To explore the enrichment of childhood obesity GWAS variants across cell types, we carried out Partitioned Linkage Disequilibrium Score Regression (LDSR) (Finucane et al., 2015) on all ATAC-seq-defined OCRs for each cell type. We assessed cell-type-specific enrichment of GWAS signals in four main categories of genomic regions (Figure 1A): (1) Total OCRs: open chromatin regions defined by ATAC-seq; (2) Promoter OCRs: the subset of OCRs overlapping a gene promoter; (3) cREs: the subset of OCRs that form chromatin loops (as determined by Hi-C/Promoter Capture-C) with a gene promoter, and are therefore considered putative enhancers or suppressors regulating gene expression; (4) cREs ±500bases: extended cREs by 500 bases in both directions. The rationale behind this approach is that different GWAS variants can influence phenotypes by regulating gene expression in a cell-type specific manner through various regulatory mechanisms. For example, they may alter enhancer function (cREs category) or affect the binding of a transcription factor at a gene’s promoter (Promoter OCRs category).

We observed that 41 of 57 cell types – including 22 metabolic, 21 immune, 7 neural cell types, and 7 independent cell lines (Supplementary file 1a) – showed at least a degree of directional enrichment with the total set of OCRs (Figure 1B – Total OCRs). However, only four cell types – two pancreatic alpha and two pancreatic beta cell-based datasets – had statistically significant enrichments (p<0.05). These enrichments were less pronounced when focusing on promoter OCRs only (Figure 1B – Promoter OCRs). To further limit the LD enrichment assessment to just those OCRs that can putatively regulate gene expression via chromatin contacts with gene promoters, we used the putative cREs (Chesi et al., 2019; Pahl et al., 2021). This reduced the number of cell types showing at least nominal enrichment (31 of 57), enlarged the dispersion of enrichment ranges across different cell types, increased the 95% confidence intervals (CI) of enrichments, and hence increased the P-value of the resulting regression score. cREs from pancreatic alpha cells derived from single-cell ATAC-seq were the only dataset that remained statically significant (Figure 1B – putative cREs).

The original reported LDSR method analyzed enrichment in the 500 bp flanking regions of their regulatory categories (Finucane et al., 2015). However, when we expanded our analysis to the ± 500 bp window for our cREs, albeit incorporating more weighted variants into the enrichment (represented by larger dots in Figure 1B – cREs ± 500 bases), this resulted in a decrease in the number of cell types yielding at least nominal enrichment (26 cell types), the enrichment range across cell types, the 95% CI, and level of significance. The pancreatic alpha cell observation also dropped below the bar for significance with this expanded window definition.

Despite the enrichments above only being limited to just a small number of cell types, it is likely that individual loci have differing levels of contributions in various cellular contexts and could not be detected at the genome wide assessment scale. As such we elected to further explore the candidate effector genes that are directly affected by cREs harboring childhood obesity-associated variants by systematically mapping the genomic positions of the LD proxies onto each cell type’s cREs. Most proxies fall within chromatin contact regions (blue area in Venn diagram Figure 2A) or OCRs (yellow area) or open chromatin contact regions (red area), or completely outside (white area) any defined region. Only 94 proxies fall within our defined cREs (overlapped area with dotted green border in Figure 2A), they clustered at 13 original loci (Supplementary file 1b). Figure 2—figure supplement 1 outlines the number of signals at each locus included or excluded based on the criteria we defined for our regions of interest. The TMEM18 locus yielded the most variants through cREs mapping, with 46 proxies for the two lead independent variants, rs7579427 and rs62104180. The second most abundant locus was ADCY3, with 21 proxies for lead variant rs4077678 (Figure 2B). The higher number of variants at one locus did not correlate with implicating more genes or cell types through mapping. The mapping frequency of various variants within a specific locus exhibited substantial differences (Figure 2—figure supplement 2).

Inspecting individual variants regardless of their locus, we found that 28 of 94 proxies appeared in cREs across multiple cell types, with another 66 observed in just one cell type (Figure 2C). 45 variants of these 66 just contacted one gene promoter, such as at the GPR1 and TFAP2B loci (Figure 2—figure supplement 3).

Overall, the number of cell types in which a variant was observed in open chromatin correlated with the number of genes contacted via chromatin loops (Figure 2—figure supplement 4). However, we also observed that some variants found in cREs in multiple cell types were more selective with respect to their candidate effector genes (Figure 2—figure supplement 5 - red arrow), or conversely, more selective across given cell types but implicated multiple genes (Figure 2—figure supplement 5 - blue arrow). Figure 2—figure supplement 6 outlines our observations at the TMEM18 locus – an example locus involved in both scenarios.

Mapping the variants across all the cell types resulted in a total of 111 implicated childhood obesity candidate effector genes (Table 1). Among these, 45 genes were specific to just one cell type (Figure 3—figure supplement 1), including 13 in myotubes and 7 in natural killer cells. Conversely and notably, BDNF appeared across 42 different cell types. Across the metabolic, neural, and immune systems and seven other cell lines, there were nine genes consistently implicated in all four categories (top panel Figure 3 – red stars, Figure 3—figure supplement 2 ‘all’), while five genes were consistently implicated in metabolic, neural, and immune systems (top panel Figure 3 – blue stars, Figure 3—figure supplement 2 ‘all_main’). Two genes, ADCY3 and BDNF, had variants both at their promoters and contacted variants in cREs via chromatin loops (Figure 3—figure supplement 3).

At the TMEM18 locus on chr 2p25.3, a highly significant human obesity locus that has long been associated with both adult and childhood obesity, we observed differing degrees of evidence for 16 genes, but noted that rs6548240, rs35796073, and rs35142762 consistently contacted the SH3YL1, ACP1, and ALKAL2 promoters across multiple cell types (Figure 2C -third and fourth column).

At the chr 2p23 locus, ADCY3 yielded the most contacts (i.e. many proxies contacting the same gene via chromatin loops), suggesting this locus acts as a regulatory hub. However, we observed a similar composition in cell types for four other genes: DNAJC27, DNAJC27-AS1 (both previously implicated in obesity and/or diabetes traits Cherian et al., 2018), AC013267.1, and SNORD14 (RF00016). ITSN2, NCOA1, and EFR3B were three genes within this locus that were only implicated in immune cell types. NCOA1 encodes a prominent meta-inflammation factor (Rollins et al., 2015) known to reduce adipogenesis and shift the energy balance between white and brown fat, and its absence known to induce obesity (Mohsen G et al., 2019).

CALCR was the most frequently implicated gene at its locus, supported by 20 cell types across all systems. While within the BDNF locus, METTL15 and KIF18A – two non-cell-type-specific genes - plus some lncRNA genes, were contacted by childhood obesity-associated proxies within the same multiple cell types as BNDF, again suggesting the presence of a regulatory hub.

At the FAIM2 locus on chr 12q13.12, we observed known genes associated with obesity, eating patterns, and diabetes-related traits, including ASIC1, AQP2, AQP5, AQP6, RACGAP1, and AC025154.2 (AQP5-AS1) along with FAIM2 (Table 1). These genes were harbored within cREs of astrocytes, neural progenitors, hypothalamic neurons, and multiple metabolic cell types. Plasmacytoid and CD1c+conventional dendritic cells were the only two immune cell types that harbored such proxies within their cREs, implicating ASIC1, PRPF40B, RPL35AP28, TMBIM6, and LSM6P2 at the FAIM2 locus.

The independent ADCY9 and FTO loci are both located on chromosome 16. Genes at the ADCY9 locus were only implicated in a subset of immune cell types. Interestingly, genes at the FTO locus were only implicated in Hi-C datasets (as opposed to Capture C), including 6 metabolic cell types and astrocytes. Most genes at the FTO locus were implicated in skeletal myotubes, differentiated osteoblasts, and astrocytes, namely FTO and IRX3; while IRX5, CRNDE, and AC106738.1 were also implicated in adipocytes and hepatocytes.

EndoC-BH1 and myotubes are the two cell types in which we implicated the most effector genes, with 38 and 42, respectively – Figure 3 side panel. This phenomenon is likely proportional in the case of myotubes, given the large number of cREs identified by overlapped Hi-C contact data and ATAC-seq open regions (Figure 1A), but not for EndoC-BH1. Albeit harboring an average number of cREs compared to other cell types, EndoC-BH1 cells were consistently among the top-ranked heritability estimates for the childhood obesity variants resulting from the EGG consortium GWAS (Figure 1) and harbored a significant number of implicated genes by the mapping of proxies. Interestingly, the pancreatic alpha cell type – shown above to be the most significant for heritability estimate by LDSC – revealed only six implicated genes contacted by the defined proxies, namely BDNF and five lncRNA genes.

Of the 111 implicated genes in total, PubMed query revealed functional studies for 66 genes. The remaining were principally lncRNA and miRNA genes with currently undefined functions (Table 1). To investigate how our implicated genes could confer obesity risk, we performed several pathway analyses keeping them either separated for each cell type or pooling into the respective metabolic, neural, or immune system sets. Figure 4—figure supplement 1 shows simple Gene Ontology (GO) biological process terms enrichment results.

Leveraging the availability of our expression data generated via RNA-seq (available for 46 of 57 cell types), we performed pathway analysis. Given that our gene sets from the variant-to-gene process was stringently mapped, the sparse enrichment from normal direct analyses is not ideal for exploring obesity genetic etiology. Thus, we incorporated two methods from the pathfindR package (Ulgen et al., 2019) and our customized SPIA (details in Materials and methods, Figure 4—figure supplement 2). The result of 60 enriched KEGG terms is shown in Figure 4A, Supplementary file 1c, with 13 genes in 14 cell types for pathfindR and 39 enriched KEGG terms shown in Figure 4, Supplementary file 1d, with 10 genes in 42 cell types for customized SPIA. There were 20 overlapping pathways between the two approaches (yellow rows in Supplementary file 1c-d) including many signaling pathways such as the GnRH (hsa04912), cAMP (hsa04024), HIF-1 (hsa04066), Glucagon (hsa04922), Relaxin (hsa04926), Apelin (hsa04371), and Phospholipase D (hsa04072) signaling pathways. They were all driven by one or more of these 5 genes: ADCY3, ADCY9, CREBBP, MMP2, and NCOA1. Interestingly, we observed the involvement of natural killer cells in nearly all the enriched KEGG terms from pathfindR due to the high expression of the two adenylyl cyclase encoded genes, ADCY3 and ADCY9, along with CREBBP. The SPIA approach disregarded the aquaporin genes (given they appear so frequently in so many pathways that involve cellular channels) but highlighted the central role of BDNF which single-handedly drove four signaling pathways: the Ras, Neurotrophin, PI3K-Akt, and MAPK signaling pathways. This also revealed the role of TRAP1 in neurodegeneration.

These two approaches did not discount the role of FAIM2 and CALCR. However, their absence was mainly due to the content of the current KEGG database. On the other hand, these approaches accentuated the role of the MMP2 gene at the FTO locus in skeletal myotubes, given its consistency within the GnRH signaling pathway (Figure 4—figure supplement 3), which is in line with previous studies linking its expression with obesity (Derosa et al., 2008; Aksoyer Sezgin et al., 2022; Nonino et al., 2021).

The GTEx consortium has characterized thousands of eQTLs, albeit in heterogeneous bulk tissues (Pejman, 2017). To assess how many observed gene-SNP pairs agreed with our physical variant-to-gene mapping approach in our multiple separate cellular settings, we performed colocalization analysis using ColocQuiaL (Chen et al., 2022).

282 genes were reported to be associated with the variants within 13 loci from our variant-to-genes analysis. We found 114 colocalizations for ten of our loci that had high conditioned posterior probabilities (cond.PP.H4.abf≥0.8), involving 44 genes and 41 tissues among the eQTLs. We extracted the posterior probabilities for each SNP within each colocalization and selected the 95% credible set as the likely causal variants (complete list in Supplementary file 1e). Despite sensitivity differences and varying cellular settings, when compared with our variant-to-gene mapping results, colocalization analysis yielded consistent identification for 21 pairs of SNP-gene interactions when considering the analyses across all our cell types, composed of 20 SNPs and 7 genes. Details of these SNP-gene pairs are shown in Figure 5A and B.

Of these 20 SNPs, 15 were at the ADCY3 locus, in LD with sentinel variant rs4077678, and all implicated ADCY3 as the effector gene in 29 cell types – 15 metabolic, 6 immune, 4 neural cell types, and 4 independent cell lines (Figure 5C). Indeed, missense mutations have been previously reported for this gene in the context of obesity (Grarup et al., 2018; Stergiakouli et al., 2014) while another member of this gene family, ADCY5, has also been extensively implicated in metabolic traits (Sinnott-Armstrong et al., 2021).

TFs regulate gene expression by binding to DNA motifs at enhancers and silencers, where any disruption by a SNP can potentially cause dysregulation of a target gene. Thus, we used motifbreakR (R package) to predict such possible events at the loci identified by our variant-to-gene mapping. Each variant was predicted to disrupt the binding of several different TFs, thus requiring further literature cross-examination to select the most probable effects. For example, rs7132908 (consistently contacting FAIM2 in 25 cell types) was predicted to disrupt the binding of 12 different transcription factors. Among them, SREBF1 (Figure 6A) was the only TF that concurred with evidence that it regulates AQP2 and FAIM2 at the same enhancer (Kikuchi et al., 2021). The full prediction list can be found in Supplementary file 1f.

To narrow down the list of putative TF binding sites at each variant position, we leveraged the ATAC-seq footprint analysis using the RGT suite (Li et al., 2019). The final set of Motif-Predicted Binding Sites (MPBS) within each cell type ATAC-seq footprints was used to overlap with the genomic locations of the OCRs, and then overlapped with our obesity variants, resulting in annotated 29 variants. Mosaic plot in Figure 6B shows the number and proportions of variants predicted by motifbreakR and/or overlapped with MPBS. Insignificant p-value from Fisher’s exact test indicated the independence of the two analyses. Only seven variants were found within the cREs for the same TF motifs predicted to be disrupted by motifbreakR (Figure 6C). Figure 6—figure supplement 1 outlines the seven variants that motifbreakR and ATAC-seq footprint analysis agreed on the TF bindings they might disrupt.

Our approach of combining RNA-seq, ATAC-seq, and promoter Capture C/Hi-C datasets with GWAS summary statistics offers a systemic view of the multi-cellular nature of childhood obesity, shedding light on potential regulatory regions and effector genes. By leveraging physical properties, such as open chromatin status and chromatin contacts, we enhanced the fine-mapping process and gained new insights into the biological pathways influencing the disease. Although further functional validation is required, our findings provide valuable leads together with their cellular contexts for future research and the development of more effective strategies to address the burden of childhood obesity.

Datasets used in prior studies are listed in Supplementary file 1a. ATAC-seq, RNA-seq, Hi-C, and Capture-C library generation for each cell type is provided in their original published study.

details of each cell type/cell line source were described in their original published study, provided in Supplementary file 1a.

Myotubes (Primary Skeletal Muscle Cells, PCS-950–010), MiaPaCa (CRL-1420), Panc-1 (CRL-1469), microglia (HMC1, CRL-3304), human fetal osteoblastic (hFOB 1.19, CRL-3602), HepG2 hepatocarcinoma (HB-8065) and Colorectal adenocarcinoma ascites derived cells (NCIH716, CCL-251) cell lines were purchased from American Type Cell Center (ATCC, Manassas, Virginia, USA). EndoC-BH1 were purchased from Univercell Biosolutions. Primary Normal Human Astrocytes (NHA) of unknown sex were obtained from Lonza as cryopreserved cells. Human embryonic stem cells (ESC H9, WA09) were obtained from WiCell Research Institute. Melanocytes isolated from foreskin healthy newborn males, were obtained from the Specialized Programs of Research Excellence (SPORE) in Skin Cancer Specimen Resource Core at Yale University. Pancreata from deceased organ donors were obtained by Human Pancreas Analysis Program (HPAP) (https://hpap.pmacs.upenn.edu), a Human Islet Research Network consortium. All were authenticated by the source and confirmed to be mycoplasma negative.

The rest of the cell types were isolated from samples, obtained from consented human donors, maintained and expanded within corresponding laboratories. Cell identities were verified based on observed global expression variation, marker genes by PCR and immunofluorescence microscopy.

This study also includes human adipocytes and pre-adipocytes from P. Seale’s lab, hepatocytes from P. Titchnell’s lab, trabecular meshwork cells from J. O’Brien’s lab, which will be described in detail in separate manuscripts currently in preparation.

The detailed configurations and technical details for each data set are provided in the original studies. In brief, open chromatin regions were called using ENCODE ATAC-seq pipeline as previously described in each study the published data provided. In brief, reads were aligned to the hg19 or hg38 genome using bowtie2; duplicates were removed, alignments from all replicates were pooled, and narrow peaks were called using MACS2. A region was considered open if it overlapped at least 1 bp with ATAC-seq peak. We lifted all coordinates from hg19 to hg38 to ensure consistency between datasets.

in brief, paired-end reads were pre-processed using HICUP pipeline (Wingett et al., 2015) with bowtie2 as aligner and hg19 for reference genome. Significant promoters’ interactions were called using unique read pairs from all baits promoter in the reference by CHICAGO (Cairns et al., 2016) pipeline. In addition to analysis of individual fragments (1frag), we also binned four fragments to improve long-distance sensitivity in interactions calling (Su et al., 2021). Interactions with CHICAGO score >5 in either 1-fragment or 4-fragment resolution were considered significant. These interactions were output as ibed format (similar to BEDPE format) in which each line represents one physical contact between fragments. Interactions from both resolutions were merged and their genomic coordinates were lifted from hg19 to hg38.

We follow the pipeline as a recent study described (Su et al., 2022). Paired-end reads from each replicate were pre-processed using the HICUP pipeline v0.7.4 (Wingett et al., 2015), aligned by bowtie2 with hg38 as the reference genome. The alignments files were parsed to pairtools v0.3.0 to process and pairix v0.3.7 to index and compress, then converted to Hi-C matrix binary format.cool by cooler v0.8.11 at multiple resolutions (500 bp, 1, 2, 4, 10, 40, 500kbp and 1Mbp) and normalized with ICE method (Imakaev et al., 2012). The matrices from different replicates were merged at each resolution using cooler. Mustache v1.0.1 Roayaei et al., 2020 and Fit-Hi-C2 v2.0.7 (Kaul et al., 2020) were used to call significant intra-chromosomal interaction loops from merged replicates matrices at three resolutions 1 kb, 2 kb, and 4 kb, with significance threshold at q-value <0.1 and FDR <1 × 10⁻⁶, respectively. The identified interaction loops were merged between both tools at each resolution. Lastly, interaction loops from all three resolutions were merged with preference for smaller resolution if overlapped.

We intersected ATAC-seq open chromatin regions (OCRs) of each cell type with chromatin conformation capture data determined by Hi-C/Capture-C of the same cell type, and with promoters –1,500/+500 bp of TSS, which were referenced by GENCODE v30.

Data on childhood obesity from the EGG consortium was downloaded from https://www.egg-consortium.org/. We used 8,566,179 European ancestry variants (consisting of 8,613 cases and 12,696 controls in stage I; of 921 cases and 1930 controls in stage II), representing ~55% of the total 15,504,218 variants observed across all ancestries in the original study (Bradfield et al., 2019). The sumstats file was reformatted by munge_sumstats.py to standardize with the weighted variants from HapMap v3 within the LDSC baseline, which reduced the variants to 1,217,311 (7.8% of total).

We used LDSC (http://www.github.com/bulik/ldsc) v.1.0.1 (Finucane et al., 2015) with --h2 flag to estimate SNP-based heritability of childhood obesity within 4 defined sets of input genomic regions: (1) OCRs, (2) OCRs at gene promoters, (3) cREs, and (4) cREs with an expanded window of ± 500 bp. The baseline model LD scores, plink filesets, allele frequencies and variants weights files for the European 1000 genomes project phase 3 in hg38 were downloaded from the provided link (https://alkesgroup.broadinstitute.org/LDSCORE/GRCh38/). The cREs of each cell type were used to create the annotation, which in turn were used to compute annotation-specific LD scores for each cell types cREs set.

19 sentinel signals that achieved genome-wide significance in the trans-ancestral meta-analysis study (Bradfield et al., 2019) were leveraged for our analyses. Proxies for each sentinel SNP were queried using TopLD Huang et al., 2022 and LDlinkR tool (Myers et al., 2020) with the GRCh38 Genome assembly, 1000 Genomes phase 3 v5 variant set, European population, and LD threshold of r² >0.8, which resulted in 771 proxies, including the 21 SNPs from the 99% credible set of the original study (Supplementary file 1g).

The detailed configurations, steps, and technical details for each data set are provided in the original studies. In brief, read fragments from fastq files were mapped to genome assembly hg19 or hg38 using STAR, independently for each replicate and condition. We used GENCODE annotation files for feature annotation and htseq-count for raw read count calculation at each feature. Read counts were transformed into TPM (transcript per million) and normalized internally between replicates/conditions in each individual study. For comparative measurements, we transformed all the expression values into 0–100 scale.

Normalized transcripts per million (TPM) of all measured genes in 46 of 57 cell types was used to perform differential analysis using DEseq2 package (Love et al., 2014), where cell type and system (immune, metabolic, neural and other) were used as variables for the modeling contrast. Because many of the genes we gathered from the variant-to-gene mapping were lowly expressed in corresponding cell types or others, causing relatively high levels of variability, we used apeglm method for effect size (logarithmic fold change estimates) shrinkage Zhu et al., 2019 to alleviate this phenomenon during the genes ranking. Weighted correlation network analysis (WGCNA) package (Langfelder and Horvath, 2008) was used to cluster genes from the variant-to-genes mapping process. WGCNA network construction power was chosen based on the analysis of scale-free topology for soft-thresholding. blockwiseModules with a power of 10 were used to create the correlation network and cluster genes into 3 modules of similarly expressed genes (Figure 4—figure supplement 4). The assigned colors were used to identify the gene modules: turquoise, grey, and blue.

We performed three analyses on the set of genes from the variant-to-genes mapping process:

Gene set over-representation analysis (ORA) was performed using clusterProfiler package Wu et al., 2021 to identify GO biology process terms (org.Hs.eg.db databese) enriched within our genes set in each cell type. A relaxed p-value cutoff was set at 0.1, and the minimum including genes was set at 2 to ensure the capture of all possible enriched terms. An adjusted p-value of 0.05 was later used to filter the significant terms.

We used the pathfindR package Ulgen et al., 2019 to identify active subnetworks in protein-protein interaction networks from Biogrid, KEGG, STRING, GeneMania, and IntAct databases, using the list of genes from the variant-to-genes mapping process. Then we provided the statistic from the differential expression analysis for pathfindR to perform enrichment analyses on the identified subnetworks, discovering enriched KEGG pathways.

The original method, proposed by Tarca in 2009 in SPIA package (Tarca et al., 2009), incorporates ORA with the adjacency matrix to measure the importance of genes within each pathway – genes that are connected to more other genes are likely more important than the downstream end-point genes. This pathway-topology approach measures the actual perturbation on a given pathway under a given condition and given differential effect size. We added more metrics to improve this method: (a) the score of gene impact among networks, (b) the neighborhood of genes that measure the importance of a gene based on its downstream effects, (c) the betweenness puts weight on the genes that act as a gateway for the network flow. The combination of metrics produces two ways of evidence – perturbance and enrichment – for each pathway similar to SPIA. Normal inverse cumulative distribution function was used to combine the p-values of this evidence, then Bonferroni and FDR correction was applied (Figure 4—figure supplement 2).

The summary statistics for the European ancestry subset from the EGG consortium GWAS for childhood obesity was used. Common variants (MAF  ≥ 0.01) from the 1000 Genomes Project v3 samples were used as a reference panel. We used non-overlapped genomic windows of ±250,000 bases extended in both directions from the median genomic position of each of 19 sentinel loci as input. We used ColocQuiaL Chen et al., 2022 to test genome-wide colocalization of all possible variants included in each inputted window against GTEx v.8 eQTLs associations for all 49 tissues available from https://www.gtexportal.org/home/datasets. A conditional posterior probability of colocalization of 0.8 or greater was imposed.

Bam files of mapped reads from replicates and samples were merged for each cell type. The merged bam files were then used for footprinting by RGT-HINT with parameters: --atac-seq, –paired-end, –organism = hg38. If bam files were generated on hg19, we performed lift-over using CrossMap.py bam and hg19ToHg38.over.chain.gz file. We then used RGT-MOTIFANALYSIS matching to scan each footprint for possible transcription binding sites from HOCOMOCO and JASPAR databases for human only with parameter --filter ‘species:sapiens;database:hocomoco,jaspar_vertebrates’. Parameter –rand-proportion 10 was used to generate random putative binding sites with sizes ten times larger than the input footprints. After performing motif matching, we evaluated which transcription factors were more likely to occur in those footprints than in background regions (generated by the previous command) using RGT-MOTIFANALYSIS enrichment with the same filtered databases and default parameters. Output included all the Motif-Predicted Binding Sites (MPBS) that occurred within the identified footprints in each cell type. We overlapped these sites with the loci of our obesity variants.

Genomic positions (0-based coordinates) and allele alternatives of each proxy (from SNPlocs.Hsapiens.dbSNP155.GRCh38 package with matching reference sequence from BSgenome package) were used to scan all position frequency matrix databases (from MotifDb package) for potential transcription factor binding disruptive effects. The motifbreakR function from motifBreakR package was used, with filterp = TRUE and setting a p-value threshold = 0.0005, information content methods method='ic' with even background probabilities of the four nucleotides bkg = c(A=0.25, C=0.25, G=0.25, T=0.25) and BPPARAM = BiocParallel::SerialParam() to allow serial evaluation.