Combining evidence from human genetic and functional screens to identify pathways altering obesity and fat distribution

Summary

Overall adiposity and body fat distribution are heritable traits associated with altered risk of cardiometabolic disease and mortality. Performing rare-variant (minor allele frequency <1%) association testing using exome-sequencing data from 402,375 participants of European ancestry in the UK Biobank for nine overall and tissue-specific fat distribution traits, we identified 19 genes where putatively damaging rare variation associated with at least one trait (Bonferroni-adjusted p < 1.58 × 10⁻⁷) and 50 additional genes at false discovery rate (FDR) ≤1% (p ≤ 4.37 × 10⁻⁵). These 69 genes exhibited significantly higher (one-sided t test p = 3.58 × 10⁻¹⁸) common-variant prioritization scores for association with body mass index (BMI), waist-to-hip ratio adjusted for BMI, and body fat percentage than genes not significantly enriched for rare putatively damaging variation, with evidence of monotonic allelic series (dose-response relationships) among ultra-rare variants (minor allele count ≤10) in 22 genes. Combining rare and common variation evidence, allelic series and longitudinal analysis, we selected 14 genes for CRISPR knockdown in human white adipose tissue cell lines. In two target genes, knockdown significantly (two-sided t test p < 0.05/14) decreased lipid accumulation: PPARG (fold change [FC] = 0.25, p = 5.52 × 10⁻⁷) and SLTM (FC = 0.51, p = 1.91 × 10⁻⁴); knockdown of COL5A3 (FC = 1.72, p = 0.0028) resulted in significantly increased lipid accumulation. Integrating across population-based genetic and in vitro functional evidence, we highlight therapeutic avenues for altering obesity and body fat distribution by modulating lipid accumulation.

Graphical abstract

Overall and tissue-specific fat accumulation are associated with altered risk of cardiometabolic disease and mortality. By combining exome-wide association analysis of traits related to obesity and fat distribution with CRISPR gene perturbation in human fat cells, this study highlights genes linked with fat accumulation, including SLTM, PPARG, and COL5A3.

Introduction

One in four adults globally are either overweight or obese.¹ While higher overall adiposity increases morbidity and mortality,^2,3 disease risk is also informed by the location and distribution of excess fat within particular depots.^4,5 Abnormal distribution of fat is often attributed to lipodystrophy syndromes, which can cause generalized or selective fat mass loss and depot-specific fat growth.⁶ Independent of overall body mass index (BMI), individuals with higher central adiposity have increased risk of cardiometabolic diseases, including type 2 diabetes (T2D) and stroke^7,8; in contrast, individuals with higher hip circumference, an indicator for gluteal adiposity, have lower risk of such outcomes. For example, a standard deviation (SD) increase in hip circumference has been shown to reduce risk of T2D by ∼40%⁹ or myocardial infarction by ∼10%.¹⁰ Previous studies indicate that fat distribution, as assessed by waist-to-hip ratio (WHR), has a strong heritable component independent of BMI, with narrow sense heritability of up to 56% in women and 32% in men.^7,11

BMI-associated genes are enriched in tissues of the central nervous system, notably the hypothalamus, which is involved in appetite regulation.¹² Indeed, blockbuster GLP1-receptor agonists that are prescribed for weight loss^13,14 act primarily through the dorsomedial hypothalamus to control food intake.¹⁵ In contrast, genome-wide association studies (GWASs) for WHR adjusted for BMI (WHRadjBMI) indicate enrichment of genes associated with fat distribution in adipose tissue.¹⁶ However, there are currently few therapeutic avenues to modify both obesity and fat distribution.¹⁷

Understanding the genetic etiology of body fat and fat distribution may drive new therapies for obesity. Mendelian genetic studies have identified rare variants in genes such as PPARG, a master regulator of adipocyte differentiation, and INSR, the insulin receptor, associated with lipodystrophies and extreme forms of central body fat distribution.^18,19 In the general population, rare protein-coding variants with large effects can point to genetic and molecular mechanisms underpinning fat distribution. For example, previous work has demonstrated that rare loss-of-function (LoF) alleles in GPR75 are protective against obesity,²⁰ and protein-truncating variants in INHBE are associated with favorable fat distribution.²¹

Indeed, while there have been several common and rare-variant association studies of obesity,^16,20,22,23 mechanistic studies of the pathways by which the identified genes affect adipogenesis have typically been limited to characterizing just one or two targets.^{24,25,26,27,28,29,30,31,32} The difficulty of manipulating human adipocytes in high-throughput experimental assays has prevented genome-wide functional studies of the genetic drivers of adipogenesis.

Here, we integrated exome sequencing (ES) and common-variant GWAS for nine obesity-related and fat distribution traits to nominate genes associated with overall obesity or central adiposity in up to 402,375 participants in UK Biobank (UKB). We then designed an in vitro functional assay for lipid accumulation in human white adipocytes to functionally characterize a selection of the identified genes using CRISPRCas9 knockdown experiments. Taken together, we demonstrate that converging multi-modal evidence from GWASs, ES, and in vitro knockdown studies in a relevant context can generate putative therapeutic targets for obesity or lipodystrophy.

Material and methods

Imputed data quality control

Sample quality control and assigning population labels

Beginning with the 487,409 individuals with phased and imputed genotype data, we restricted to unrelated individuals with low autosomal missingness rates used for principal components (PCs) by Bycroft et al.³³ We then used genotyping array data, subset to LD-pruned autosomal variants, from these samples to project into the PC space defined by the 1000 Genomes dataset,³⁴ ensuring that we correctly account for shrinkage bias in the projection.³⁵ Next, we used the “super-population” labels (AFR, Africans; AMR, admixed Americans; EAS, East Asians; EUR, Europeans; SAS, South Asians) of the 1000 Genomes dataset to train a random forest (RF) classifier, using the randomForest (4.6) library in R (https://cran.r-project.org/web/packages/randomForest/randomForest.pdf) and predicted the super-population for each of the UKB samples. Samples with classification probability >0.99 for the European super-population were retained for downstream analysis.

Following our ancestry-based filtering regime, we remove samples who withdrew from UKB participation as well as those individuals who were omitted from phasing and imputation. For GWAS, we further subset individuals passing imputed data quality control (QC) to those who also passed ES QC (described below), resulting in a set of 397,315 individuals.

Variant QC

Over 92 million imputed variants across the autosomes and chromosome X are available for analysis. As a starting point for our initial collection of GWAS, we subset to variants with minor allele frequency (MAF) >0.1% in the subset of individuals defined in the sample QC procedure and an INFO score >0.8 from the UKB single nucleotide polymorphism (SNP) manifest file. Following this collection of filtering steps, 16.7 million variants were retained for common-variant GWAS. After performing GWAS, two additional filters were performed, retaining variants with Hardy-Weinberg p > 1 × 10⁻¹⁰, calculated on the whole QCed subset of individuals, and retaining variant-level summary statistics if MAF > 0.1%, calculated for each GWAS. This results in 13,117,850 variants for downstream common-variant analysis.

ES data QC

ES data QC summary

Using the UKB Research Analysis Platform (RAP) we accessed the Genome Aggregation Database (gnomAD) variant call format (VCF) files containing Genome Analysis Toolkit (GATK)-called genotypes for 454,671 individuals. We followed an identical sample QC protocol to Karczewski et al.,³⁶ filtering individuals on covariate-regressed individual-level metrics using median absolute deviations (MADs). We removed variants flagged by the gnomAD QC team for failing one or more of their filters (AC0, RF, MonoAllelic, InbreedingCoeff). Genotype calls were set to missing if they failed filters for genotype quality, depth, and allele balance (described in Karczewski et al.³⁶). We filtered to European ancestry using the super-population ancestry labels assigned with an RF classifier, described above for imputed data QC.

Variant and sample-level QC

We defined “high-quality” variants as those MAF > 0.1% and call rate ≥0.99 falling within the UKB capture intervals plus 50 bp padding. These variants were used to evaluate sample-level metrics of mean call rate and depth and retained samples satisfying all of the following:

• (1)Genetic sex inferred as XX or XY (specifically, genetic sex is defined).
• (2)Mean call rate ≥0.99 among high-quality variants.
• (3)Mean coverage ≥20× among high-quality variants.
• (4)Not withdrawn.

Next, we removed variants satisfying at least one of the following criteria:

• (1)The variant lies outside the UKB capture plus 50 bp padding.
• (2)The variant lies within a low-complexity region.
• (3)The variant lies within a segmental duplication.

In this (sample, variant) set, we ran Hail’s (https://github.com/hail-is/hail) sample_qc() to remove samples lying outside the median ± 4 MADs within each super-population (see above section on imputed data QC). The QC protocol was split by UKB ES tranche (50k, 200k, 450k) to guard against batch effects, as tranches were sequenced in separate runs. The following metrics were used for QC:

• (1)Number of deletions (n_deletion).
• (2)Number of insertions (n_insertion).
• (3)Number of SNPs (n_snp).
• (4)Ratio of insertions to deletions (r_insertion_deletion).
• (5)Ratio of transitions to transversions (r_ti_tv).
• (6)Ratio of heterozygous variants to homozygous alternate variants (r_het_hom_var).

Following MAD filtering (Figure S2; Table S1), 402,375 European-ancestry samples were retained for analysis. For each sample, we excluded non-passing sites as described in Karzcewski et al.³⁶ Briefly, an RF classifier was trained to distinguish true positives from false-positive variants using a collection of allele and site annotations. Variants were assigned “PASS” to maximize sensitivity and specificity across a series of readouts in trio data and precision-recall in two truth samples, after which samples with excess heterozygosity (defined as inbreeding coefficient <−0.3) were removed. Next, we removed low-quality genotypes by filtering to the subset of genotypes with depth ≥10 (5 among haploid calls), genotype quality ≥20, and minor allele balance > 0.2 for all alternate alleles for heterozygous genotypes. Following this filter, we remove variants that were not called as “high quality” among any sample. The resulting high-quality European call set consisted of 402,375 samples and 25,229,669 variants.

Variant consequence annotation

We annotated ES variants using Variant Effect Predictor (VEP) v105 (corresponding to GENCODE v39)³⁷ with the LoF Transcript Effect Estimator (LOFTEE) v1.04 GRCh38³⁸ and dbNSFP³⁹ plugins, annotating variants with Combined Annotation Dependent Depletion (CADD) v1.6,⁴⁰ and Rare Exome Variant Ensemble Learner (REVEL) using dbNSFP4.3⁴¹ and LoF confidence using LOFTEE. We provide code and instructions for this step in our VEP 105 LOFTEE repository (https://github.com/BRaVa-genetics/vep105_loftee), which contains a Docker/Singularity container for reproducibility of annotations. Next, we ran SpliceAI v1.3⁴² using the GENCODE v39 gene annotation file to ensure alignment between VEP and SpliceAI transcript annotations. For variant-specific annotations we use “canonical” transcripts. We separated variants by transcript using bcftools +split-vep and filtered to MANE Select⁴³ protein-coding transcripts. If the gene lacked a MANE Select transcript we selected the canonical transcript defined by GENCODE v39. Using this collection of missense, predicted LoF (pLoF), and splice metrics, and annotations of variant consequence on the canonical transcript, we then determine a set of variant categories for gene-based testing.

Variant consequence categories

• (1)High-confidence pLoF: high-confidence LoF variants, as defined by LOFTEE³⁸ (LOFTEE HC).
• (2)
  Damaging missense/protein altering: at least one of:

  • (a)
    Variant annotated as missense/start-loss/stop-loss/in-frame indel and (REVEL ≥ 0.773 or CADD ≥ 28.1 or both).
  • (b)
    Any variant with SpliceAI delta score (DS) ≥0.2 where SpliceAI DS is the maximum of the set {DS_AG, DS_AL, DS_DG, DS_DL} for each annotated variant, where DS_AG, DS_AL, DS_DG, and DS_DL are delta score (acceptor gain), delta score (acceptor loss), delta score (donor gain), and delta score (donor loss), respectively.
  • (c)
    Low-confidence LoF variants, as defined by LOFTEE (LOFTEE LC).
• (3)Other missense/protein altering: missense/start-loss/stop-loss/in-frame indel not categorized in (2).
• (4)Synonymous: synonymous variants with SpliceAI DS <0.2 in the gene.

REVEL and CADD score cutoffs are chosen to reflect the supporting level for pathogenicity (PP3) from the American College of Medical Genetics and Genomics and the Association for Molecular Pathology (ACMG/AMP) criteria.⁴⁴

Phenotype curation

We used the following nine phenotypes, either directly measured by UKB or derived from UKB phenotypes: BMI (UKB code: 21001, n = 401,059 [sample size after subsetting to 402,375 European-ancestry individuals who pass ES QC]), WHR (derived from waist circumference [UKB code: 48] and hip circumference [UKB code: 49], n = 400,948) with BMI regressed out, body fat percentage (UKB code: 23099, n = 395,147), android tissue fat percentage (UKB code: 23247, n = 39,671), gynoid tissue fat percentage (UKB code: 23264, n = 39,671), ratio of android to gynoid tissue fat percentage (derived from android and gynoid tissue fat percentage phenotypes, n = 39,671), total tissue fat percentage (UKB code: 23281, n = 39,671), visceral adipose tissue volume (UKB code: 22407, n = 21,253), and abdominal fat ratio (UKB code: 22434, n = 20,687).

Genetic association testing

All variant- and gene-level associations were performed in the European-ancestry subset of the UKB using Scalable and Accurate Implementation of Generalized Mixed Model (SAIGE),⁴⁵ a mixed model framework that accounts for sample relatedness.

In SAIGE step 0, we constructed a genetic relatedness matrix (GRM) using the UKB genotyping array data. The genotyped data is linkage disequilibrium (LD) pruned using PLINK (--indep-pairwise 50 5 0.05),⁴⁶ and the sparse GRM is calculated using the createSparseGRM.R function within SAIGE, using 5,000 randomly selected markers, with relatedness cutoff of 0.05. To generate a variance ratio file for subsequent steps in SAIGE, we selected 2,000 variants from the genotyping array data to define a PLINK dataset. For testing common variants in imputed data, we extracted 2,000 variants with minor allele count (MAC) ≥20. For testing the ES data, we extracted two sets of 1,000 variants with 10 ≤ MAC < 20 and MAC ≥ 20, and we combined these sets of markers.

In SAIGE step 1 for each trait, the null model is fit using the curated phenotype data and sparse GRM, with no genetic contribution. Default parameters were used in SAIGE, except --relatednessCutoff 0.05, --useSparseGRMtoFitNULL TRUE, and --isCateVarianceRatio TRUE. In the sex-combined analyses, we account for age, sex, age², age × sex, age² × sex, the first 21 PCs, UKB assessment center, and ES tranche (50,000,⁴⁷ 150,000,⁴⁸ or 250,000⁴⁹ tranches) as fixed effects. For sex-specific analyses, we exclude sex-dependent terms (sex, age × sex, age² × sex) when accounting for fixed effects. All continuous traits were IRNTed using the --invNormalize flag in SAIGE. For SAIGE step 2, we always use the flags --LOCO FALSE and --isfastTest TRUE.

Common-variant association testing in imputed data

We performed single-variant genetic association testing of the UKB imputed data using SAIGE version 1.1.6.3 on the Oxford Biomedical Research Computing cluster. For consistency across analyses, the imputed data were subset to individuals who passed QC in the ES data, our primary dataset for discovery.

Following null model fitting, we carried out variant testing in SAIGE step 2 with default parameters, except for --isFirthbeta TRUE and --pCutoffforFirth = 0.1.

Fine-mapping

Using summary statistics from the common-variant association tests, fine-mapping loci are identified for each trait. Starting with the most significant variant in the pool of genome-wide significant variants, a 1-Mb window centered around the variant is created. All genome-wide variants falling in this window are considered part of the locus defined by the most significant variant and removed from the pool of genome-wide significant variants. Then we proceed to the next most significant variant in the pool of genome-wide significant variants and repeat until the pool of genome-wide significant variants is empty. Loci with overlapping windows are merged.

LD matrices were calculated with LDstore v2.0,⁵⁰ using the same set of individuals used in the GWAS for each trait. Fine-mapping was performed with FINEMAP v1.4,⁵⁰ using shotgun stochastic search and allowing a maximum of 10 causal SNPs per locus.

Rare-variant and gene-level testing in ES data

We carried out rare-variant and gene-level genetic association testing in the European-ancestry subset of the UKB ES data using SAIGE. All analyses involving the ES data were carried out on the UKB RAP using SAIGE version wzhou88/saige:1.1.9.⁴⁵

Following null model fitting, we carried out variant- and gene-level testing in SAIGE step 2 using the variant categories described above, with the --is_single_in_groupTest = TRUE flag. All other parameters were set to default, except –maxMAF_in_groupTest = 0.0001, 0.001, 0.01, although we only report results for maximum MAF of 0.01. We included the following collection of group tests, using the annotations defined above (see section “variant consequence annotation”).

• (1)High-confidence pLoF
• (2)High-confidence pLoF or damaging missense/protein altering

For genes with no damaging missense/protein-altering variants, only the pLoF variant consequence mask was considered when performing Bonferroni multiple-testing correction. Variant-level and gene-burden tests are always two sided.

Obesity and fat distribution trait-gene-level association tests

For each gene and for each obesity and fat distribution trait we tested the optimal sequence kernel association test (SKAT-O) association of rare-variant (MAF < 1%) variation. Exome-wide statistical significance (p < 1.58 × 10⁻⁷) for gene-level tests was defined using Bonferroni correction for 315,996 unique phenotype-gene-consequence mask combinations. To estimate the strength and direction of effect of damaging variation in genes on phenotypes, we also performed gene burden tests, however, the p values from the burden analysis were not used to determine significance.

To expand the set of obesity and fat distribution-associated genes to consider for functional screening, we followed a false discovery rate (FDR)-control process similar to Zhou et al.⁵¹ We selected the gene-level association result which had the lowest SKAT-O p value (across all phenotypes and both variant masks) and applied the Benjamini-Hochberg (BH) procedure to identify 83 significant genes when controlling FDR to 1% (corresponding to a p value of 4.37 × 10⁻⁵). If we had performed the FDR-controlled selection using the full set of results instead of taking the result with the lowest p value, this would be equivalent to an FDR of 12.0%.

We re-ran the gene-level association tests for the 83 FDR-significant genes, conditional on the top fine-mapped common (MAF > 0.1%) variants for fine-mapped loci (described above). For each trait, we identified fine-mapped common variants for each trait that have the highest Bayes factor for being a causal variant (i.e., strongest evidence for causality) in their respective loci. Variants tied for the highest Bayes factor are all selected. There are a total of 915 fine-mapped trait-variant pairs (846 unique variants) on which to condition. We used the --condition flag in SAIGE to condition on the selected common variants when performing gene-level tests for the corresponding phenotype. We perform the gene-level association tests for each chromosome independently and only conditioned on common variants from the same chromosome as the genes being tested.

After conditioning, seven genes (ACVR1C, PRPH, PYGM, SEC16B, TNFRSF6B, TRIM40, and TUBE1) were no longer significant at the predefined SKAT-O p ≤ 4.37 × 10⁻⁵ threshold, leaving 76 genes. Of these 76, seven genes (DEFB112, CHMP4B, FEZF2, GLP1R, PCBD2, TM4SF20, and VGF) were flagged as having very low MAC (MAC < 10), leaving 69 genes to be considered for functional screening.

Sex-difference analyses

We use sex-specific estimated burden effect sizes ($\beta$) and standard errors (SEs) to calculate Z scores corresponding to sex-differential effects using the heterogeneity test described by Martin et al.⁵²:

$$z=\frac{{\beta }_{\text{female}}-{\beta }_{\text{male}}}{\sqrt{S{E}_{\text{female}}^2+S{E}_{\text{male}}^2}.}$$ (Equation 1)

The sex-differential p (p_diff) is estimated under a two-sided hypothesis:

$$p_{\text{diff}}=2\times \mathrm{\Phi }\left(-\left|z\right|\right);\mathrm{\Phi }=CD{F}_{\text{Normal}}\text{.}$$ (Equation 2)

Genes were considered to exhibit significant sex-differential effects if p_diff < 2.67 × 10⁻⁶, Bonferroni adjusted for 18,737 genes tested for sex-differential effects.

Age at diagnosis longitudinal analyses

We curated age at diagnosis of obesity from the UKB linked primary care and hospital record data using mapping tables generated by Kuan et al.,⁵³ resulting in 311,575 individuals for longitudinal analysis. Any codes related to “history of obesity,” for which accurate age at diagnosis could not be extracted, were excluded. We left-truncated observations at the age of the first record (of any code) in either primary care or hospital data, and right-censored at the age of the last record. For each of the 69 FDR-significant genes identified from ES analyses, we performed Cox proportional hazards modeling⁵⁴ to estimate differences in lifetime risk of developing obesity between carriers of pLoF, damaging missense, other missense, synonymous, or non-coding variants (at MAF < 1%) and wild-type individuals (reference group). All effects were adjusted for sex (in sex-combined analyses), the first 10 genetic PCs, birth cohort (in 10-year intervals from 1930 to 1970), and UKB assessment center. Cox modeling was performed using the R package survival v3.3.1.⁵⁵ We visualized age at onset probabilities using Kaplan-Meier survival curves with the R package survminer v0.4.9.⁵⁶

Common-variant gene prioritization scores

We used polygenic priority score (PoPS) calculated by Weeks et al.⁵⁷ for UKB BMI, WHRadjBMI, and body fat percentage. Scores were downloaded from https://www.finucanelab.org/data. There were 34,635 gene-level PoPSs available each for BMI, WHRadjBMI, and body fat percentage, totaling 103,905 gene-trait PoPS.

All of Us gene-level summary statistics

We downloaded gene-level meta-analyzed summary statistics from the All by All browser (https://allbyall.researchallofus.org/), for the two adiposity-related traits which were also available in the All of Us (AoU)⁵⁸ dataset: BMI and body fat percentage. Only cross-ancestry meta-analyzed MAF < 1% pLoF results corresponding to our FDR≤ 1% genes were used. The replication significance threshold (SKAT-O p < 3.62 × 10⁻⁴) was Bonferroni corrected for 138 tests (two traits, 69 genes).

Genebass gene-level summary statistics

Gene-level association test summary statistics were calculated by Karczewski et al.³⁶ and downloaded as a Hail Matrix Table from https://app.genebass.org/downloads. Only MAF < 1% pLoF-associated corresponding to our 69 FDR ≤ 1% genes were used. Associations were significant if Genebass SKAT-O p ≤ 9.98 × 10⁻⁶, controlled for FDR ≤ 1% using the BH procedure.⁵⁹

Genetic correlation estimates

Genetic correlation estimates were accessed from the Neale lab’s UKB genetic correlation browser: https://github.com/astheeggeggs/UKBB_ldsc_r2. These genetic correlations were estimated using LD score regression (LDSC)⁶⁰ v1.0.1, using summary statistics from common (MAF > 0.1%) imputed variant GWAS. Genetic correlation p values less than or equal to 1 × 10⁻³⁰⁸ are reported in Table S5 as zero due to numerical precision limits in Microsoft Excel.

Selecting target genes

To choose target genes for CRISPR-Cas9 knockdown, we consider all MAC ≥ 10 gene association results with FDR < 0.01 which remain significant (p ≤ 4.37 × 10⁻⁵) after conditioning on common variation significantly associated with the trait (see section “Rare-variant and gene-level testing in ES data”). We then compiled multiple lines of evidence for each gene:

• (1)Sufficiently expressed in wild-type hWAT cells (RNA sequencing [RNA-seq] transcript counts >10 at eight and 24 days after differentiation).
• (2)Monotonic allelic series indicating a dose-response relationship with gene dosage and an obesity or fat distribution trait.
• (3)Significant obesity age-of-onset association.
• (4)Published functional work involving gene knockout or knockdown implicates the role of the gene in adiposity.

Using these lines of evidence, we separated the genes into categories: positive controls (genes previously implicated in obesity pathways by functional work in human adipocytes), plausible candidates (genes that have not been implicated in adiposity pathways by functional work in human adipocytes but have suggestive evidence from other functional work or from GWAS/burden association results), and impossible candidates (RNA-seq counts are too low or gene seems to be associated with a pathway that we are not testing). From these categories we chose 14 target genes for knockdown (ABCA1, COL5A3, DENND5B, EXOC7, HERC1, INSR, MFAP5, MLXIPL, PCSK1, PLIN1, PPARG, SLTM, TRIP10, and UBR2).

Wet lab methods

Cell culture and reagents

Human white adipose tissue cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM) (Sigma Aldrich, catalog [Cat] #D6546) with 10% fetal bovine serum (FBS) (Thermo Fisher Cat #A5256801) in a 37°C humidified incubator with 5% CO₂.⁶¹ Passage 40 was used for the experiments.

Cell line generation

For each gene target, four guide RNAs (gRNAs) were commercially synthesized and cloned into the lentiviral vector pLentiCRISPRv2 (GenScript, UK). gRNA sequences are listed in Table S9. Lentivirus for individual gRNAs was produced. Briefly, CRISPR plasmids were cotransfected in HEK293T with packaging vectors pMD2.G (Addgene, 12259) and psPAX2 (Addgene, 12260) using FuGENE HD Transfection Reagent (Promega Cat #E2311). Viral supernatant was collected at 48 h and 72 h post transfection for each gRNA. The lentivirus from the four gRNAs was pooled. hWAT cells were transduced with the pooled virus and with 2 μg/mL polybrene. The selection was performed for 3 days in 1 μg/mL puromycin with medium changes as required.

Differentiation of human white adipose tissue cells

Human white adipose tissue (hWAT) adipocytes were derived from an established preadipocyte immortalized cell line⁶¹ generated from human neck-fat samples. The wild-type hWAT and single-gene knockdown (including the negative-control cell line with the empty Cas9 vector) cells were cultured in DMEM containing 10% FBS. To differentiate into adipocytes via addition of differentiation mixture, the cells were induced with 0.5 mM 3-isobutyl-1-methylxanthine (IBMX, Sigma Cat #5879) and 0.1 μM dexamethasone (Sigma Cat #1756), 33 μM Biotin (Sigma Cat #4639), 17 μM pantothenate (Sigma Cat #5155), 0.5 μM insulin (Sigma Cat #2643), 2 nM 3,3′,5-triiodo-L-thyronine (T3, Sigma, Cat #6397), and 1 μM rosiglitazone (APE Cat #A4304) for 3 days. Next, culture medium was replaced to DMEM supplemented with 10% FBS and 33 μM biotin, 17 μM pantothenate, 0.5 μM insulin, 2 nM 3,3′,5-triiodoL-thyronine, and 1 μM rosiglitazone for 3 days. Medium was subsequently changed every 3 days for the following days. The cells were scrupulously maintained below 70%–80% confluence at all assay stages to preserve their differentiation capacity.

Lipid staining with BODIPY

Differentiated hWAT and single-gene knockdown cells were stained with 2.9 μM boron dipyrromethene (BODIPY) 505/515 (Thermo Fisher Cat #4639) and Hoechst (Hoechst 33342, trihydrochloride, trihydrate, 10 mg/mL solution in water, Thermo Fisher Cat #H3570) for 15 min at 37°C in complete medium, washed with phosphate-buffered saline. Cells were analyzed for fluorescence intensity using an Opera Phenix High-Content Screening System (Revvity, UK) and Harmony high-content analysis software v4.9 (Revvity, UK) provided by the Cellular High Throughput Screening Group. For each knockdown condition, there were six replicates. All cell lines and replicates were processed in the same batch, on the same plate.

RNA-seq of human white adipose tissue cells

Wild-type hWAT and single-gene knockdown cells were seeded in six-well plates and collected after either (1) zero (undifferentiated), 3, 8, or 24 days of differentiation when assessing sufficient expression in wild-type hWAT; or (2) 14 days of differentiation when performing RNA-seq for cell lines involved in the CRISPR knockdown experiment (wild-type hWAT, and single-gene knockdown cells, including Cas9empty cells). Following collection, RNA was isolated using TRIzol (Thermo Fisher Scientific, Carlsbad, CA, USA). Subsequently, RNA samples were purified using the Direct-zol RNA Miniprep protocol as per manufacturer’s (Zymo Research, Cat #R2050) recommendations. Quantification of RNA was performed using the Nanodrop 1000 Spectrophotometer. All RNA samples were sequenced using Illumina 2x150bp paired-end sequencing. Read alignment was performed with STAR⁶² v2.7.10b. Transcript quantification estimation was performed with RSEM⁶³ v1.3.2. Normalized transcript counts were calculated with DESeq2⁶⁴ v1.42.0.

cDNA generation and RT-qPCR

For reverse transcription-quantitative polymerase chain reaction (RT-qPCR), total RNA was extracted for gene expression analysis using Direct-zol RNA Miniprep, (Zymo Research, Cat #R2050). After measurement of the RNA (Nanodrop, Thermo Scientific, USA), the cDNA synthesis was performed using the High-Capacity cDNA Reverse Transcription Kit (Applied Biosystems, Cat #4368814) according to the manufacturer’s instructions. RT-qPCR was carried out with 1 μL of the 1:5 -diluted reverse transcription sample with 10 μL of 2× SYBR Green Master Mix (iQ SYBR Green Supermix, Cat #1708882, BioRad) and 100 nM specific gene primer pairs in a 10-μL total volume in 96-well microtiter plates. The primer sequences for this study are listed in Table S11. LightCycler 96 (Roche, USA) was used to measure CT values of each sample. Each experiment was performed at least in triplicate. β-Actin messenger RNA (mRNA) was employed as an internal standard.

The expression level of each gene was determined by RT-qPCR and normalized against β-actin mRNA level. Knockdown was confirmed by one-sided t test ($\alpha$ = 0.05) using the 2-ΔΔCT method⁶⁵ for amplification performed in separate wells, with the alternative hypothesis that knockdown decreased gene expression.

Differential expression analysis of CRISPR adipocyte RNA-seq

Differential expression analysis was performed using DESeq2⁶⁴ v1.42.0 on RNA-seq data from CRISPR adipocytes. Genes with a raw count less than 10 were removed. The Cas9-empty cell line was used as the baseline condition. Each knockdown or wild-type hWAT was then separately compared to Cas9 empty.

Statistical analysis of lipid accumulation assays

We quantified lipid accumulation using whole-well mean cytoplasm fluorescence of the BODIPY stain, as measured by the Opera Phenix High-Content Screening System. For each knockdown and the wild-type hWAT, we combined replicates with the baseline negative control, Cas9 empty, and regressed readout (mean cytoplasm fluorescence) against knockdown status:

$$\text{readout}\sim \text{is}\_\text{knockdown}+1$$

For wild-type hWAT replicates, is_knockdown is true to separate Cas9 empty and the wild-type replicates. After regression, we calculated Cook’s distance⁶⁶ for each observation. Outliers were removed if Cook’s distance exceeded the median of the F distribution with degrees of freedom n and n − p, where n is the number of observations included in the regression (lipid accumulation n = 12) and p is the number of predictors, including the intercept (p = 2).⁶⁷ Only one observation was flagged as an outlier, for the SLTM knockdown: cytoplasm mean fold change (FC) relative to Cas9 empty = 1.618, Cook’s distance = 0.865. We removed this outlier and re-ran the regression for SLTM knockdown lipid accumulation. We consider the knockdown to have a significant effect on the readout if p < 0.05 for the is_knockdown term from the linear regression. Note that this is the equivalent of a two-sided t test of mean readout between observations from Cas9 empty and a single-gene knockdown or the wild-type hWAT. In Table S13 we report the FC of mean fluorescence of each single-gene knockdown relative to the Cas9 empty cell line and the p value from the two-sided t test, after removing outliers identified using Cook’s distance.

Pathway enrichment in CRISPR RNA-seq data

We used gene set enrichment analysis (GSEA)⁶⁸ v4.3.3 to test for pathway enrichment in CRISPR-Cas9 knockdown adipocyte RNA-seq data. Regularized log-transformed mRNA counts calculated by DESeq2 are provided to GSEA. The pathway enrichment for each knockdown or wild-type hWAT is compared against the baseline control, the empty Cas9 vector, such that pathways with more positive normalized enrichment scores (NESs) are those enriched in Cas9 empty and those with more negative NESs are those enriched in the knockdown or hWAT.

We used pathway sets from Kyoto Encyclopedia of Genes and Genomes (KEGG) Legacy, KEGG Medicus, REACTOME, GO:Biological Process, and GO:Molecular Function, downloaded from GSEA’s Molecular Signatures Database v2023.2.Hs.⁶⁸ We always used 10,000 permutations when running GSEA.

Druggability evidence

Druggability evidence was collated for each knockdown target gene using OpenTargets profiles,⁶⁹ including both known drugs and drug tractability. Here we provide a summary of terms within the “Tractability” columns:

Small molecule; Structure with Ligand: target has been co-crystallized with a small molecule.⁷⁰ HighQuality Ligand: target with ligand(s) (PFI ≤ 7, SMART hits ≤ 2, scaffolds ≥ 2).⁷¹ High-Quality Pocket: target has a DrugEBIlity⁷¹ score of ≥0.7. Med-Quality Pocket: target has a DrugEBIlity score between 0 and 0.7. Druggable Family: target is considered druggable as per the Finan et al.⁷² druggable genome pipeline.

Antibody; UniProt loc high conf: high confidence that the subcellular location of the target is plasma membrane, extracellular region/matrix, or secretion.⁷³ GO CC (Gene Ontology Cellular Component) high conf: high confidence that the subcellular location of the target is either plasma membrane, extracellular region/matrix, or secretion.⁷⁴ UniProt loc med conf: medium confidence that the subcellular location of the target is plasma membrane, extracellular region/matrix, or secretion.⁷³ UniProt SigP or TMHMM (transmembrane hidden Markov model helix prediction)⁷⁵: target has a predicted signal peptide or transmembrane regions and is not destined to organelles (source: Uniprot SigP, TMHMM⁷⁵). GO CC med conf: medium confidence that the subcellular location of the target is plasma membrane, extracellular region/matrix, or secretion.⁷⁴ Human Protein Atlas loc: high confidence that the target is located in the plasma membrane.⁷⁶

PROTAC; Literature: target mentioned in a set of PROTAC-related publications curated by OpenTargets.⁶⁹ UniProt Ubiquitination: target tagged with the UniProt keyword “Ubl conjugation [KW-0832],” which indicates that the protein has a ubiquitination site, based on evidence from the literature.⁷³ Database Ubiquitination: target has reported ubiquitination sites in PhosphoSitePlus,⁷⁷ mUbiSiDa,⁷⁸ or Kim et al.⁷⁹ Half-life Data: target has available half-life data⁸⁰). Small Molecule Binder: target has a reported small-molecule ligand in ChEMBL with a measured activity of at least 10 μM in a target-based assay.⁷¹

Research ethics statement

This study used data from the UK Biobank. The North West Multi-centre Research Ethics Committee (MREC) gave approval to UK Biobank as a Research Tissue Bank (RTB) approval. This approval means that researchers do not require separate ethical clearance and can operate under the RTB approval. This study used a human preadipocyte immortalized clonal cell line derived by Xue et al.,⁶¹ whose study followed the institutional guidelines of and was approved by the Human Studies Institutional Review Boards of Beth Israel Deaconess Medical Center and Joslin Diabetes Center.

Results

Discovery of genes associated with obesity and fat distribution through exome-wide association tests

We performed gene-level discovery testing across 1,827,504 rare variants (MAF < 1%) in 18,788 genes for association with three obesity or fat distribution related traits, i.e., overall obesity (BMI and body fat percentage) and central fat distribution (WHRadjBMI) in up to 402,375 participants of European ancestry in the UKB (Figure S1; Tables S1 and S2). We additionally assessed the effects of these genes on tissue-specific fat-component phenotypes derived from dual-energy X-ray absorptiometry (DXA) and magnetic resonance imaging (MRI) scans (such as android tissue fat percentage, abdominal fat ratio, and visceral adipose tissue volume), in up to 39,671 participants of European ancestry in the UKB (Table S2). We defined damaging rare variants as those annotated to be high-confidence pLoF or damaging missense using a combination of existing effect prediction tools (LOFTEE,³⁸ CADD,⁴⁰ REVEL,⁴¹ and SpliceAI⁴²; material and methods). The reported gene-level effect sizes were estimated by burden testing, while p values were estimated using the SKAT-O for improved power (material and methods). Finally, to account for potential confounding due to nearby common variants, we conditioned gene-level associations on fine-mapped GWAS loci on the same chromosome (material and methods).

We identified 19 unique genes carrying rare damaging variation associated with BMI, body fat percentage, or WHRadjBMI at exome-wide significance (SKAT-O p < 1.58 × 10⁻⁷, Bonferroni adjustment for 315,996 unique tests) (Table 1; Figure 1A). Four genes associated with body fat percentage are also exome-wide significant for BMI with the same direction of effect (PTPRG and GPR151) or WHRadjBMI with opposing directions of effect (PDE3B and PLIN4) (Figures 2 and S3). PLIN4 is also negatively associated with gynoid fat percentage (SKAT-O p = 2.16 × 10⁻⁶), suggesting that the association of PLIN4 with lower-body fat percentage may be driven by reduced gynoid fat deposits, which in turn results in a positive association with WHRadjBMI (an anthropometric proxy for the ratio of android to gynoid fat accumulation). We see a qualitatively similar effect of burden of rare variation in PDE3B, which is associated with higher overall obesity as measured by body fat percentage, but reduced WHRadjBMI (exome-wide significant), android to gynoid ratio, abdominal fat ratio, and visceral adipose tissue volume (all p < 0.05), indicating that the increased body fat percentage in these individuals may be driven by higher subcutaneous rather than visceral fat accumulation (Figures 2 and S3).

Table 1.

Exome-wide significant gene-level associations

Gene (genomic location)	Phenotype	Consequence	Variants in mask	Cumulative MAF (MAC)	SKAT-O p value	Burden p value	Effect size [95% CI]	Effect size, unweighted
ANKRD12 (18:9136228)	WHRadjBMI	pLoF	127	0.00038 (308)	1.71 × 10−9	6.67 × 10−8	0.010 [0.006, 0.013]	0.238
APBA1 (9:69427532)	BMI	pLoF	59	0.00016 (128)	6.25 × 10−8	5.18 × 10−8	0.019 [0.012, 0.026]	0.475
BLTP1 (4:122152331)	BMI	pLoF	282	0.00075 (600)	2.15 × 10−10	9.19 × 10−11	0.011 [0.007, 0.014]	0.258
COL5A3 (19:9959561)	WHRadjBMI	pLoF plus damaging missense	362	0.00388 (3,111)	5.23 × 10−10	2.98 × 10−9	0.003 [0.002, 0.004]	0.078
DIDO1 (20:62877738)	BMI	pLoF	22	0.00003 (25)	1.14 × 10−7	1.14 × 10−7	0.042 [0.026, 0.057]	1.050
GPR151 (5:146513144)	Body fat %	pLoF	37	0.01039 (8,209)	6.60 × 10−8	2.51 × 10−7	−0.002 [−0.003, −0.001]	−0.043
	BMI	pLoF	37	0.01036 (8,313)	1.35 × 10−7	2.02 × 10−7	−0.003 [−0.004, −0.002]	−0.058
INHBE (12:57452323)	WHRadjBMI	pLoF	23	0.00119 (954)	3.27 × 10−9	1.37 × 10−7	−0.005 [−0.007, −0.003]	−0.123
INSR (19:7112255)	WHRadjBMI	pLoF	55	0.00019 (151)	6.38 × 10−8	4.20 × 10−7	−0.012 [−0.017, −0.007]	−0.299
KEAP1 (19:10486125)	WHRadjBMI	pLoF plus damaging missense	88	0.00034 (273)	1.76 × 10−11	2.52 × 10−12	0.012 [0.009, 0.016]	0.305
MC4R (18:60371062)	BMI	pLoF	23	0.00023 (187)	9.30 × 10−13	1.87 × 10−13	0.021 [0.016, 0.027]	0.532
PDE3B (11:14643804)	WHRadjBMI	pLoF	68	0.00132 (1,057)	5.91 × 10−10	7.09 × 10−10	−0.007 [−0.009, −0.005]	−0.166
	body fat %	pLoF	66	0.00132 (1,046)	1.34 × 10−7	2.09 × 10−7	0.005 [0.003, 0.007]	0.102
PKD1 (16:2088708)	WHRadjBMI	pLoF	97	0.00022 (176)	7.68 × 10−11	8.50 × 10−7	0.012 [0.007, 0.017]	0.311
PLD1 (3:171600404)	WHRadjBMI	pLoF plus damaging missense	240	0.01163 (9,324)	4.69 × 10−10	2.88 × 10−9	0.002 [0.001, 0.003]	0.044
PLIN1 (15:89664367)	WHRadjBMI	pLoF	48	0.00104 (836)	8.94 × 10−23	8.98 × 10−21	−0.009 [−0.011, −0.007]	−0.233
PLIN4 (19:4502192)	WHRadjBMI	pLoF	121	0.00444 (3,560)	1.00 × 10−13	3.25 × 10−14	0.004 [0.003, 0.005]	0.091
	body fat %	pLoF	120	0.00444 (3,506)	1.23 × 10−10	2.11 × 10−10	−0.003 [−0.004, −0.002]	−0.080
PPARG (3:12287368)	body fat %	pLoF plus damaging missense	66	0.00021 (168)	2.90 × 10−8	4.34 × 10−7	−0.012 [−0.016, −0.007]	−0.292
PTPRG (3:61561569)	BMI	pLoF	120	0.00024 (194)	1.37 × 10−8	8.31 × 10−8	0.015 [0.01, 0.021]	0.377
	body fat %	pLoF	120	0.00024 (192)	1.35 × 10−7	8.25 × 10−7	0.011 [0.006, 0.015]	0.267
RIF1 (2:151409883)	body fat %	pLoF	66	0.00011 (89)	1.41 × 10−7	1.41 × 10−7	0.017 [0.01, 0.023]	0.416
SLC12A5 (20:46021690)	BMI	pLoF	13	0.00002 (13)	1.21 × 10−7	1.21 × 10−7	0.058 [0.036, 0.079]	1.446

Gene-level exome-wide significance threshold: SKAT-O p < 1.58 × 10⁻⁷. Genomic location indicates the chromosome and base pair coordinates for the start of the gene in Genome Reference Consortium Human Build 38. Variants in the mask are those that are included in the SKAT-O test for a given gene. They include all variants in the gene with MAF < 1% and consequence annotation matching the specific mask. Total MAF is the sum of MAF for all variants in the mask. In parentheses, MAC is the total MAC for the gene among individuals included in the association test. The effect size of a gene is estimated using burden testing and is in units of the phenotype’s SD. The 95% CI is defined by an interval centered on the effect size point estimate, ±1.96 SEs. The unweighted effect size is a version of the burden effect size that does not weight variants by Beta (MAF; 1, 25), the default weighting used by SAIGE. Using the unweighted effect sizes puts the gene-level effect on the same scale as variant-level effects. See Note S1 for calculation of unweighted effect size. WHRadjBMI, waist-to-hip ratio adjusted for BMI; BMI, body mass index; pLoF, predicted loss of function.

Figure 1.

Consequences of rare missense variation on obesity or fat distribution related traits in UKB

(A) Gene- and variant-level effects as a function of aggregated minor allele frequency. Only gene-level results that are exome-wide or FDR ≤ 1% significant are shown. Only fine-mapped common (MAF > 1%) variants with posterior inclusion probability ≥0.9 are shown. Effect sizes for BMI and body fat percentage are converted to the kg/m² and fat percentage scale, respectively, by multiplying the effect size in SD units by the empirical SD of each trait (4.75 kg/m² for BMI, 8.51% for body fat percentage). Genes selected for knockdown are outlined in black and labeled. Genes with highest and lowest effect size for each trait are also labeled. Point sizes are scaled by effect size.

(B) Allelic series for exome-wide significant genes with monotonic relationships between effect size and consequence severity in ultra-rare (MAC ≤ 10) variant burden tests. 95% CIs for effect size were defined as ±1.96 SEs. For allelic series of all FDR ≤ 1% genes, see Figure S4.

(C) Enrichment of PoPS⁵⁷ among exome-wide and FDR ≤ 1% significant genes compared to other genes. Significance of one-sided t test with alternate hypothesis that a significant gene category has higher average PoPS than the non-significant gene category: ^∗∗∗p < 0.001, ^∗∗p < 0.01, ^∗p < 0.05. Results of t test for BMI PoPS: exome-wide significant gene PoPS vs. non-significant gene PoPS t-statistic (P) = 2.79 (2.65 × 10⁻³), FDR ≤ 1% vs. non-significant = 3.17 (7.64 × 10⁻⁴). Results of t test for WHRadjBMI PoPS: exome-wide significant vs. non-significant = 4.96 (3.60 × 10⁻⁷), FDR ≤ 1% vs. non-significant = 5.19 (1.06 × 10⁻⁷). Results of t test for body fat percentage PoPS: exome-wide significant vs. non-significant = 2.04 (2.07 × 10⁻²), FDR≤ 1% vs. non-significant = 2.46 (7.02 × 10⁻³). The mean PoPS for each gene category is indicated by black bars. For BMI: exome-wide significant genes PoPS mean [95% CI from ±1.96 SEM] (number of genes) = 0.324 [0.086, 0.562] (7), FDR ≤ 1% = 0.250 [0.070, 0.430] (15), non-significant = −0.009 [-0.012, −0.006] (34,613). For WHRadjBMI: exome-wide significant = 0.430 [0.161, 0.698] (10), FDR ≤ 1% = 0.353 [0.207, 0.500] (16), non-significant = −0.014 [−0.017, −0.011] (34,609). For body fat percentage: exome-wide significant = 0.242 [−0.060, 0.544] (6), FDR ≤ 1% = 0.226 [0.023, 0.429] (10), non-significant = −0.006 [−0.009, −0.003] (34,619).

Figure 2.

Gene burden effects across all nine obesity and fat distribution traits for 19 genes with exome-wide significant burden associations

Heatmap is colored by the burden test t-statistic. Only the result of the consequence mask with the lowest SKAT-O P of that gene is shown for each trait-gene pair. Significance of SKAT-O test: ^∗∗∗exome-wide significant (p < 1.58 × 10⁻⁷), ^∗∗FDR ≤ 1% significant (p ≤ 4.37 × 10⁻⁵), ^∗nominal significance (p < 0.05). Gray squares indicate that the gene-level test could not be performed due to insufficient allele counts. WHRadj, waist-to-hip ratio adjusted for BMI; BF%, body fat percentage; And%, android fat percentage; Gyn%, gynoid fat percentage; A/G, android-gynoid fat percentage ratio; Tot%, total fat percentage; VAT, visceral adipose tissue; AbdR, abdominal fat ratio.

While the 19 exome-wide significant genes were not significantly associated (all p > 4.38 × 10⁻⁴, Bonferroni adjustment for 19 unique genes × 6 phenotypes) with any other tissue-specific fat components, we observed that nominally significant associations (p < 0.05) generally shared the expected direction of effect on the primary and supplemental phenotypes (Figures 2 and S3). Genes that are positively associated with BMI (MC4R) and body fat percentage (RIF1) are also positively associated with android fat percentage, gynoid fat percentage, android/gynoid ratio, and total tissue fat percentage; PPARG is negatively associated with body fat percentage as well as gynoid and total tissue fat percentage (Figures 2 and S3). Similarly, COL5A3 is associated with increased WHRadjBMI and higher android to gynoid ratio, while PLIN1 is associated with lower WHRadjBMI and higher gynoid fat percentage as expected.

Finally, we found evidence of monotonic allelic series, that is, increasingly large effects of increasingly damaging variants in a dose-response relationship, among ultra-rare variants (MAC ≤ 10) in COL5A3, DIDO1, INSR, PLIN1, PTPRG, PPARG, and SLC12A5 (Figure 1B).

Convergence of rare- and common-variant evidence for genes associated with obesity and fat distribution

Genes associated with obesity and fat distribution in rare-variant gene-level tests overlapped with those highlighted by common-variant associations, demonstrated by a significant enrichment of PoPS⁵⁷ (one-sided t test p < 0.05; Figure 1C) among genes harboring exome-wide-significant damaging rare-variant association signals. Moreover, 50 rare-variant gene-level associations that do not reach exome-wide significance but pass a less stringent significance threshold (SKAT-O p ≤ 4.37 × 10⁻⁵, using the BH procedure⁵⁹ for FDR ≤ 1% on minimum SKAT-O p value across traits and two variant masks; material and methods) also have significantly higher PoPS compared to non-significant genes (one-sided t test, p < 0.01; Figure 1C), indicating value in examining these additional genes as potential therapeutic targets for obesity and lipodystrophy (Table S3). Seven additional gene-level associations also passed the FDR ≤ 1% significance threshold but were supported by MACs less than 10 (Note S2). Combining PoPS across all gene-trait pairs (for traits with PoPS available: BMI, WHRadjBMI, body fat percentage) which are exome-wide or FDR ≤ 1% significant and comparing to PoPS of non-significant gene-trait pairs, we observe strong enrichment of scores (one-sided t test: t = 8.47, p = 1.26 × 10⁻¹⁷, mean [95% confidence interval (CI) using ±1.96 × SEM] PoPS of 64 exome-wide or FDR ≤ 1% significant gene-trait pairs = 0.307 [0.223, 0.392], mean of 103,841 non-significant gene-trait pairs = −0.010 [−0.012, −0.008]), demonstrating again the convergence of rare and common-variant evidence when FDR ≤ 1% significant genes are included.

Of the 31 genes linked with severe monogenic obesity or lipodystrophy,^6,81 three (PPARG, PLIN1, and MC4R) are associated with an obesity or fat distribution trait at the exome-wide level (p < 1.58 × 10⁻⁷) in our rare-variant gene-level association testing, representing a 125-fold enrichment over random chance (Fisher’s exact test p = 3.87 × 10⁻⁶). Of the 69 FDR-significant genes, 23 have previously been reported to have rare-variant gene-level associations with traits related to obesity or fat distribution in UKB participants, across a variety of variant annotation and gene-level association testing methods (Table S8).^20,21,23,82 We looked up the effects of pLoF variant burden in FDR-significant genes with MAC ≥ 10 in AoU⁵⁸ results meta-analyzed across ancestries for available phenotypes: BMI and WHRadjBMI. Accounting for the multiple-testing burden of 69 genes and two phenotypes (Bonferroni adjustment, p < 3.62 × 10⁻⁴), we replicated, with matching effect directions, the associations of SLTM, MC4R, and KIF1B with BMI, and PKD1 and PLIN4 with WHRadjBMI (Table S4). We also found that HECTD4 is positively associated with body fat percentage in UKB and BMI in AoU; an additional 18 of our FDR-significant genes are nominally associated (p < 0.05) with an obesity-related trait in AoU (Table S4).

Lastly, to evaluate the phenome-wide effects of genes related to obesity and fat distribution, we scanned 4,529 phenotypes in Genebass³⁶ summary statistics for associations with the 69 FDR-significant genes. We found 549 significant associations (Genebass SKAT-O p ≤ 9.98 × 10⁻⁶, FDR ≤ 1%) across 211 traits for 41 out of 69 genes (Figure S5; Table S5). The most significant Genebass associations were among blood biochemistry lipid phenotypes (Figure S5A; Table S5). We also observed shared associations across FDR-significant genes for phenotypes significantly genetically correlated to BMI and body fat percentage, such as fat mass and distribution traits. For example, eight FDR-significant genes (BLTP1, DIDO1, GPR151, MC4R, PDE3B, PLIN4, RIF1, and UBR2) are associated with arm fat mass, leg fat mass, arm fat percentage, whole-body fat mass, and hip circumference (all traits have common-variant LDSC⁸³ r_g > 0.80, p ≤ 1 × 10⁻³⁰⁸ with BMI and body fat percentage; Table S5). We see some evidence for pleiotropy, as defined by associations with phenotypes genetically uncorrelated with BMI or body fat percentage, among results with common-variant genetic correlations available: PDE3B and STAB1 are associated with platelet distribution width (BMI r_g = 0.025, p = 0.274; body fat percentage r_g = 9.7 × 10⁻⁴, p = 0.958) and ANKRD12 is associated with neutrophil percentage (BMI r_g = −0.029, p =.117; body fat percentage r_g = −8.8 × 10⁻³, p = 0.639) (Table S5).

Sex- and age-specific effects of genes associated with obesity and fat distribution

We performed all association tests in sex-specific strata, identifying three genes (INSR, PDE3B, and PLIN4) with significant differences (sex-difference two-sided heterogeneity test of effect sizes P_diff < 2.67 × 10⁻⁶, Bonferroni adjusted for 18,737 genes tested for sex-differential effects; material and methods) in burden effect sizes. We observed opposing sex-specific effects for INSR on WHRadjBMI: burden $\beta$_female (SE) = −0.00458 (9.44 × 10⁻⁴), $\beta$_male = 0.00195 (0.00100), P_diff = 1.66 × 10⁻⁶. We found that female-specific effects may be driving the reported sex-combined associations between WHRadjBMI and PDE3B: $\beta$_female = −0.0169 (0.00169), $\beta$_male = −0.00288 (0.00179), sex difference p = 8.02 × 10⁻⁹) and PLIN4 ($\beta$_female = 0.00822 (9.15 × 10⁻⁴), $\beta$_male = 0.00108 (9.86 × 10⁻⁴), P_diff = 8.19 × 10⁻⁸ (Figure S6A; Table S6). Within the sex-specific analysis, we find two genes with female-specific associations (p < 2.67 × 10⁻⁶, Bonferroni adjusted for 18,737 genes tested for sex-specific effects) of pLoF variation on obesity and fat distribution (CASQ1 with body fat percentage, female SKAT-O p = 1.12 × 10⁻⁶; PLXND1 with WHRadjBMI, female p = 9.69 × 10⁻¹⁰), and two genes with male-specific associations, both for the ratio of android to gynoid tissue fat percentage (ZNF841 male p = 1.04 × 10⁻⁶; TINF2 male p = 4.84 × 10⁻⁷) (Figure S6B; Table S6), which do not reach significance at the FDR ≤ 1% threshold (p ≤ 4.37 × 10⁻⁵) in sex-combined analyses.

We leveraged the age at diagnosis of obesity in longitudinal health records linked to UKB participants to identify that a burden of rare missense variation (MAF < 1%) in five genes was associated (p < 2.42 × 10⁻⁴, Bonferroni adjusted for 207 gene-consequence mask pairs) with elevated lifetime risk of obesity (Figure S7; Table S7). Individuals carrying rare missense variants in MC4R (Cox proportional hazard ratio [HR] [SE] = 1.46 [0.0843], p = 7.75 × 10⁻⁶), men with pLoF variants in SLTM (HR = 5.37 [0.354], p = 2.05 × 10−6) and damaging missense variants in PCSK1 (HR = 1.86 [0.165], p = 1.63 × 10⁻⁴), and women with pLoF variants in DIDO1 (HR = 11.2 [0.578], p = 2.85 × 10⁻⁵) and SLC12A5 (HR = 14.8 [0.708], p = 1.45 × 10⁻⁴) were at risk of earlier age at onset of obesity (Figure S7; Table S7).

Nominating target genes for in vitro functional characterization through CRISPR knockdown

As adipose tissue is enriched for the expression of fat distribution genes identified from common-variant GWAS,¹⁶ we assessed the mRNA counts of overall and tissue-specific fat distribution-associated genes from our exome-wide gene-level association testing in differentiated hWAT derived pre-adipocytes in vitro (Figure 3A). Genes with FDR-significant results in gene-level tests had higher expression in hWAT 8 and 24 days after differentiation compared to non-significant genes (one-sided t test p = 1.55 × 10⁻⁴³, comparing mRNA normalized counts; Figure 3A). From the 56 of 69 FDR-significant genes sufficiently expressed in differentiated hWAT (mean normalized mRNA count > 10 at 8 and 24 days after differentiation), we selected for functional characterization 14 genes (ABCA1, COL5A3, DENND5B, EXOC7, HERC1, INSR, MFAP5, MLXIPL, PCSK1, PLIN1, PPARG, SLTM, TRIP10, and UBR2; Figures 3B and S8), which (1) demonstrated monotonic allelic series indicating a dose-response relationship with gene dosage (ABCA1, COL5A3, DENND5B, INSR, PCSK1, PLIN1, and PPARG) (Figure S4), (2) associated with obesity age-of-onset (SLTM and PCSK1) (Figure S7; Table S7), or (3) are implicated in lipid or glucose metabolism pathways by mouse whole-body or adipocyte-specific gene perturbation (Table S8).

Figure 3.

CRISPR knockdown strategy

(A) Mean normalized mRNA counts across four time points (days after differentiation: 0, 3, 8, 24) in human white adipocyte tissue for 69 genes passing FDR ≤ 1% significance (SKAT-O p ≤ 4.37 × 10⁻⁵) versus all non-significant genes (n = 60,605). Error bars indicate 95% CIs as defined by ±1.96 × SEM normalized mRNA count.

(B) Selection strategy for genes to include in conditional knockdown experiment. Numbers in parentheses indicate the number of genes in a given group.

(C) Experimental design for CRISPR knockdown, adipocyte cell culture, and imaging.

(D) Confirmation of knockdown efficacy using relative mRNA expression in differentiated adipocytes. Error bars indicate 95% CIs defined by ±1.96 × SEM. One-sided Wald test comparing mRNA count in knockdown cell line relative to count in Cas9 empty, with the alternative hypothesis that the count is lower in the knockdown: ^∗∗∗p < 0.001, ^∗∗p < 0.01, ^∗p < 0.05.

(E) Confirmation of knockdown efficacy using relative RNA expression quantified with RT-qPCR in undifferentiated adipocytes.

Error bars indicate 95% CIs defined by ±1.96 × SEM. One-sided t test comparing mRNA count in knockdown cell line relative to count in Cas9 empty, with the alternative hypothesis that the count is lower in the knockdown: ^∗∗∗p < 0.001, ^∗∗p < 0.01, ^∗p < 0.05. Knockdown confirmation for PPARG was performed by checking two isoforms, PPARG1 and PPARG2. ^†PPARG knockdown confirmation was performed in differentiated adipocytes because expression of PPARG isoforms was too low in undifferentiated adipocytes.

We selected four gRNAs per gene target to generate 14 knockdown hWAT cell lines (material and methods; Figure 3C; Table S9). We additionally generated a negative-control cell line with Cas9 empty vector. We confirmed gene knockdown through a two-step process: (1) RNA-seq expression of the target gene in knockdown cell lines relative to expression in Cas9-empty cell line and observed significantly decreased (p < 0.05, one-sided Wald test) expression for COL5A3, DENND5B, EXOC7, HERC1, INSR, MFAP5, MLXIPL, SLTM, TRIP10, and UBR2 (Figure 3D; Table S10), and (2) RT-qPCR measurements of RNA for genes whose knockdown could not be confirmed via RNA-seq and observed significantly decreased (p < 0.05, one-sided t test) for ABCA1 and PPARG (Figure 3E; Table S12). Expression of PCSK1 quantified by RT-qPCR was not significantly decreased compared to the control (RT-qPCR FC [95% CI] = 0.624 [0.373, 1.045], one-sided t test p = 0.0665) but was suggestive of a knockdown. Another gene knockdown that could not be confirmed was PLIN1 (RT-qPCR FC [95% CI] = 0.933 [0.390, 2.234], one-sided t test p = 0.450). Therefore, we excluded PCSK1 and PLIN1 from consideration in downstream analyses.

Knockdown of obesity- and fat distribution-associated genes alters in vitro adipogenesis and lipid accumulation

Abnormal adipogenesis and lipid accumulation are hallmarks of adipose tissue dysfunction.⁸⁴ As the process of adipocyte differentiation is considered a necessary precursor to lipid accumulation,⁸⁵ we use BODIPY staining⁸⁶ (material and methods) to measure both lipid accumulation and adipogenesis more generally. The BODIPY assay was conducted for six replicates of each knockdown cell line.

We observed significantly reduced (p < 3.57 × 10⁻³ in two-sided t test of mean FC in fluorescence between knockdown and control Cas9-empty cell line, Bonferroni adjusted for 14 knockdown cell lines) lipid accumulation in the PPARG knockdown cell line (FC = 0.245, p = 5.52 × 10⁻⁷) (Figure 4; Table S13). Inactivation of PPARG through knockout has previously been shown to lower lipid accumulation in human and mouse pre-adipocytes,^25,87,88 providing validation of the assay used here.

Figure 4.

Effect of CRISPR single-gene knockdown on lipid accumulation in human adipocytes

(A) Effect of single-gene knockdown on lipid accumulation measured by whole-well mean cytoplasm intensity of BODIPY dye fluorescence. Points represent measurements from separately cultured cells (biological replicates) from the same knockdown cell line. Each knockdown cell line has six replicates, except SLTM, which had one outlying replicate removed (material and methods).

(B) Differentiated human adipocyte knockdown cell lines stained with BODIPY (green) and Hoechst (blue), imaged at 10× magnification. The elements of the boxplots are median (horizontal line in box), upper (Q3) and lower (Q1) quartiles marking the ends of the box, with whiskers extending to the most extreme points within the range [Q1 − IQR, Q3 + IQR], where IQR = Q3 − Q1. All points are shown.

Two-sided t test of mean FC relative to Cas9-empty cells: ^∗∗p < 0.05/14, ^∗p < 0.05. Summary statistics are available in Table S13. hWAT, human white adipose tissue.

We also noted significantly reduced lipid accumulation in knockdown of SLTM (FC = 0.514, p = 1.91 × 10⁻⁴) (Figure 4A; Table S13). SLTM encodes the SAFB-like transcription modulator, which regulates the GLI family of transcription factors in mice.⁸⁹ These GLI transcription factors in turn control expression of lipid metabolic genes, including critical adipogenesis transcription factors PPARG and C/EBP (α, β, γ, and δ).⁹⁰ Notably, the SLTM pLoF burden association was replicated in a multiancestry meta-analysis conducted in AoU (https://allbyall.researchallofus.org/) as the top association for BMI (AoU meta-analyzed p = 2.75 × 10⁻⁷). While this suggests a potential therapeutic target for altering fat deposition, more work is needed to understand why individuals carrying rare pLoF variants in SLTM have higher BMI (burden $\beta$ [SE] = 0.0191 [0.00455], SKAT-O p = 2.65 × 10⁻⁶), and the target tractability of SLTM, which may be amenable to small-molecule binding (Table S14).

We observed significantly increased lipid accumulation relative to Cas9-empty cells in knockdown of COL5A3(FC = 1.72, p = 0.0028) and nominally significant increased lipid accumulation in knockdowns of EXOC7 (FC = 1.35, p = 0.0096) and TRIP10 (FC = 1.39, p = 0.0157) (Figure 4A; Table S13). The products of these genes are all variously involved in lipid uptake, synthesis, and adipogenesis. EXOC7 is a component of the exocyst complex, which regulates the uptake of free fatty acids by adipocytes.⁹¹ TRIP10 is an interactor of the thyroid hormone receptor (TR-β1),⁹² which regulates de novo fatty acid synthesis.⁹³ TR-β1 has been proposed as a target to treat dyslipidaemia,⁹⁴ and the TR-β1 agonist resmetirom is currently in phase III clinical trials for treating non-alcoholic fatty liver disease.⁹⁵ However, there are no known drugs directly targeting TRIP10 (Table S14). We also observed increased lipid accumulation in knockdowns of COL5A3 (FC = 1.72, p = 2.78 × 10⁻³) (Figure 4A; Table S13)—as a ubiquitous component of the extracellular matrix, type V collagen informs the proper differentiation and development of adipocytes.⁹⁶

Transcriptome-wide effects of gene knockdowns

Obesity-gene knockdowns had wide-ranging effects on the hWAT transcriptome, with between eight and 1,904 genes differentially expressed (|log₂(FC)| > 1, FDR-adjusted p < 0.05) in knockdown cell lines as compared to Cas9-empty negative controls (Figure 5; Table S15). We examined the molecular effects of these knockdowns through pathway enrichment analyses for pathways in the KEGG,⁹⁷ REACTOME,⁹⁸ and GO:Biological Process and GO:Molecular Function⁹⁹ databases.

Figure 5.

Effect of gene knockdown on differential gene expression in human adipocytes

(A) Gene differential expression results from RNA-seq data performed on hWAT wild-type and single-gene knockdowns. Gene differential expression analysis was performed with DESeq2,⁶⁴ comparing each cell line (wild-type hWAT or knockdown) to the Cas9-empty cell line. All significance thresholds use p_adj, a p value corrected for multiple testing calculated in DESeq2. Blue points indicate genes that are significantly downregulated (p_adj < 0.05 and log₂FC < 1). Red points indicate genes that are significantly upregulated (p_adj < 0.05 and log₂FC > 1). Gray points are all other genes. Genes are labeled if they are in the top five most significant (lowest p_adj) red or top five most significant blue points, in the top five most significant regardless of point color, or have |log₂FC| > 10. The horizontal and vertical dashed lines indicate the significance cutoffs of p_adj < 0.05 and |log₂FC| > 1, respectively.

(B) Counts of significantly downregulated (blue, p_adj < 0.05 and log₂FC < −1) and upregulated (red, p_adj < 0.05 and log₂FC > 1) genes for each single-gene knockdown and hWAT. The y axis is sorted by the total number of differentially expressed genes for each knockdown condition.

Knockdown of PPARG, a master regulator of adipogenesis, significantly alters 95 pathways (family-wise error rate [FWER] p < 0.05), including (1) transcriptional regulation through PRC2-driven DNA methylation (number of genes [n] = 59, p < 0.0001), histone acetylation (n = 138, p < 0.0001), and small RNAs (n = 102, p = 0.0001); (2) protein synthesis by modulating rRNA expression through SIRT1 (n = 63, p < 0.0001), ERCC6 (CSB), and EHMT2 (G9a) (n = 71, p < 0.0001) and the B-WICH chromatin remodeling complex (n = 86, p = 0.0001); and (3) functions critical to cell division such as assembly of the origin recognition complex at the origin of replication (n = 64, p < 0.0001), condensation of prophase chromosomes (n = 69, p < 0.0001), and cell-cycle checkpoints (n = 286, p < 0.0001) (Table S16). As expected, PPARG knockdown also alters pathways important for adipogenesis, such as the transcription of androgen-receptor regulated genes (n = 62, p < 0.0001), RUNX1 regulation¹⁰⁰ (n = 89, p = 0.0001), formation of the beta-catenin T cell factor trans-activating complex¹⁰¹ (n = 87, p = 0.0001), and pre-NOTCH expression and processing¹⁰² (n = 104, p = 0.0003) (Table S16).

We observed enrichment of several mitotic, transcription, and translation pathways in knockdowns of lipid accumulation-altering genes EXOC7 and TRIP10 (Table S16) but did not find mechanisms directly implicated in lipid accumulation. On the other hand, while MLXIPL knockdown did not affect lipid accumulation in our assays, we nevertheless observed enrichment (FWER p < 0.05) of pathways involved in adipogenesis, such as interferon α/β signaling¹⁰³ (n = 66, p = 0.0016) and lipopolysaccharide binding¹⁰⁴ (n = 32, p = 0.0284). MLXIPL encodes the carbohydrate-responsive element-binding protein (ChREBP), which is thought to regulate gene expression in response to glucose and reduces adipose tissue mass in mouse knockout models.¹⁰⁵ While this may be a potential therapeutic target for fat distribution, mice deficient in MLXIPL/ChREBP do not tolerate fructose in their diet and develop severe diarrhea and irritable bowel syndrome.¹⁰⁶

Discussion

Through a series of population-level genetic and in vitro functional genomics investigations, we demonstrate a model to map the biological mechanisms of obesity and body fat distribution, from variant to molecular function to systemic phenotype. We utilized complementary rare-variant gene-based analyses and common-variant prioritization to nominate 69 genes with robust associations to BMI, WHRadjBMI, body fat percentage, and six tissue-derived fat components in up to 402,375 European-ancestry participants in UKB. By evaluating the sex-specific and phenome-wide associations of these genes, we built a comprehensive understanding of their systemic effects. Combining multiple lines of genomic, transcriptomic, and prior functional evidence, we selected 14 genes for functional characterization by CRISPR knockdown in hWAT cell lines. Our systematic, multi-gene approach to functional characterization represents a step forward in comprehensive evaluation of adipogenesis-associated genes, given the difficulty of manipulating adipocytes in high-throughput experiments. We observed significantly reduced lipid accumulation in PPARG and SLTM knockdown, significantly increased lipid accumulation in COL5A3 knockdown, and nominally significant increased lipid accumulation in knockdowns of EXOC7 and TRIP10. While some of these genes have previously been identified through large-scale genetic analyses, we show their functional effects on adipogenesis and highlight potential therapeutic avenues to regulate body fat mass and/or distribution. Taken together, our population-based and in vitro genetics investigations highlight molecular mechanisms of, and therapeutic avenues to alter, fat distribution.

Large-scale biobanks with genetic and phenotypic data power discovery of genetic loci relevant to disease. Among biobanks, UKB has led the way in many aspects, but is not without limitations in the context of our work. Our study focused on individuals of European ancestry, who make up the majority of UKB. Future studies should consider individuals from other ancestry groups, as replication in these groups would improve the generalizability of findings. As diverse biobanks increase in size, power to discover associations in non-European ancestries will also increase.

While there was high sample coverage for the three main traits of study (BMI, WHRadjBMI, and body fat percentage; n > 395,000), the sample sizes of traits derived from DXA (n = 39, 671) and MRI scans (n > 20,000) were substantially smaller. These DXA/MRI scan traits were included in our analysis as they can provide depot-specific granularity to the nature of adipogenesis; however, future analysis would benefit from increased sample sizes.

The recent growth in sample sizes of exome-sequenced participants in biobanks has accelerated the development of gene-level analyses.^107,108 By annotating rare variants with their putative functional consequences and then collapsing these across each gene, gene-level testing has enabled the discovery of novel genes associated with complex human traits.^36,109 While gene-level testing is a powerful strategy for gene discovery, it remains limited by the requirement of high-quality functional annotations of rare variants. Here, we combined evidence from CADD,⁴⁰ REVEL,⁴¹ and SpliceAI⁴² annotations to generate high-confidence sets of variants and integrated evidence from common-variant studies to prioritize putative risk genes for functional follow-up. Using these annotation categories, we built allelic series that showed the effect of each category of variants within a given gene. In assessing the monotonicity of these allelic series, we made the simplifying assumption that missense variants not predicted to be LoF or highly damaging would act in the same direction as pLoFs or damaging variants. However, missense variants may also result in gain of function.¹¹⁰ Categorizing missense variants into gain of function and LoF before assessing monotonicity could result in identifying more genes with monotonic allelic series. Furthermore, our analysis was also only focused on coding variation. In the future, improved rare-variant masks, such as those that better stratify missense variants or incorporate non-coding rare variation outside exonic regions, may help identify additional therapeutic targets to alter fat distribution.

Genome-wide common-variant and exome-wide rare-variant analyses can provide complementary lines of evidence to nominate genes associated with human traits.^109,111 We found that genes implicated for obesity and body fat distribution by a burden of rare damaging variation were (1) associated with rare Mendelian forms of obesity and lipodystrophies, and (2) also prioritized by common-variant prioritization scores⁵⁷ for BMI, WHRadjBMI, and body fat percentage. Converging evidence from across the allele-frequency and phenotypic-severity spectra can therefore identify high-confidence genes for a range of human traits. Assessing the phenome-wide consequences of these genes can provide evidence for therapeutic potential, for example by revealing possible side effects on platelet distribution or neutrophil percentage. Considering the effect of sex is also important. Our analysis revealed significant sex heterogeneity in burden effects for several genes: INSR, PDE3B, and PLIN4. Functional validation of this heterogeneity requires more complex in vivo models in sexually dimorphic organisms. While this was not the focus of our study, such models would provide a deeper mechanistic understanding of the role of sex in fat accumulation.

We measured the effects of CRISPR gene knockdowns on lipid accumulation in hWAT cell lines using the BODIPY assay, which we performed in six technical replicates per knockdown cell line to ensure replicability. We limited the assay to 14 genes due to the difficulty of manipulating adipocytes, which would have made a high-throughput whole-genome CRISPR scan impossible. However, with advances in methodology or technology, such an approach may become feasible.

As the process of adipocyte differentiation is considered a necessary precursor to lipid accumulation,⁸⁵ we interpreted the BODIPY readout as a measure of both lipid accumulation and adipogenesis more generally. However, we acknowledge that lipid accumulation is only one of many pathways involved in adipocyte behavior, alongside glucose uptake, stem cell dynamics, adipocyte lifespan, depot-specific effects, etc. Furthermore, the culture conditions of hWAT cell lines lacked endocrine and neural inputs that can affect overall and depot-specific adiposity. Thus, a failure to find alterations in lipid accumulation and adipogenesis due to gene knockdown in vitro does not imply that the gene is not relevant for obesity nor that it is an unattractive drug target. Genetic variation associated with obesity will likely affect other components of the adipose tissue niche, including innervation, immune cells, or stromal components in a depot-specific manner.^112,113,114 Indeed, BMI-associated genetic variation is known to alter the expression of genes in the central nervous system,¹² and blockbuster GLP1-receptor agonists that are prescribed for weight loss^13,14 act primarily through the dorsomedial hypothalamus to control food intake.¹⁵ Therapeutic targets against obesity will thus need to span the range of mechanisms through which excess body fat may accumulate.

Our confirmed knockdown of INSR did not significantly affect lipid accumulation as measured by the BODIPY assay despite this gene having a generally well-characterized role in body fat accumulation.^115,116 However, INSR was still partly expressed (RNA-seq expression FC [95% CI] = 0.521 [0.452, 0.601]), which may explain why the knockdown did not result in significant changes to lipid accumulation.

Rare missense variation in PPARG is associated with decreased body fat percentage. Concordantly, the disruption of PPARG in hWAT cell lines resulted in decreased lipid accumulation, an example of concordant effects of gene perturbation at both cell and human levels (Figure S9). However, some genes demonstrated contrasting effects. Of the two genes affecting central obesity (WHRadjBMI or ratio of android to gynoid tissue fat percentage) with confirmed knockdowns increasing lipid accumulation, COL5A3 was correspondingly associated with increased central obesity, whereas TRIP10 was nominally associated with decreased central obesity (Figure S9). Similarly, of the two genes carrying rare missense variation associated with increased BMI and with confirmed knockdowns, EXOC7 knockdown nominally increased lipid accumulation in hWAT cell lines while SLTM knockdown significantly decreased lipid accumulation (Figure S9).

Explaining the discordance of SLTM, one of the three gene knockdowns to significantly affect lipid accumulation in hWAT, is of particular interest. One possible explanation could be that SLTM variants annotated as pLoF could actually be gain of function; however, this is unlikely as previous work has corroborated both rare- and common-variant associations between decreased SLTM expression and increased BMI.⁸² Saturation editing of SLTM could provide clarity on effects of specific variants. Another explanation for SLTM’s apparent discordance could be that perturbation of SLTM has different effects across cell types, possibly owing to its broader role as a transcription modulator. For instance, like fellow genes associated with earlier age at onset of obesity such as MC4R and PCSK1, SLTM may affect neurohormonal pathways that were not directly tested by our functional assay. Future work could involve gene perturbation in mouse models, where the functional consequence of SLTM knockdown could be assessed at the whole-organism level. With both concordant and apparently discordant results, our work demonstrates the importance of systematically characterizing the functional effects of gene perturbation at the tissue level, which may not follow the expected direction of effect based on missense variation in humans.

In summary, we demonstrate that the convergence of evidence across rare and common genetic variation can help identify high-confidence target genes for overall adiposity and body fat distribution. We provided in-depth functional characterization through CRISPR knockdowns in human adipocytes, allowing us to nominate candidate genes for therapeutic targets. Through functional readouts and transcriptomic analyses, we also highlighted several molecular mechanisms by which genetic variation may impact adipogenesis. Our results provide a model by which future work can integrate genetic and functional evidence to identify, design, and evaluate potential drug targets to alter overall and tissue-specific fat distribution.

Data and code availability

• •Full summary statistics and transcriptomic datasets are available on Zenodo: SAIGE gene and variant-level summary statistics and fine-mapped GWAS loci (https://doi.org/10.5281/zenodo.15730316); wild-type hWAT transcriptomics (https://doi.org/10.5281/zenodo.15729466); CRISPR knockdown cell line transcriptomics (https://doi.org/10.5281/zenodo.15729561).
• •All code used in this study is available on GitHub: ES QC (https://github.com/lindgrengroup/ukb_wes_450k_qc); gene- and variant-level tests with SAIGE and imputed data QC (https://github.com/lindgrengroup/ukb_wes_450k_gwas); fine mapping of common-variant GWAS loci (https://github.com/lindgrengroup/ukb_wes_450k_finemapping); CRISPR knockdown cell line differential expression and gene set enrichment with GSEA (https://github.com/lindgrengroup/crispr_ko_rna_seq).
• •All other analysis of hWAT cell lines, including wild-type RNA-seq, BODIPY assay, and RT-qPCR confirmation are available on GitHub (https://github.com/lindgrengroup/obesity_hwat).

Acknowledgments

We would like to thank Dr. Val Millar and the Ebner group for expert help in the use of PerkinElmer Opera Phenix/High Content Imaging. We also acknowledge wet lab resources from Prof. Benedikt Kessler (Mass Spectrometry for Drug Target Discovery Lab) and Dr. Adan Pinto Fernandez (Translational Ubiquitomics-Cancer Immunology Lab). This research was conducted using the UK Biobank resource under application number 11867. We thank UK Biobank participants for their contribution.

N.A.B. is supported by the Pembroke College Oxford-Bendich Graduate Scholarship, the Clarendon Fund, and Wellcome Trust grant number 224890/Z/21/Z. S.S.V. is supported by the Rhodes Scholarships, Clarendon Fund, and the Medical Sciences Doctoral Training Centre at the University of Oxford. F.H.L. is supported by the Wellcome Trust (award 224894/Z/21/Z) and the Medical Sciences Doctoral Training Centre at the University of Oxford. M.C. is the Weissman Family MGH Research Scholar Award and supported by the Novo Nordisk Foundation (NNF21SA0072102). C.M.L. is supported by the Li Ka Shing Foundation, NIHR Oxford Biomedical Research Centre, Oxford, NIH (1P50HD104224-01); Gates Foundation (INV-024200); and a Wellcome Trust Investigator Award (221782/Z/20/Z). This research was supported by the Wellcome Trust Core Award grant number 203141/Z/16/Z with additional support from the NIHR Oxford BRC. The views expressed are those of the authors and not necessarily those of the NHS, the NIHR, or the Department of Health.

Author contributions

Conceptualization, C.M.L.; funding acquisition, C.M.L.; data curation, N.A.B., B.H., and F.H.L.; methodology, N.A.B., D.S.P., P.D.C., and M.C.; software, N.A.B., S.S.V., B.H., and F.H.L.; formal analysis, N.A.B., S.S.V., F.H.L., and B.H.; investigation, I.S.E., S.R., and E.N.-G.; writing – original draft, N.A.B., I.S.E., S.S.V., and D.S.P.; writing – review & editing, N.A.B., S.S.V., and D.S.P.; visualization, N.A.B. and S.S.V.; project administration, N.A.B., C.M.L., and D.S.P.; supervision, C.M.L. and D.S.P.

Declaration of interests

S.R. is currently employed by e-therapeutics PLC but while she conducted the research described in this manuscript was only affiliated with the University of Oxford. M.C. is on the SAB of Nestle and SixPeaks Bio, and she further reports grants from Novo Nordisk and Calico. C.M.L. reports grants from Bayer AG and Novo Nordisk, has a partner who works at Vertex, is a part-time employee of PHP, and owns equity in PHP and its subsidiaries.

Published: September 4, 2025