Recently, we characterized the in vitro binding preferences of the three PPAR subtypes on a panel of 39 systematic single nucleotide variations of the consensus DR1-type PPRE (AGGTCAAAGGTCA) 18. Based on this analysis we subdivided the single nucleotide variants into three classes (Table 1). Sequences in class I are bound by the PPAR subtypes with a strength of 75 ± 15% of that of the consensus PPRE; sequences in class II are bound with a strength of 45 ± 15% of that of the consensus PPRE; and sequences in class III are bound with a strength of 15 ± 15% of that of the consensus PPRE. Although the overall binding pattern of the three PPAR subtypes showed no major differences, some variations gave rise to a PPAR subtype-specific classification. We observed that the number and class of variations seem to correlate with experimental binding. Therefore, we decided to take the concept further to create a classifier for PPREs based on binding data. We sorted a total of 136 DR1-type response elements (REs; including combinations of multiple variations) according to the number and class of variations (Figure 1). The in vitro binding strength to these REs in relation to the consensus DR1-type PPRE was determined by gelshift assays for the RXR heterodimers of all three PPAR subtypes. For each category in Figure 1 the average of the relative binding strength was determined (based on 6 to 47 RE/PPAR subtype combinations). REs with 1/0/0, 2/0/0 and 0/1/0 variations (where the numbers indicate the number of variations for the classes I, II and III, respectively) bound the receptor strongly (67%, 43% and 39% relative binding, respectively), REs with 3/0/0, 1/1/0 and 0/0/1 variations were medium PPREs (29%, 22% and 20%, respectively) and REs with 0/2/0, 2/1/0, 1/0/1, 3/1/0 and 4/0/0 variations were considered to be weak PPREs (8%, 4%, 3%, 1% and 1%, respectively). We set 1% as a cut-off limit. Representative DR1-type REs with increasing numbers of more drastic variations were examined as well (Additional data file 1), but these elements were not considered as functional PPREs. Please note that the published PPRE of the acyl-CoA oxidase 1 (ACOX1) gene 19 belongs to the latter list.

The performance of the classifier in predicting novel binding sites was simulated by random sampling of the experimental data in Figure 1 and Additional data file 1 into a training set that was used to re-calculate the category averages at each initialization (approximately 10% of data was used in training) and a validation set that can be used in testing (rest of the data). Representative data from 10 rounds of simulation are shown in Additional data file 2. Interestingly, the category averages were relatively robust to changes in the set of sequences used to calculate the average. This suggests that the introduction of further sequences that belong to the same category will not drastically affect the classifier performance.

In order to compare the classifier to the traditional matrix methods, we created a position-specific weight matrix (PSWM) and a position-specific affinity matrix (PSAM) using the PPARγ data from Figure 1 and Additional data file 1. For the PSWM we took all medium and strong PPREs with multiple variations from Figure 1, calculated base pair frequencies and converted these to matrix values by logarithmic transformation, where an equal background frequency was assumed and a pseudocount of 0.01 was included for non-observed base-pairs (bp). We chose not to include the systematic single nucleotide variation screen data, since this would have biased the matrix strongly towards the consensus PPRE. In total, 20 sequences were used to construct the matrix, which is in the order of known binding sites typically used as a basis of such matrices in databases, such as JASPAR or TRANSFAC. The PSAM was chosen to represent a matrix method utilizing the single nucleotide screening data, in order to see if these data are sufficient to capture the binding preferences of multiple variation data.

We compared the three methods first on the level of their ability to detect binding. True positive and false positive rates (TPRs and FPRs, respectively) were calculated using different cut-off values for each method and are represented in the form of a receiver operating characteristic (ROC) curve (Figure 2a). The line of no discrimination is indicated as a diagonal line; perfect performance would give a TPR of 1 and FPR of 0. For all methods an optimum performance was detected with FPR from 20-30% and TPR varying from around 90% for the PPRE classifier to 75% for the PSAM. For clarity, one representative classifier curve out of ten calculated is shown.

Next we wanted to know whether the scores correlated with experimental binding when comparing single and multiple variation data. We examined this with correlation plots using the PPARγ data as shown in Additional data file 3. In parallel, we set a tolerance interval of 15% relative to the consensus sequence for a match between predicted binding strength and experimental binding (5%, if the predicted binding was less than 15%) and calculated predictions by the different methods. The equations of the lines fitted to the single nucleotide data (Additional data file 3) were used to correlate matrix scores with binding strength. The ideal cut-off values based on the ROC curves were used in the scoring and produced respective data points in the ROC space (Figure 2b), this time with TPR reflecting correct predictions (no underestimation, if 1) and FPR reflecting overestimated values. Several data points are given for the classifier, representing ten separate initializations with the sampling of training and validation sets.

When comparing the performance of the PSWM between the different datasets (Additional data file 3), a rather clear distinction between the scores of single nucleotide variations (medium and strong PPREs) and the non-binding PPRE classes was observed. The partition of single nucleotide data into two groups of data points shows that the matrix handles variations that were not included in the PPRE set by penalizing these with a constant negative score. Values above 6 still separated quite well from the data points of the last panel. However, the multiple variation data that include weak to strong PPREs were not well resolved by the matrix. Instead a large amount of weak binding sites received high matrix scores, which seems to cause the high FPR rate.

Despite the fit to the single nucleotide data, the PSAM did not offer a significant improvement to the prediction of multiple variations and also had problems differentiating the non-binding PPREs. This is evident by examining the data points between matrix values 0.000001 and 0.000002. This interval includes weak to strong PPREs with identical matrix scores leading to an increased FPR rate. The classifier correlation was weaker for single nucleotide data compared to the PSAM, but the same variation was preserved for multiple variation data. A clear separation between weak PPREs and those of medium and strong strength was achieved. The ability to use a PPRE prediction that also correlates with binding strength is a clear advantage for the evaluation of putative binding site content of target genes. Based on the different comparisons, we chose the PPRE classifier as most suited for the follow-up analysis of PPAR target genes.

We tested the performance of our PPRE binding strength prediction scheme on eight primary PPAR target genes. We selected the well-known up-regulated human genes ACOX1 19, carnitine palmitoyl transferase (CPT) 1B 20 and PPARα 21 and the established down-regulated gene apolipoprotein (APO) C3 22. The genes angiopoietin-like 4 (ANGPLT4) 23, sulfotransferase (SULT) 2A1 24 and Rev-ErbAα (RVRα) 25 were chosen because their PPREs were at unusual positions, such as in an intron or more than 5 kB upstream of their TSS, or of unusual structure, such as a direct repeat with two intervening nucleotides (DR2). Finally, the gene UCP3 26 was included, because despite being an established PPAR target, no PPRE had yet been characterized within its previously studied regulatory regions. Therefore, the latter gene was a specific challenge to our PPRE prediction approach. By real-time quantitative PCR we confirmed the inducibility of all eight genes by PPAR ligands (Additional data file 4) and demonstrated in parallel that our experimental systems, the human cell lines HEK293 (embryonal kidney) and HepG2 (hepatocarcinoma), with the exception of the APOC3 gene in HEK293 cells, are well suited for the investigation of these genes.

For the eight PPAR target genes we performed an in silico PPRE search, which spanned 10 kB upstream and downstream of the respective TSS (Figure 3). All PPRE categories that included PPREs with 5% or more binding strength for each subtype are shown. The categories resulting in 1-5% of binding (1/0/1, 3/1/0 and 4/0/0) were indicated only when the PPREs were conserved in the mouse genome. Based on sequence alignments of the human and mouse genome, the evolutionary conservation of all putative REs was evaluated on the level of the RE itself and the level of its flanking sequence (± 50 bp). As a result, we found 5 REs in each of the genes ACOX1, CPT1B, SULT2A1 and ANGPTL4, 9 in the APOC3 gene, 4 in the PPARα gene, 7 in the RVRα gene and 6 in the UCP3 gene, giving rise to a total of 46 REs in the 160 kB genomic sequence examined. The distribution of the putative REs, relative to the TSS, was roughly equal, since 21 and 25 were found in the upstream regions and downstream areas, respectively. In a cross-species comparison (mouse to human), 10 of the 46 REs were found to be evolutionarily conserved and a further 6 REs were located in conserved regions. Our in silico screening found the published PPREs of the genes ANGPTL4, APOC3 and CPT1B as evolutionarily conserved REs and the published PPREs of the genes SULT2A1 and PPARα as non-conserved. As mentioned above, the published RE of the ACOX1 gene did not pass our in silico screening parameters and we confirmed by gelshift assays that it does not bind PPARs (Additional data file 1). This observation concurs with a previous report 27. However, in that study it was claimed that the human ACOX1 gene may not be an active PPAR target, whereas here we show that the gene is regulated by PPARs and suggest five new binding sites, of which one is located in an evolutionarily conserved area of intron 1.

The in silico binding strength predictions were confirmed by gelshift assays for the six REs of the UCP3 gene (novel sequences that had not been used for average calculations in Figure 1). Comparing the experimentally determined and the calculated values, all predicted binding sites match the experimentally determined binding strength with a deviation of less than 15%.

Taken together, in silico screening predicts that, for each of the eight tested PPAR target genes, there are four to nine PPREs within 10 kB of their respective TSSs, of which at least one is a strong PPRE. The example of the UCP3 gene demonstrates the good correlation between in silico prediction of PPREs and actual in vitro binding of PPAR-RXR heterodimers.

We selected within the regulatory regions of the eight PPAR target genes 10 proximal REs (within 1 kB of the TSS), 10 REs further upstream and 10 REs further downstream (the element of the APOA1 promoter element was counted as a proximal RE because the gene is a known responding gene 28). These REs are contained within 23 genomic regions (each approximately 300 bp in length; for locations see Figure 3 and Table 2), which we cloned by PCR and fused with the thymidine kinase promoter driving the luciferase reporter gene. We included the ACOX1 published region, in which we do not predict a PPRE, as a negative control. The activity of the constructs in the absence or presence of PPAR subtype expression vectors in response to PPAR subtype-specific ligands was tested by reporter gene assays in HEK293 and HepG2 cells (Figure 4). Nine of the genomic regions are located within 1 kB of their respective TSSs (Figure 4a,d). With the exception of the RVRα gene TSS, which contains a reported DR2-type PPRE, eight of these regions displayed, in at least one of the two cell lines, significant inducibility by PPAR ligands. The region of the human CPT1B gene was inducible by all three PPAR subtype-specific ligands in both cell lines, whereas the seven other regions show PPAR subtype- and cell type-specific profiles. An increase in the basal activity compared to empty cloning vector and its subsequent loss due to PPAR over-expression were observed with the proximal regions of the genes APOC3 and UPC3 in both cell lines as well as in HepG2 cells with the intron 1-containing region of the ACOX1 gene and the proximal region of the APOA1 gene. This effect may reflect the attraction of constitutively active transcription factors, such as other nuclear receptors that recognize DR1-type REs, for example, HNF4α, to the respective genomic regions and their subsequent displacement 22. The cellular context may permit stronger activation by the displaced transcription factor, for example, due to higher expression of favored coregulator interaction partners. This switching of activating transcription factor to the binding site could offer one explanation for the observed change in the basal expression level.

Of the nine upstream regions, the region of the SULT2A1 gene was shown to be the most active (Figure 4b,e). In both cell lines over-expression of PPARs clearly increased this fragment's basal activity as well as significant inducibility by all three PPAR ligands. A similar observation was made in HepG2 cells for the distal region of the UCP3 gene, an effect that was far more modest in HEK293 cells. In contrast, neither the distal regions of the genes ACOX1 and RVRα nor the region containing the published PPRE of the gene ACOX1 displayed any inducibility by PPAR ligands in either of the two cells lines. Therefore, they can be considered as negative controls. In addition, the distal regions of the genes ANGPTL4 and APOC3 were only inducible in HEK293 cells, whereas the PPARα gene's putative PPRE-containing region responded only in HepG2 cells to GW501516 treatment. Interestingly, in HEK293 cells, the distal regions of the genes ANGPLT4, APOC3 and PPARα showed the already described effects of increased basal activity with endogenous activators and subsequent suppression of the activity by PPAR subtype over-expression.

Of the five downstream regions, the intron 2 region of the ACOX1 gene and the cluster region of the ANGPTL4 gene (containing four putative PPREs) displayed a clear response to all three PPAR ligands in both cell lines. In contrast, the inducibility of the intronic region of the APOC3 gene was far more modest (Figure 4c,f). Individual mutagenesis of the ANGPLT4 REs was carried out and this resulted in reduced activity, thus demonstrating that the other REs, in addition to the published PPRE, contribute to the responsiveness of this region (data not shown). Finally, the cluster and intronic region of the UCP3 gene responded only in HEK293 cells to GW501516 treatment.

In summary, of the 23 investigated genomic regions containing putative PPREs, up to 17 display significant inducibility in the presence of PPAR ligands (Table 3).

The same 23 genomic regions of the eight PPAR target genes were investigated by chromatin immuno-precipitation (ChIP) assays with chromatin extracts from HEK293 cells (or from HepG2 cells for regions from the APOC3 gene) that were treated with solvent or for 120 minutes with the PPARα ligand GW7647 (Figure 5). We assessed these regions for the binding of PPARα, its partner receptor RXRα and pPol II (the latter as a sign for a direct connection between the RE-containing region and the TSS). Chromatin templates were analyzed by quantitative real-time PCR and the specificity of the antibodies for the three proteins was compared with the non-specific background binding to IgG. Of the 23 tested regions, the region of the CPT1B gene, the distal and published region of the ACOX1 gene, the distal 1, distal 2 and intronic region of the APOC3 gene and the cluster of the UCP3 gene did not show specific binding of any of the three proteins. For the two regions of the ACOX1 gene this result confirmed their failure in the previous functionality test (Figure 4). The 16 other regions showed a significant association with PPARα in the presence of ligand. When comparing the relative association levels of PPARα under these conditions, we found that the most prominent binding was to the region of the SULT2A1 gene, followed by the regions of the RVRα TSS and the proximal region of the PPARα gene (Figure 5d). Interestingly, the latter two regions as well as the proximal regions of the genes APOA1 and UCP3, the distal region of the RVRα gene and the distal and intronic region of the UCP3 gene even displayed ligand-independent binding of PPARα. Similarly, a GW7647-independent association of RXRα was found on the published region of the APOC3 gene, on the proximal regions of the genes APOA1, PPARα and UCP3 and on the distal regions of the genes ANGPTL4 and UCP3. In contrast, no statistically significant binding of pPol II, irrespective of the presence of ligand, was found on the published region of the APOC3 gene and in the distal regions of the genes ANGPTL4 and RVRα.

Taken together, PPARα and RXRα associate in living cells with 16 of the 23 genomic regions. Thirteen of these regions also associate with pPol II, twelve of which show functionality in reporter gene assays (Figure 4, Table 3). With the exception of the CPT1B gene, the tested PPAR target genes possess one to three of these tested regions. The regions show neither positional bias nor do they preferentially contain evolutionarily conserved PPREs. However, each of them contains at least one predicted medium or strong PPRE. Three of the twelve fully functional regions, the ANGPTL4 cluster region, the proximal region of the PPARα gene and the distal region of the SULT2A1 gene, were already known, but we identified two alternative regulatory regions for each of the genes ACOX1 and APOC3, and one for the RVRα gene. We also found one additional regulatory region for the PPARα gene. Finally, for the UCP3 gene, for which no regulatory regions had so far been described to account for the effect of PPAR ligands on its mRNA transcription, we identified three functional areas.

The common feature of the eight investigated PPAR target genes appears to be a prevalence for strong PPREs at a distance of up to 10 kB from the TSS. With the aim of extending this conclusion, we next compared all human genes that are known as primary PPAR targets. The genes were selected according to the following criteria: mRNA or protein level changes were reported for human cells or tissues; a PPRE was described for the human gene and it was experimentally confirmed by either gelshift, reporter gene or ChIP assay. This resulted in 30 additional genes, for which we performed in silico analysis for putative REs up to a distance of 10 kB from their respective TSSs (as done for the first eight genes; see Figure 3). In addition, for all of the 38 genes, the orthologous mouse gene was investigated in the same way (Figures 6 and 7). From these data, overview figures for each gene were constructed that show the location of the PPREs (x-axis) and their respective predicted binding strength (y-axis). In order to reveal further characteristics of the genes and their PPREs, such as overall similarity of their patterns and evolutionary conservation, they were clustered using a self-organizing map (SOM) algorithm followed by Sammon's mapping to illustrate the clusters. The input dataset of the SOM consisted of six variables, which are the sum of the predicted binding strength (BS), the number of conserved strong and medium binding sites (CS) and the number of weak binding sites (CW) both for the human and the mouse gene (Additional data file 5). An initial map resulted in four clusters, which were then each separated in the final map in up to five subclusters (Figures 6 and 7).

While the evolution of protein-coding sequences within genomes is well understood, the same cannot be said of the regions regulating transcription. The conservative model, often utilized as a limit for in silico binding site predictions, requires the strict preservation of binding site sequence and location. However, recent research on enhancer evolution has challenged this view. From these data, it appears possible to maintain overall functional conservation of regulatory elements over long evolutionary times, despite having high structural architecture turnover rates 16. This feature has not been widely explored with human target genes, but is important to consider for target gene prediction. We therefore sought to identify traces of binding site evolutionary patterns in the clusters of this dataset. In general, clusters I and II contain genes that are well conserved between human and mouse, while the genes in cluster III are more divergent and those in cluster IV show no evolutionary conservation of PPREs.

Cluster I contains genes that carry multiple conserved PPREs (Figure 6a). Cluster IA contains the genes ANGLPTL4, CPT1A 29, lipoprotein lipase (LPL) 30 and phosphoenolpyruvate carboxykinase (PEPCK) 1 31, which have, in both human and mouse, multiple conserved regions with strong REs. While each gene contains a conserved medium/strong element, a subset of genes expanded this set: human CPT1A and mouse LPL show significant enrichment of de novo binding sites compared to their orthologs. Cluster IB contains the genes glycerol kinase (GK) 32 and UCP3, which have multiple conserved RE regions. The distal UCP3 PPRE is conserved as a strong PPRE but is outside the 10 kB window in mouse. The cluster of PPREs in the human GK promoter seems to have lost significance in the mouse. The lipoprotein receptor-related protein 1 (LRP1) 33 gene represents cluster IC, in which multiple conserved REs are strong in mouse but weaker in human. Together with the genes GK and LRP1, the genes found in cluster ID, caveolin 1 34 and insulin-like growth factor binding protein 1 (IGFBP1) 18, exhibit a retain-loss pattern concerning conserved PPREs, where only one species retained a cluster of strong PPREs. In the case of the LRP1 gene, this appears to have arisen in the context of several compensating de novo binding sites.

Cluster II differs from cluster I by having one or two strong or medium conserved REs in human, which are found in a comparable strength and location in mouse (Figure 6b). This cluster is subdivided into two clusters. Cluster IIA contains the genes APOC3, CPT1B, CPT2 35, cytochrome P450 (CYP) 1A1 36, 3-hydroxy-3-methylglutaryl-CoA synthase 2 (HMGCS2) 37 and scavenger receptor B1 (SRB1) 38, which have relatively comparable PPRE content. In contrast, cluster IIB contains the genes adipose differentiation-related protein (ADRP) 39, APOA1, G0/G1 switch gene 2 (G0S2) 40, liver X receptor (LXR)α 41 and spermidine/spermine N1-acetyltransferase (SSAT) 42, which exhibit an increase in PPRE content in the mouse gene.

Cluster IIIA, which contains the genes cyclooxygenase 2 (COX2) 43 and semaphorin (SEMA) 6B 44, extends the pattern observed above with an opposite trend; the human ortholog contains one or two medium/strong REs, which are conserved but only weak in the mouse (Figure 7a). Possible compensating elements appeared in the mouse SEMA6B gene, while this is not the case for the mouse COX2 gene. Also, cluster IIIB contains one or two conserved REs, but they are weak in both human and mouse. This cluster comprises the genes APOA2 45, APOA5 46, fatty acid desaturase 1 (FADS2) 47, pregnane X receptor (PXR) 48, RVRα and solute carrier (SLC) 10A2 49. Interestingly, these genes each have novel binding sites at nearly similar locations.

Cluster IV contains genes that carry one or more REs, but none of them is conserved (Figure 7b). The ACOX1 gene represents cluster IVA, in which multiple strong, but non-conserved, REs are found in both species. The genes resistin 50 and SULT2A1 form cluster IVB; they have one or two strong non-conserved REs in human and multiple REs in mouse. The genes APOE 51 and PPARα are in cluster IVC, which is characterized by one strong RE in the mouse ortholog and one or more non-conserved REs in the human gene. In cluster IVD are the genes CYP27A1 52, glutathione S-transferase (GST) A2 53, transferrin 54 and UDP-glycosyltransferase (UGT) 1A9 55, which carry one or two medium, non-conserved REs in both species. Finally, the cyclin-dependent kinase inhibitor 2A (CDKN2A) 56 gene represents cluster IVE, in which strong or medium non-conserved REs are found in mouse and but only weak REs are found in human.

In summary, SOM clustering of the 38 presently known human PPAR target genes sorts them into four clusters, of which the first three contain different numbers of evolutionarily conserved REs, while the 10 genes of cluster IV are characterized by having non-conserved REs. Interestingly, although for some genes a conservation of the PPRE pattern is evident, significant diversity in the composition of PPREs is visible as well.

In order to explore the evolutionary preservation patterns of PPREs further, the genes ACOX1 and ANGPLT4 from the genomes of chicken, chimpanzee, dog, rat and zebrafish were also analyzed (Figure 8). In the genome of the chimpanzee (Pan troglodytes), the closest relative to human, four conserved PPREs were located in the ACOX1 gene, but the intron 2 region is missing. The respective human PPRE is not conserved in any of the species analyzed, suggesting that it is human-specific. Mouse and rat (Rattus norvegicus) share two PPREs, although both also contain a unique set of further REs. The analysis of the ACOX1 gene in chicken (Gallus gallus), dog (Canis familiaris) and zebrafish (Danio rerio) is in accordance with the overall pattern of relatively species-specific composition of PPREs. Within mammals, the cluster of intronic REs of the ANGPLT4 gene is rather well conserved. The closest PPRE pattern in comparison to the human gene is observed in the dog. Two intronic PPREs are also present in rat, but a significant expansion in PPREs seems to have occurred in this species, including a distal consensus RE. The zebrafish ANGPLT4 gene is also profoundly enriched with strong PPREs, whereas the chicken gene has a quite poor PPRE content. Moreover, the loss of the published PPRE is observed in the chimpanzee gene. Therefore, while this gene is an example of a much more preserved PPRE pattern, significant diversification is evident amongst the genomes analyzed and not all functional PPREs are conserved.

In conclusion, the SOM analysis and further genome comparisons together suggest that functional RE choice is evolutionarily flexible with respect to first gaining and then maintaining responsiveness to PPARs. In effect, integration of the stabilizing selection model into target gene identification and characterization may more faithfully identify PPAR targets.

For the purpose of target gene identification, the SOM analysis (Figures 6 and 7) indicated that either the presence of at least one strong PPRE or more than two medium PPREs within the 20 kB surrounding the annotated TSS of a gene is a strong indication for a PPAR target gene. In this way, 28 out of the 38 human genes (74%) would have been identified as PPAR targets. Similarly, for 29 of these 38 genes (76%) the analysis of their murine ortholog would have come to the same conclusion. A combination of these two criteria (passing the threshold in either the human or mouse ortholog) would have identified 37 of the 38 genes (97%) as PPAR targets. UGT1A9 could be detected from the list of genes having two medium REs. While location clearly is not a major determinant of PPRE functionality, the screening of these regions in 500 bp windows indicates modest enrichment in the proximal promoter (24% of genes have a medium/strong PPRE close to the TSS; data not shown).

To explore this concept further, we selected the gene-dense human chromosome 19 (63.8 MB, 1,445 known genes in Ensembl, release 41) and its syntenic mouse regions (956 genes have known orthologs) and screened both species for medium and strong PPREs (based on a PPARγ prediction). We extracted from four human microarray datasets 8575859 those genes located on chromosome 19 that were shown to be regulated and determined whether these passed the criteria derived from the 38 target genes (that is, having a strong PPRE, a proximal medium PPRE or more than two medium PPREs). Typically, each dataset contained 6 to 8 genes from chromosome 19, out of which at least 5 (71-100%) passed the criteria in either or both species, and typically 1 or 2 genes had 2 medium PPREs only. This implies that the sensitivity of detecting targets based on our screen for medium and strong PPREs is high.

The background frequency of strong PPREs calculated by the total number of predicted strong PPREs divided by the length of chromosome 19 was found to be 0.66 for the 20 kB region considered for each gene, whereas that of medium or strong PPREs is 1.3. Twenty percent of genes of chromosome 19 contain a co-localizing strong PPRE and an additional four percent have more than two medium PPREs or a proximal medium PPRE. These numbers suggest a total of 4,000 to 5,000 targets for PPARs in the human genome, if no false positives are assumed. A complete evaluation of the selectivity is complicated by the restricted expression profiles of the predicted genes, which prevents simple read-outs from individual target tissues (regulation of only two genes from chromosome 19 was detected on more than two microarrays used earlier). We chose a focused list for further inspection, by requiring detection in human and mouse. In total, 116 genes (12.1%) from chromosome 19 pass the filter (Table 4). Furthermore, all 1,445 human genes were screened for high enrichment of PPREs, that is, for assemblies of at least three medium or strong PPREs (of which at least one was strong). These even more stringent criteria were fulfilled by 50 genes of both species (bold entries in Table 4) and an additional 75 human genes (Additional data file 6), a sample that represented 8.7% of all genes from human chromosome 19.