eRNA transcription (Figure 3A). Thus, strain-selective PPARγ binding in WAT correlated with functional enhancer activity defined by eRNA transcription. Moreover, ChIP-seq for coactivators in 3T3-L1 adipocytes (Step et al., 2014) revealed imbalanced occupancy of p300 (Figure 3B) and Med1 (Figure S4A) at heterozygous sites with the same imbalance in PPARγ binding. Thus, allele-selective PPARγ occupancy robustly predicts cofactor occupancy and differential enhancer function.

Figure 3. Mouse strain-selective PPARγ sites drive differential WAT gene expression.

(A) eRNA transcription in B6 WAT was measured by GRO-seq. By Chi-squared test, P<0.0001 for eRNA distribution, enriched in B6-selective sites. (B) 373 SNPs in mouse 3T3-L1 adipocytes had adequate read depth (10+) in both PPARγ and p300 ChIP-seq, and allelic imbalance in p300 occupancy correlated with imbalance in PPARγ occupancy (linear regression r²=0.4 and P<0.0001 for a non-zero slope). Inset shows that, at 2-fold threshold, ~60% of SNPs with imbalanced PPARγ binding had p300 imbalance in the same direction, while only ~3% had p300 imbalance in the opposite direction. (C) Four RNA-seq experiments identified ~13,000 genes expressed in WAT, and 432 genes showed significant strain-differential expression >1.5-fold in at least 3 experiments. (D) For the three gene classes, the nearest strain-selective binding site within 50kb of the TSS was identified, demonstrating that strain-differential genes enrich for sites selective for the same strain. By Chi-squared test, P<0.0001 for strain-similar versus B6- or 129-higher, and for B6- versus 129-higher. (E) Representative RNA-seq and ChIP-seq browser tracks are shown at the Sgcg gene, with 129-higher expression and a 129-selective PPARγ site (yellow). (F) The consensus PPARγ/DR1 motif logo is shown, above the motif sequences in both strains at the Sgcg binding site, with SNP alleles colored and showing their difference in motif agreement. See also Figure S4–S5, Table S2.

To test for function of strain-selective PPARγ sites in gene regulation, RNA-seq was performed in four separate experiments using WAT from B6 and 129 mice (Table S2), revealing 432 genes with consistently higher expression in one strain (Figure 3C). RNA-seq was also performed in WAT from F1 mice, and genes with strain-differential expression showed the predicted imbalanced expression in F1s, indicating cis effects (Figure S4B). Strain differences in gene expression observed in WAT were also found in primary adipocytes differentiated in cell culture (Figure S4C), further supporting genetic determination.

To integrate strain-differential gene expression with PPARγ binding, we identified the nearest strain-selective site within 50kb in either direction from each TSS. For genes with B6-higher expression, the nearest selective PPARγ site was B6-selective 4 times as often as 129-selective (Figure 3D). Conversely, a 5-fold enrichment for 129-selectivity of the nearest PPARγ site was observed for genes whose expression was higher in 129 WAT. By contrast, strain-similar genes were equally unlikely to be near to B6- or 129-selective sites. There was also a marked effect of distance, such that if a strain-differential gene TSS was within 5kb of the strain-selective site, gene expression and PPARγ binding nearly always favored that same strain (Figure S4D). If the nearest strain-selective site was 5 to 50kb away, there was ~2-fold same-strain preference for gene expression, but this was lost for distances 50 to 100kb. The consistent direction of effects and their proximity dependence strongly suggest that PPARγ binding is causing differential gene expression. For example, Sgcg expression is much higher in 129 WAT, and in its first intron ~600bp from the TSS lies a 129-selective PPARγ site (Figure 3E) with a 129-stronger DR1 motif SNP (Figure 3F). In F1 WAT, expression is intermediate, with PPARγ binding and gene expression both demonstrating allelic imbalance highly favoring 129 alleles. Although no characteristic of strain-selective PPARγ binding was fully predictive of differential gene expression, sites with strain-selectivity in both binding and nearest gene expression were on average stronger (Figure S4E), nearer to the gene’s TSS (Figure S4F), associated with fewer nearby and potentially redundant sites (Figure S4G), and enriched for a helix-loop-helix TF-binding motif (Figure S4H) compared to control sites (non-selective or strain-selective without differential nearest gene expression).

Functional PPARγ site SNPs have potentially powerful applications to genetic studies. B6 and 129 mice differ in their susceptibility to obesity, and several quantitative trait loci (QTLs) for this phenotype have been mapped (Lin et al., 2013). One effort used computational tools to narrow QTL intervals (Su et al., 2008), and within these we find 5 cases of strain-selective WAT gene expression attributable to PPARγ binding (Figure S5A). For instance, the Zdhhc2 gene encodes a palmitoyltransferase and a 129-selective PPARγ site is in nearby gene intron, with a DR1 motif-altering SNP favoring 129 (Figure S5B–C). Zdhhc2 expression is 129-higher, and there is also the predicted F1 imbalance in expression and PPARγ binding. The PPARγ site was validated by ChIP and quantitative PCR showing high binding in 129, absent binding in B6, and intermediate binding in F1 WAT (Figure S5D). Furthermore, an allelic imbalance assay on the F1 ChIP PCR product confirmed the imbalance observed by ChIP-seq (Figure S5E). Thus, genetically-determined adipose PPARγ binding that affects gene expression nominates mechanistic candidates for several mouse obesity QTLs.

Strain-selective PPARγ binding mediates differential responses to antidiabetic drugs

Rosiglitazone (rosi) is an antidiabetic TZD drug that functions as a high-affinity activating ligand for PPARγ (Lehmann et al., 1995). To determine whether genetic differences in PPARγ binding affect TZD response, B6 and 129 mice were treated with rosi for 2 weeks, then RNA-seq of WAT was performed (Table S2). Rosi up-regulated (Figure 4A) and down-regulated (Figure S6A) genes were classified by statistical significance in both strains versus B6-only or 129-only. Integration of these genes with nearby strain-selective PPARγ binding revealed that genes rosi-up only in B6 mice were much more likely be near a B6-selective sites, and genes that were rosi-up only in 129 mice likewise showed more nearby 129-selective sites (Figure 4B). By contrast, there was no such discrimination in four control gene classes: those rosi-up in both strains were equally likely to be near B6- and 129-selective sites (Figure 4B), as were all groups of rosi-down genes (Figure S6B). This latter result is remarkably consistent with the recent finding that adipocyte gene repression by rosi is unrelated to nearby PPARγ binding (Step et al., 2014). As an example, the Abhd3 gene, encoding a phospholipid lipase (Long et al., 2011), was rosi-up by 2-fold only in 129 mice (Figure 4C), consistent with an upstream 129-selective PPARγ site (Figure 4D) with a 129-stronger DR1 motif-altering SNP (Figure 4E). Other examples of strain-selective rosi-induced genes are shown in Figure S6C. Allele-dependent response to rosi was also measured in 3T3-L1 adipocytes by analyzing GRO-seq data (Step et al., 2014) for allelic imbalance in gene body transcription. The rosiglitazone-induced gene Dhrs9 has two haplotypes (reference and other) in 3T3-L1 cells, with 50 heterozygous SNPs in the gene (Figure 4F). In the absence of rosi, these SNPs showed imbalanced transcription favoring reference alleles (Figure 4G). Rosi elicited a selective increase in transcription of reference but not other alleles, and thus significantly increased the imbalance (Figure 4G–H). This haplotype-selective effect of rosi on Dhrs9 transcription correlated with PPARγ occupancy, as two sites near the TSS harbor SNPs with imbalanced PPARγ binding strongly favoring reference alleles (Figure 4F). These data demonstrate the pharmacogenomic role of regulatory SNPs in determining transcriptional response to a drug, in this case by altering PPARγ genomic binding.

Figure 4. Mouse strain-selective PPARγ sites drive differential gene activation by rosiglitazone.

(A) RNA-seq identified rosi-regulated genes in WAT, and the heat map shows those with statistically significant increases in both strains (both EDGE M P<0.001) versus only one (P<0.001 but P>0.01 in the other strain). (B) For the three classes of genes, the nearest strain-selective binding site within 50kb of the TSS was identified. By Chi-squared test, P<0.01 for B6-only versus 129-only, and P<0.05 for the overall distribution. (C) The gene Abhd3 was rosi-up only in 129 WAT. A 129-selective PPARγ site (yellow) is upstream of the Abhd3 TSS (D), with a motif-altering SNP favoring the 129 allele (E). (F) In 3T3-L1 adipocytes, PPARγ ChIP-seq showed two sites near the Dhrs9 TSS (*), each with a heterozygous central SNP imbalanced favoring the reference allele by 3–4 fold. GRO-seq showed a significant 40% rosi-induction of positive strand transcription of Dhrs9, with 50 heterozygous SNPs in the gene body. (G) Counting reference and other alleles at these SNPs showed baseline imbalance favoring reference alleles, and that rosiglitazone selectively activated only the reference haplotype. (H) At these SNPs, rosi significantly increased the average amount of imbalance. In C and H, mean and SEM, with ***P<0.0001, **P<0.01, *P<0.05, NS=not significant, by unpaired T-test, 2-tailed). See also Figure S6, Table S2.

Motif-altering SNPs determine PPARγ occupancy and gene expression in human fat

Humans are outbred and any two unrelated individuals differ at SNPs in ~0.1% of the genome (Shastry, 2002), similar to the difference between B6 and 129 mice. The above mouse findings are not directly applicable to humans, since the great majority of adipocyte PPARγ sites are not retained at syntenic genomic positions between the species (Mikkelsen et al., 2010; Soccio et al., 2011), and entirely different SNPs exist as they arose after speciation. Therefore, to determine the effects of SNPs on PPARγ genomic occupancy, ChIP-seq was performed on human subcutaneous adipose tissue from five individuals (Table S3). SNPs were identified based on three criteria: (1) annotated in dbSNP141 with minor allele frequency >1%, (2) located in a PPARγ binding site identified by ChIP-seq in one or more subjects, and (3) altered an identifiable PPARγ or C/EBP motif. To test for effects of motif alterations, a subset of SNPs was identified as heterozygous in at least one subject as defined by detection of both alleles in ChIP-seq reads; imbalanced PPARγ binding in heterozygous subjects was investigated by the same method used in F1 mice and 3T3-L1 cells. Remarkably, and similar to mouse, human SNPs altering motifs for PPARγ (Figure 5A) or C/EBP (Figure 5B) led to imbalanced binding in heterozygotes favoring the alleles with stronger motifs. Thus, motif-altering SNPs determine PPARγ binding in human fat. Indeed, this effect was sizable even though heterozygous cases with 100% imbalance, which appear homozygous in the ChiP-seq data, could not be identified by this analysis.

Figure 5. Human motif-altering SNPs affect PPARγ occupancy and nearby gene expression.

Human PPARγ sites were found by ChIP-seq in WAT from 5 subjects, and SNPs in these sites were identified that altered PPARγ/DR1 or C/EBP motifs. (A) PPARγ-motif altering SNPs heterozygous in one or more subjects were assessed for 2-fold allelic imbalance in ChIP-seq reads, demonstrating that SNPs with larger motif effects showed enrichment for higher PPARγ binding to the stronger motif allele. (B) The same analysis for C/EBP motif-altering SNPs, which also affected PPARγ binding. By Chi-squared test, the difference between SNPs with large and minimal motif effects was P=0.0002 in A and P=0.002 in B. (C) Motif-altering SNPs were interrogated for effects on nearby gene expression in human WAT eQTLs, and the best candidates with large motif effects or imbalance in heterozygotes (as in A-B) showed significant enrichment for eSNPs. (D) For PPARγ or C/EBP motif-altering SNPs with eQTLs, the direction of association was tested, and for the best candidates the stronger motif alleles were associated with higher gene expression. In C and D, **P<0.01, *P<0.05 by Fisher’s Exact Test, one-tailed. See also Table S3.

To test for function regulating nearby human genes, all candidate PPARγ site motif-altering SNPs (regardless of heterozygosity) were integrated with human eQTLs identified in fat biopsies from 1,381 Finnish men in the METSIM cohort (Stancáková et al., 2012). These subjects had genotyping and adipose gene expression analyses, and cis eQTLs fell within 1 Mb of a gene and were significantly associated with its expression level (see Extended Procedures). Within each eQTL, the expression SNP (eSNP) was defined as most significantly associated and thus most likely to be causal. There was a significant 3-fold enrichment of eSNPs in strong candidate motif-altered SNPs relative to an internal control set of PPARγ site SNPs with only minimal motif effects (Figure 5C). SNPs with small or medium motif effects showed intermediate 2.3-fold enrichment, while the strongest candidate PPARγ site SNPs (those from Figure 5A–B showing imbalance in heterozygous subjects favoring the stronger motif) were dramatically and significantly enriched 7-fold for eSNPs (Figure 5C).

The direction of each eQTL was tested to determine whether the stronger motif allele was associated with higher gene expression. For all SNPs in the eQTL dataset, there was an even distribution with 50.09% of reference alleles positively associated with expression. Analysis PPARγ and C/EBP motif-altering eSNPs revealed 70–80% positive eQTL association for those strongest candidate SNPs with observed imbalance in heterozygotes (Figure 5D). By contrast, control eSNPs with minimal motif effects showed the ~50% positive association expected by chance, while greater motif changes led to more positive association with gene expression, strongly indicating causality (Figure 5D). Based on interrogation of this large and powerful eQTL dataset, motif-altering SNPs affecting PPARγ occupancy are functional in determining the expression of nearby genes in human fat.

For example, rs568867 is a highly significant eSNP for TMEM170B gene expression (odds ratio 2.25, P=3.3e-86), and higher expression with the G allele was confirmed in fat biopsies from 23 human subjects (Figure 6A). TMEM170B encodes an uncharacterized transmembrane protein, and its mRNA is strongly induced during mouse and human adipocyte differentiation (Figure 6B) consistent with PPARγ gene regulation. The SNP falls in a PPARγ site 12kb upstream, in a DR1 with the G allele giving a stronger motif (Figure 6C), and ChIP from a heterozygous subject confirmed selective PPARγ binding to the G allele (Figure 6D). Furthermore, an exonic SNP in the TMEM170B 3’ UTR allowed measurement of allelic imbalance in mRNA expression. 11 subjects were heterozygous for this mRNA SNP, and imbalanced TMEM170B expression was only observed in those 3 subjects G/A heterozygous at the PPARγ binding site SNP (Figure 6E). This correlation is highly suggestive that imbalanced PPARγ binding upstream of TMEM170B causes imbalanced gene expression, and other human regulatory SNPs likewise affect PPARγ binding and function.

Figure 6. Human TMEM170B gene expression driven by a polymorphic PPARγ site.

(A) In 23 subjects, the G:A eQTL SNP rs568867 was genotyped and WAT TMEM170B gene expression was measured by Q-RT-PCR, validating genotype effects on expression (*P<0.05 versus G/A by Mann-Whitney test). (B) Murine 3T3-L1 or human SGBS cultured cells were differentiated to adipocytes and treated with rosiglitazone, with effects on TMEM170B gene expression. (C) Location of rs568867 PPARγ motif-altering SNP and a weakly linked (r²=0.513) SNP rs295051 in the mRNA 3’ UTR. (D) ChIP allelic imbalance assay in a G/A heterozygous subject. (E) Allelic imbalance in TMEM170B mRNA was dependent on the rs568867 PPARγ site genotype. In B, D, and E, ***P<0.0001, **P<0.01, *P<0.05 by T-test, 2-tailed, unpaired.

SNP-dependent PPARγ binding underlies human metabolic phenotypic differences

Human PPARγ site motif-altering SNPs were interrogated in published GWAS meta-analyses for 8 metabolic traits (see Extended Procedures). SNPs that altered PPARγ or C/EBP motifs in PPARγ sites were highly enriched among SNPs associated with triglyceride (TG) and HDL cholesterol (HDL) levels (Figure 7A). For TG, the odds ratio was 1.84 (95% confidence interval of 1.27–2.66), and for HDL the odds ratio was 1.76 (1.22–2.56). In both cases, this nearly 2-fold enrichment was beyond that expected by chance, even accounting for multiple testing.

Figure 7. PPARγ site motif altering SNPs affect human metabolic GWAS traits.

(A) Candidate human SNPs in PPARγ sites with more than minimal effects on PPARγ or C/EBP motifs were interrogated for enrichment in loci with a threshold P<0.01 for 8 GWAS traits: lipid levels (TG, HDL, LDL), body mass index (BMI), waist hip ratio corrected for BMI (WHR), T2D, coronary heart disease (CHD), and systolic blood pressure (SBP). (*Chi-squared P<0.00625, significance threshold with Bonferroni correction for multiple testing of 8 traits). (B) One HDL-associated PPARγ motif-altering SNP (rs392794) fell ~300kb from the ANKRD55 gene, with the T allele giving a stronger motif. (C) Luciferase reporters with the T and C alleles at this SNP were transiently transfected into 293T cells with or without expression plasmids for PPARγ and RXRα, and the T allele reporter gave higher PPARγ-dependent activity (*P<0.01 by unpaired T-test, two-tailed). (D) Locus zoom plot showing this SNP has a genome-wide significant association with HDL, and it was not linked to the nearby previously-described lipid association at SNP rs9686661, with a low r² value and a recombination hot spot (red arrow) between the SNPs. (E) In addition to the HDL association (orange), this SNP had significant associations with other metabolic traits (P<0.005 in green). For D and E, associations for rs392794 are via the near perfect proxy rs459193 (r²=0.957) which was genotyped in the GWAS.

Since PPARγ site motif-altering SNPs are enriched for dyslipidemia trait associations, some may be mechanistically implicated. We focused on an HDL-associated PPARγ motif-altering C:T SNP (rs392794), for which the T allele favors PPARγ binding (Figure 7B). There was significantly higher PPARγ-dependent activation of a reporter with a T allele (31-fold) versus the C-allele (7-fold, Figure 7C). The T allele that gives higher PPARγ transcriptional activity is associated with lower HDL, and this association reached genome-wide significance at P=8.1e-09 (below the standard threshold of P=5e-08). This novel association is independent of a reported dyslipidemia locus ~55kb away marked by another SNP (Global Lipids Genetics Consortium et al., 2013) which is not tightly linked (Figure 7D, red arrow indicates a recombination hotspot between the SNPs). Furthermore, conditional analysis to remove the effects of the previously reported SNP showed that the PPARγ motif-altering SNP maintained a strong independent association with HDL (P=3.3e-06). In addition to HDL, this PPARγ motif-altering SNP rs392794 was also found to be associated with other metabolic traits, including serum TG, waist-hip ratio, and systolic blood pressure (Figure 7E). Thus, PPARγ occupancy and transcriptional activity at this site in human WAT may provide the mechanism for association with multiple traits that are part of the metabolic syndrome (Grundy et al., 2004).

DISCUSSION

Non-coding regulatory variants are key to understanding differences among individuals. Here we show that SNPs affect TF genomic occupancy, gene expression, and drug response in adipose tissue. Moreover, PPARγ site SNPs in human fat may modulate genetic risk of metabolic disease.

Previous studies have shown that TF binding among mammalian species differs in numerous cases even when the motif is intact (Mikkelsen et al., 2010; Schmidt et al., 2010; Soccio et al., 2011), and many are accounted for by variants affecting co-bound TFs (Schmidt et al., 2011; Stefflova et al., 2013). Here we show this applies within a species, as SNPs affect PPARγ occupancy in WAT by altering motifs for PPARγ itself or other cooperating TFs. A similar phenomenon was reported in macrophages from inbred mouse strains, where SNPs altering PU.1 motifs and binding also affected binding of C/EBPα, and vice versa (Heinz et al., 2013). The mechanism likely involves cooperativity by assisted loading, which has been shown for recombinant PPARγ and C/EBPα in immortalized cell lines (Madsen et al., 2014). The present study demonstrates that these principles apply in living tissues in mice and man.

Motif-altering SNPs accounted for ~20% of strain-selective PPARγ binding between two mouse strains, even when only the top four TF motifs in PPARγ binding regions were considered. The majority of SNPs in strain-selective sites failed to alter recognizable motifs, consistent with a recent report that GWAS causal variant SNPs are highly enriched in tissue-specific enhancers, yet only 10–20% directly alter motifs while most are “motif adjacent” (Farh et al., 2015). There was also clear strain-selective binding at non-polymorphic PPARγ sites, and we provide evidence for a novel mechanism based on the observed clustering of strain-selective sites: a causal polymorphic site may drive differential binding at other sites by long range interactions within a regulatory locus. Conversely, we observed numerous cases of motif-altering SNPs that failed to give strain-selective PPARγ binding, consistent with prior reports and presumably attributable to factors such as specific motif alterations, distance to peak center, presence of alternative motifs, and “buffering” by site context (Heinz et al., 2013; Maurano et al., 2012).

We found many examples in which genes nearest to mouse strain-selective PPARγ sites showed differential expression, and this enrichment was evident genome-wide for basal or rosiglitazone-induced adipose gene expression. The consistent direction of PPARγ occupancy and gene regulation effects (i.e. B6-stronger motif SNPs with B6-selective binding and B6-higher gene expression) argues strongly for causality. Similarly in humans, motif-altering PPARγ site SNPs showed enrichment for eQTLs in the predicted direction. Our approach used SNPs as “experiments of nature” to identify candidate functional binding sites among the multitude of sites that surround some genes, but further experiments will be necessary to elucidate the mechanisms underlying “functional” versus “non-functional” binding events, which is a major unanswered question in the field and may involve enhancer activity and/or 3D-looping events. Indeed, the majority of strain-selective sites did not have a detectable effect on the nearest expressed gene (though the actual targets may be more distant). Apparently non-functional regulatory variation may result from multiplicity and redundancy of enhancers around target genes (Cusanovich et al., 2014), as adipocyte gene expression has been correlated with the number of nearby PPARγ sites (Step et al., 2014). Just as PPARγ regulatory variation is not always functional, only a subset of differential genes are explained by PPARγ regulatory variation, as myriad other factors could affect adipose gene expression or response to TZD – yet our genome-wide methods had the power to identify the important cases of functional regulatory variation.

In addition to regulating basal adipose tissue gene expression, genetic variation at PPARγ sites also determines response to PPARγ ligand rosiglitazone. While there were more strain-selective TZD-induced genes than could be accounted for by PPARγ sites, and thus other strain effects drive some of the observed variability in drug response, genetically determined differential PPARγ binding emerged as a clear mechanism for some strain-selective rosiglitazone effects. This provides proof-of-concept that naturally occurring genetic variation can affect nuclear receptor ligand-mediated gene activation and, more generally, drug response in living animals. This has special significance for TZDs, which have powerful antidiabetic effects but limited clinical utility due to adverse effects including bone loss and edema (Soccio et al., 2014). The target genes responsible for TZD efficacy and side effects are unknown, but genetically-determined differences in TZD activation of such genes may underlie inter-individual differences in drug response. PPARγ cistromes are cell-type specific (Lefterova et al., 2010), and thus a given SNP might differentially impact the therapeutic and harmful effects of TZDs, particularly since many TZD side effects likely reflect non-adipose tissues. Understanding regulatory variation could lead to personalized PPARγ agonist therapy based upon individual profiles of SNPs that alter PPARγ binding.

There is great potential linking regulatory variation to human genetics of disease susceptibility revealed by GWAS. A recent effort to identify causal GWAS SNPs and their tissue effects showed that SNPs affecting TG and HDL fell in enhancers from adipose tissue (Farh et al., 2015). The present work shines a bright light on the role of PPARγ in this context, as the polymorphic PPARγ sites that we identified in human fat were enriched for the same dyslipidemic GWAS traits.

Detailed study of one PPARγ motif-altering SNP, rs392794, revealed its association with several metabolic syndrome traits. Of note, in GWAS meta-analyses a near-perfect proxy (rs459193) has been identified as the lead SNP in finely-mapped loci for both T2D (Morris et al., 2012) and waist-hip ratio (Shungin et al., 2015). Thus, the PPARγ site SNP identified here could be responsible for these associations with T2D and visceral adiposity, as well as the other metabolic traits reported here. While the GWAS locus is named for ANKRD55, this gene is quite distant (TSS 279kb away) and not expressed in adipose tissue, and we were unable to link the metabolism-associated PPARγ site to ANKRD55 or any other gene. Thus the target gene for this PPARγ-bound enhancer and the GWAS effect is unknown, and further studies will be necessary to understand how differential SNP-dependent PPARγ occupancy at this site may alter systemic metabolic phenotypes.

In sum, SNPs affect adipose PPARγ genomic occupancy and the basal and drug-induced expression of nearby genes. SNPs that alter PPARγ binding are enriched in loci affecting metabolic traits, such that they contribute to an individual’s metabolic disease risk. They also impact personalized pharmacogenomics, as better understanding of the tissue-specific functions of such SNPs has the potential to improve the risk/benefit ratio for TZD therapy in individual patients based on their genomes. The implications of this work go beyond PPARγ to all drug targets that function directly at the genome to regulate physiology in health and disease.

EXPERIMENTAL PROCEDURES

Mouse and human adipose tissue samples

Male wild type inbred C57Bl/6J and 129S1/SvImJ mice, in addition to the F1 intercross progeny (B6129SF1/J), were purchased from Jackson Laboratories, and all care and use procedures were approved by the Institutional Animal Care and Use Committee of the University of Pennsylvania. Human fat samples (Table S3) were obtained from the Human Adipose Resource (HAR) of the Penn Institute for Diabetes, Obesity, and Metabolism, which obtains pre-operative informed consent from surgical patients for biopsies to be taken, banked, and distributed to investigators with de-identified patient characteristics. All HAR protocols were approved by the University of Pennsylvania’s Institutional Review Board.

ChIP-seq, RNA-seq, and GRO-seq

Three related experiments compared PPARγ ChIP-seq in WAT from male mice of the B6 and 129 strains, with minor differences in age, depot (epididymal versus inguinal), and diet (chow versus low fat control diet, Table S1). The great concordance among the three experiments allowed high confidence identification of genetic effects. The ~5.3 million known SNPs differing between B6 and 129 mice (Keane et al., 2011) were incorporated into SNP-sensitive GSNAP read alignments (Wu and Nacu, 2010). The HOMER software suite (Heinz et al., 2010) was used for peak identification and quantification. Motif identification and determination of SNP effects on motifs was performed as described in the Extended Procedures. Four related experiments compared WAT RNA-seq from male mice of the B6 and 129 strains, with one also measuring the effect of rosiglitazone treatment for 2 weeks (Table S2). GRO-seq was performed, and eRNAs were identified and quantitated as previously described (Fang et al., 2014), with minor modifications for mouse WAT samples. All deep sequencing (single end 50 or 100bp reads) was performed by the Functional Genomics Core of the Penn Diabetes Research Center using Illumina HiSeq2000 and the Solexa Analysis Pipeline. Allelic imbalance was assayed in aligned sequence read BAM files using an allele counting PERL script.

Other methods

Quantitative PCR was performed using ABI 7500 Fast Real-Time PCR System and Power SYBR Green PCR Master Mix (Applied Biosystems). Allelic imbalance was confirmed at individual loci using the SNaPshot assay (Applied Biosystems). See Extended Procedures.

Statistical methods

Prism (Graphpad) was used for graphing and statistical tests, all of which are described in figure legends.