According to the previously reported linkage results at the 5q locus , we selected 45 subjects (39 obese and six non-obese subjects) in families that contributed to the linkage with obesity. The association study was performed using at first a set of 368 non-obese and normoglycemic subjects ([mean ± SD] age, 57.1 ± 13.5 years, BMI = 22.9 ± 2.3 kg/m2, sex ratio : women/men 221/147) and a set of 292 morbidly obese French individuals (age, 44.7 ± 10.8 years, range 24 to 74, BMI = 47.5 ± 7.8 kg/m2, sex ratio : women/men 232/60). SNPs showing a significant association were genotyped in an extended set of 329 morbidly obese French subjects (age, 48.5 ± 10.5 years, range 24 to 74, BMI = 48.1± 7.1 kg/m2, sex ratio : women/men 229/100) from the same population as the first set of morbidly French obese subjects were extracted from. The 385 severely obese subjects from Zurich, Switzerland that were studied were consecutive, unrelated, Caucasian subjects ([mean ± SEM] age, 43.5 ± 0.5 years, range 24 to 69; BMI = 43.4 ± 0.4 kg/m2; sex ratio : women/men 302/83) referred to the ? clinic for refractory obesity from January 1999 to December 2000; informed written consent was obtained . The set of 619 moderately obese French subjects were characterized as following: age 49.9 ± 13.4 years, range 24 to 88, BMI = 34.5 ± 2.9 kg/m2, sex ratio: women/men 354/265. Two additional sets of control subjects from the French general populations were studied : 546 subjects from the SUVIMAX population  ([mean ± SD] age 55 ± 6 years, BMI = 22 ± 1.8 kg/m2, sex ratio women/men 246/300) and 186 subjects from the WHO-MONICA Lille population  ([mean ± SD] age, 60.6 ± 3.1 years, BMI = 24.7 ± 2.9 kg/m2, sex ratio : women/men 98/88). Further analysis was performed by pooling these data.
Screening and SNP map of the CART gene
We screened for SNPs in 3.7 kb of the plausible promoter region, as well as the exons and the introns of the CART gene by direct sequencing. PCR primers and annealing temperatures are available from authors on request. The protocol was carried out using the 96 capillary ABI PRISM® 3700 DNA Analyzer (Applied Biosystems, Foster City, CA) with the Big Dye Terminator Cycle Sequencing Ready Reaction Kit, as previously described . SNP positions were assigned according to the A of the translation initiation codon (ATG ; Figure 1). The list of identified SNPs and corresponding rs numbers are presented in the additional file (Table 2).
Several genotyping methods were used. E32K (+94G>A), IVS1+114C>T, IVS1+224G>A and IVS1-31C>T were genotyped with PCR-RFLP using the Sac1, Blp1, Apa1 and Mae3 restriction enzymes respectively (New England Biolabs). Promoter fragments containing more than two SNPs were genotyped by direct sequencing. All other SNPs were genotyped with the LightCycler™ assay (Roche, Mannheim, Germany) based on hybridization of probes labeled by two different dyes allowing Fluorescence Resonance Energy Transfer (FRET) . A genotyping quality control was performed by introducing duplicates in the PCR plates and by genotyping all individuals twice. Sequences of primers and conditions of LightCycler assays are available on request.
Statistical analysis for association studies
Hardy-Weinberg proportions for cases and controls were tested by the χ2 test. Allelic and genotypic frequencies differences between cases and controls were assessed by χ2. A region-wide empirical p-value was calculated through permutation. This involved the individual genotype as a whole and the individual's status being shuffled. This method preserves the correlation between SNPs (LD) while breaking the relation between status and the genotypes. For each replicate or permutation each SNP was tested for association and the most significant p-value was stored. We could then compare this p-value to the best observed p-value.
For the haplotype analysis, identification of the minimal set of SNPs that could account for the genotypic diversity was made by systematic enumeration in each block. Haplotypes frequencies were calculated and, after skipping the haplotypes with a frequency lower than 0.02, each SNP and set of SNPs in turn were removed. Thus we found the SNP combination that preserves the marginal haplotype frequencies. This method is implemented in the STRATEGY software. Effects of haplotype were tested using the THESIAS (Testing Haplotype Effects in Association Studies) software. The objective of the program is to perform haplotype-based association analysis in unrelated individuals. This program is based on the maximum likelihood model described in Tregouet et al. (2002) and is linked to the SEM algorithm . The effect of haplotypes with a frequency lower than 1% was not included in the analysis. THESIAS allows the simultaneous estimation of haplotype frequencies and of their effects on the phenotype of interest. It is possible to get the log-likelihood of the data under a specific hypothesis concerning haplotype effects by setting some appropriate constraints on regression parameters. The notation β(h) will refer to the regression parameter characterizing the effect of haplotype h. This option is useful for testing for the equality of haplotype effects and to observe the SNP effects on a haplotype. For example, to test the effect of the second SNP among the four existent haplotypes, we could note two equations β(11) = β(12) and β(21) = β(22). If the difference between global likelihood and likelihood for tested SNP is significant, then the SNP tested had an important role in the haplotype combination. Significance of the model was confirmed through permutation with disease. As the Hardy Weinberg disequilibrium observed in controls could induce errors for haplotype analysis, we tested the robustness of this analysis. As a result, the permutation of status among individuals allowed us to confirm the significance of the result. Secondly, the analysis using the UNPHASED/COCAPHASE program  on the individuals having unambiguous haplotypes, which does not rely on Hardy Weinberg equilibrium hypotheses was carried out and confirmed a strong association.
Cell line and treatment
Rat islet somatostatin-producing RIN-1027-B2 cells were grown in Dulbecco's modified Eagles medium (DMEM) supplemented with 10% fetal bovine serum, penicillin (10 U/ml), streptomycin (10 mg/ml) and incubated at 37°C under a 5% CO2 atmosphere. To analyse CHOP DNA-binding activity in stressed cells, a set of confluent cells was treated for six hours with 2 μg/ml tunicamycin (Sigma), as previously described . Nuclear extracts were prepared as previously described .
Electrophoretic Mobility Shift Assay (EMSA) experiment and Western blot
For EMSA, protein concentrations were determined by the Bio-Rad protein assay with bovine serum albumin as a standard. Double-stranded DNA probes of 23 bp (forward strand : 5'-gctcactgcaaT/Cctctgccctgc-3') containing the -3608T>C polymorphisms were labeled with T4 polynucleotide kinase using [γ32P]ATP, purified with the mini-Quick Spin Columns system (Roche Applied Science, Basel, Switzerland). The labelled probes had a specific activity of ~1 × 106 cpm/pmol DNA. Binding reactions were performed at least three times to replicate results and in the presence of homologous, heterologous and unrelated control competitors They were carried out in a total volume of 20 μl, containing 35000 cpm of radio labeled probe, 10 μg of nuclear extracts, 2 μg of poly (dI-dC) and a buffer with 20 mM potassium phosphate (pH 7.9), 70 mM KCL, 1 mM DTT, 0.3 mM EDTA and 10% glycerol. Gels were exposed to X-ray films (Kodak, Rochester, New York, United States). Quantitation of the label was performed with the NIH Image software http://rsb.info.nih.gov/nih-image/. Nuclear extracts were subjected to SDS-Page, Western blotting and immunolabeling using rabbit anti-CHOP (R-20), rabbit anti-GATA-3 (H-18) or rabbit anti-OCT-1 (C-21) polyclonal antibodies (200 μg/ml) from Santa Cruz Biotechnology, Inc., California, U.S.A. When EMSA was performed with antibodies, the binding mix (listed above) was incubated with antibody for 30 min; where after the radiolabeled probe was incubated for 30 min.