FNIP1 variants are associated with favourable metabolism in 1 million humans

Altered energy and lipid metabolism contribute to cardiometabolic diseases such as coronary artery disease (CAD), obesity, type 2 diabetes and metabolic-dysfunction-associated steatotic liver disease (MASLD), which collectively represent the leading cause of death worldwide and account for a substantial and growing burden of morbidity^1,10. The TG:HDL ratio has been proposed as a biomarker of the metabolic state of the body and its levels are associated with insulin resistance, type 2 diabetes and CAD^{7,11,12,13,14}.

Large exome-sequencing association studies of rare coding variants have led to the identification of therapeutic targets^{8,9,15,16,17,18}. Here we use data from over 1 million people with exome sequencing, genome-wide common variant imputation and detailed health phenotypes from 11 cohorts across the world (Fig. 1 and Supplementary Table 1) to validate the TG:HDL ratio as a biomarker of the energy state of the body and risk of various cardiometabolic diseases.

We also perform a genetic discovery analysis of the TG:HDL ratio, implicating 59 genes of which coding variants are independently associated with this metabolic biomarker, including known and new therapeutic targets. With a combination of human genetics, omics data, and in vitro and in vivo validation in experimental models, we implicate the folliculin interacting protein 1 (FNIP1) and folliculin (FLCN) pathway in energy metabolism and the aetiology of cardiometabolic disease in humans as an illustrative example of how massive rare-coding-variant discovery analyses of this circulating metabolic biomarker can lead to translational and therapeutically relevant insights.

In an in-depth epidemiological analysis of refined anthropometric and metabolic traits, a higher TG:HDL ratio was associated with higher total and visceral adiposity across body imaging modalities (Fig. 2a, Supplementary Fig. 2 and Supplementary Table 2). This was accompanied by greater ectopic fat deposition in the liver and skeletal muscle, higher levels of liver damage biomarkers (for example, alanine aminotransferase (ALT)) and higher levels of liver-biopsy-defined steatosis, inflammation and fibrosis (Fig. 2b,c). There were statistically significant associations with higher fasting insulin, higher glycated haemoglobin, higher blood pressure and higher circulating C-reactive protein (CRP; Fig. 2d). Finally, a higher baseline TG:HDL ratio was predictive of an increased risk of incident type 2 diabetes, myocardial infarction, MASLD and liver cirrhosis (Fig. 2e). These associations were consistently observed across African, admixed American, East Asian, South Asian and European ancestries (Supplementary Figs. 3–7).

Thus, the TG:HDL ratio is a simple and powerful biomarker of energy metabolism and cardiometabolic disease risk, which can be used for genetic and biologic discovery.

We next performed a genetic discovery analysis of the TG:HDL ratio in a diverse set of 1,032,116 individuals from 11 cohorts and three continents (Fig. 1 and Supplementary Table 1).

In multiancestry genome-wide association and fine-mapping analyses of common variants (minor allele frequency ≥ 1%), we identified 1,617 independent signals at 801 physical loci associated with the TG:HDL ratio (Methods, Supplementary Note 1, Supplementary Tables 3 and 4 and Supplementary Fig. 8). In ancestry-specific fine-mapping analyses, we identified 1,406 independent signals in European, 209 in admixed American, 63 in South Asian, 44 in African and 9 in East Asian analyses (Supplementary Note 1, Supplementary Table 5 and Supplementary Figs. 9–13). Common variant signals showed enrichment for genes expressed in adipose and liver tissues (Supplementary Fig. 14) and we found strong genetic associations with anthropometric and metabolic phenotypes (Supplementary Fig. 15).

In an exome-wide gene-level association analysis of rare coding variants (minor allele frequency < 1%), 60 genes were associated with the TG:HDL ratio at the exome-wide level of statistical significance (P < 1.04 × 10⁻⁷), in an analysis adjusted for the 1,617 common variant signals identified in the same individuals (Fig. 3a,b, Extended Data Table 1 and Supplementary Note 2). Of the 60 genes, 59 showed an independent association when adjusting for other gene-burden signals (Supplementary Tables 6 and 7 and Supplementary Note 2), whereas the association of the AC126283.2 gene disappeared in the conditional analysis, reflecting the sequence overlap with PLA2G12A (Supplementary Tables 6 and 8 and Supplementary Note 2).

Of the 59 independent genes, 15 (25%) had been identified in the previous largest rare variant analysis of TG:HDL ratio in around 170,000 people¹⁹, while the remaining 44 (75%) were novel rare-coding-variant associations with the TG:HDL ratio (Supplementary Tables 7 and 9 and Supplementary Note 2).

The associations of the 59 genes with the TG:HDL ratio were robust and consistent in a series of sensitivity and subgroup analyses using different common variant adjustment, ancestry adjustment, phenotype definition approaches, as well as sex-, ancestry- and fasting-state subgroups analyses (Supplementary Tables 10–16, Supplementary Figs. 16–18 and Supplementary Note 2). Gene-burden associations were directionally consistent across ancestry groups, and the higher frequency in non-European ancestry groups of rare variants in some of the genes drove the identification of several associations (for example, LIPE, PCSK9 and GALNT2; Supplementary Table 11, Supplementary Fig. 19 and Supplementary Note 2). In sex-stratified analyses, we noted an interesting sex interaction for PDE3B (I² = 93%; P_interaction = 1.8 × 10⁻⁴; β_women, −0.40 versus β_men, −0.28 s.d. units per allele; Supplementary Table 13, Supplementary Fig. 17 and Supplementary Note 2)—a fat-distribution-associated gene⁹.

In a replication analysis in 114,942 participants from the All of Us Research program, association estimates for the 59 genes showed strong correlations with those of our discovery analysis (Pearson’s r = 0.97, P = 2.1 × 10⁻³⁸; Supplementary Table 14, Supplementary Fig. 18 and Supplementary Note 2).

Using proteomics and immunoassay data, we studied the impact on circulating protein levels for predicted loss-of-function (pLOF) variants in 13 genes with available measurements and found strong associations with lower levels of the encoded proteins for all genes, consistent with a loss-of-function effect (Supplementary Fig. 20). We also intersected protein quantitative trait loci (pQTL) with common variant signals at the gene-burden loci and identified 25 colocalizing signals between pQTLs and TG:HDL fine-mapped peaks, 14 of which converged on the exome-wide significant genes from the rare coding variant analysis (Supplementary Table 17).

Among the 59 genes, there was a considerable over-representation of genes with enriched expression in liver and adipose tissues (odds ratio (OR) for enrichment, 13.6; P_enrichment = 9 × 10⁻¹⁶ for liver; OR of 31.2 and 18.1, P = 1 × 10⁻⁸ and P = 1 × 10⁻⁴ for subcutaneous and visceral adipose tissue, respectively; Fig. 3c,d), further emphasizing the enrichment for the same energy metabolism tissues also observed in the common variant analysis (Supplementary Fig. 14).

The 59 associated genes included: (1) 11 (19%) genes for which coding variants are known to cause monogenic diseases of lipid metabolism (Supplementary Table 7); (2) 31 (53%) genes encoding known drug targets, including 23 for drugs in clinical development in humans or approved by health authorities (Supplementary Table 7); (3) 7 genes in the lipoprotein lipase (LPL) pathway, which regulates tissue energy availability (Extended Data Fig. 1 and Supplementary Table 7); and (4) 5 genes encoding apo-lipoproteins regulating the transport of lipids in the circulation, including a novel association for rare coding variants in the APOC1 gene with a lower TG:HDL ratio (Extended Data Table 1 and Supplementary Tables 6 and 7).

We performed pathway enrichment analyses using DEPICT²⁰, which revealed 761 enriched pathways (at a Bonferroni-corrected P < 5.9 × 10⁻⁶; Supplementary Table 18) dominated by energy metabolism pathways, including lipid and glucose homeostasis, oxygen and energy consumption, lipoprotein metabolism, insulin resistance, liver fat accumulation, lipolysis regulation and fat storage (Supplementary Table 18 and Supplementary Note 2). For these pathways, an average of 16 (interquartile range, 11–19) of our 59 genes contributed to the pathway enrichment association (Supplementary Table 18 and Supplementary Note 2), reflecting a high concentration of true biological signal. Pathway analysis for gene subsets defined by the relative contribution of triglycerides versus HDL-C to the associations with the TG:HDL ratio revealed enrichment for pathways relating to the respective lipoprotein particles (Supplementary Tables 19–21 and Supplementary Note 2).

More than half of the 59 TG:HDL-ratio-associated genes showed gene-burden associations with glycaemic, liver or atherosclerosis related phenotypes (Supplementary Fig. 21), consistent with their central role in energy balance, storage and metabolism (Extended Data Fig. 1). There were also extensive associations with refined lipoprotein particle composition phenotypes measured using nuclear magnetic resonance in 469,068 individuals (Extended Data Fig. 2).

Ultra-rare protein-truncating variants in FNIP1 were observed in around 1 out of every 7,000 people who we sequenced (combined alternative allele frequency (AAF) = 0.01%) and were associated with around 0.5 s.d. lower TG:HDL ratio (P = 1.1 × 10⁻¹⁰), representing a novel rare variant association and one of the largest effect associations identified in our analysis (Fig. 3, Extended Data Table 1 and Supplementary Tables 6 and 7). Associations with favourable phenotypes and disease protection for rare pLOF variants have highlighted a number of therapeutic targets^{21,22,23,24,25,26}, leading to new approved medicines^27,28,29,30, so we examined this association in more depth.

A total of 86 distinct ultra-rare pLOF variants in FNIP1 contributed to the association (Supplementary Table 22), including a validated³¹ loss-of-function variant implicated in immunodeficiency-93 and hypertrophic cardiomyopathy (Online Mendelian Inheritance in Man ID (MIM) 619705), an autosomal recessive disease caused by complete loss of FNIP1 function^31,32,33 (Supplementary Fig. 22a). Associations with TG:HDL ratio were robust when restricting to variants predicted to cause nonsense-mediated decay and for both N- and C-terminal variants (Supplementary Fig. 22b), consistent with a loss-of-function effect.

In over 1 million people with available exome-sequencing and clinical data, FNIP1 pLOF variants were associated with lower atherogenic blood lipids, lower body-mass index, a lower visceral–gluteofemoral fat volume ratio, lower body fat percentage, higher body lean percentage, lower glycated haemoglobin, lower liver fat and lower liver enzyme levels, highlighting a favourable cardiometabolic phenotype across a broad range of traits (Fig. 4a and Extended Data Fig. 3).

In 227,636 cases and 265,114 controls, FNIP1 pLOF variants were associated with around 60% lower odds of cardiometabolic disease outcomes (that is, CAD, type 2 diabetes, MASLD and liver cirrhosis combined; OR = 0.39; 95% confidence interval (CI) = 0.22–0.69; P = 0.0011; Fig. 4a), in the heterozygous state.

Heterozygous FNIP1 variants did not show a statistically significant association with a composite electronic health record phenotype for the clinical features of immunodeficiency due to biallelic loss-of-function of FNIP1 (refs. ^31,33,34), consistent with the recessive inheritance of this Mendelian syndrome (Supplementary Note 3, Supplementary Table 23 and Supplementary Figs. 23 and 24).

FNIP1 encodes folliculin-interacting protein 1, which interacts with folliculin (encoded by FLCN) to inhibit mitochondrial biogenesis, oxidative phosphorylation and energy expenditure downstream of AMPK³⁵.

In our human genetic analysis, FLCN heterozygous pLOF plus deleterious missense variants were associated with a lower TG:HDL ratio (per-allele β = −0.13 ratio units; 95% CI = −0.23 to −0.02; P = 0.017) and with an overall pattern of association with cardiometabolic traits that was similar in directionality, but weaker in magnitude compared with FNIP1 variants (Supplementary Fig. 25). Heterozygous variants of FLCN cause autosomal dominant Birt–Hogg–Dubé syndrome (MIM 607273), an adult-onset disorder characterized by pulmonary cysts, pneumothorax and tumour predisposition, particularly renal³⁶. Compared with non-carriers, heterozygous carriers of pLOF variants in FLCN had an 18-fold higher risk of pneumothorax and a 9-fold higher risk of kidney cancer and were at increased risk for a composite phenotype of Birt–Hogg–Dubé symptoms (Supplementary Figs. 23 and 26a, Supplementary Table 24 and Supplementary Note 3).

We did not observe associations with the TG:HDL ratio for pLOF variants in FNIP2 (Supplementary Fig. 27), which encodes another member of the folliculin pathway with functions that have been hypothesized to be similar to those of FNIP1.

Mechanistically, the FNIP1–FLCN complex inhibits TFEB and TFE3, which are transcription factors that drive mitochondrial and lysosomal gene expression^37,38,39,40. Accordingly, FNIP1–FLCN pathway inhibition promotes energy use and expenditure³⁵ (Extended Data Fig. 4a).

In this context, our human genetics findings suggest that heterozygous loss-of-function of FNIP1 and, to a lesser degree, FLCN may induce mitochondrial biogenesis and energy expenditure in metabolically active tissues resulting in a favourable phenotype. Consistent with this hypothesis, previous literature has proposed a role for Flcn in mouse models of MASLD and liver inflammation, in which liver-specific knockdown has shown benefit^41,42, whereas Fnip1 has not been studied in such models.

FNIP1 and FLCN are expressed in human and mouse hepatocytes (Supplementary Fig. 28), suggesting that at least part of the metabolic benefit observed in human variant carriers may be due to liver energy metabolism. This is important as there is precedent for the successful development and regulatory approval of N-acetylgalactosamine (GalNAc)-conjugated short interfering RNA (siRNA) therapeutics that effectively, durably and specifically silence a variety of genes selectively in hepatocytes in humans^{29,30,43,44,45}. Thus, selective silencing of FNIP1 in hepatocytes with siRNA may be a tractable approach for exploiting the favourable metabolic effects of variants in this pathway, while avoiding potential side effects of systemic loss-of-function.

To validate this hypothesis, we used siRNA to knockdown FNIP1 in primary human hepatocytes. FNIP1 silencing (>90% mRNA reduction) led to induction of lysosomal and lipid catabolism genes (Fig. 4b, Extended Data Fig. 5 and Supplementary Fig. 1), consistent with the reported biology of this pathway^35,46 and with the hypothesis that knockdown of FNIP1 in human hepatocytes affects metabolic gene expression.

We next turned to in vivo mouse models to evaluate the impact on whole body energy homeostasis of FNIP–FLCN pathway inhibition in the liver. We used guide RNAs (gRNAs) delivered by adeno-associated virus serotype 8 (AAV8-gRNAs) to ablate Flcn, Fnip1, Fnip2 or Fnip1 together with Fnip2 in the liver of Cas9-expressing mice challenged with a high-fat, high-fructose diet (HFHFD). DNA editing, mRNA and protein analysis confirmed the targeting approach (Extended Data Fig. 4b,c,e), which led to robust induction of mitochondrial and lysosomal genes in liver (Extended Data Fig. 4d and Extended Data Fig. 6a–c). Mice with hepatic Flcn deletion were substantially protected against weight and fat gain and had a higher proportion of lean mass (Fig. 4c,d and Extended Data Fig. 7a–e). Individually, Fnip1 or Fnip2 inhibition had no impact on weight gain, but combined Fnip1 + Fnip2 inhibition protected against diet-induced obesity similarly to Flcn inhibition (Fig. 4c,d and Extended Data Fig. 7a–e). The lack of phenotype in mouse with selective Fnip1 liver inhibition probably reflects interspecies differences between mice and humans, given that (1) in our human genetics analysis, none of the 155 carriers of pLOF variants in FNIP1 had any FNIP2 pLOF variants, as expected from their genotype frequency (Supplementary Table 25), which means that the phenotype is solely due to FNIP1 variants; (2) FNIP2 variants did not show robust associations with metabolic traits; and (3) human hepatocyte experiments showed the induction of lipid catabolism and lysosomal genes with the specific knockdown of FNIP1.

Mice with liver Flcn or Fnip1 + Fnip2 inhibition exhibited improved insulin tolerance, lower circulating insulin, reduced liver triglyceride content and lower ALT levels on an HFHFD (Fig. 4e,f and Extended Data Fig. 7f,g). These mice also displayed metabolically beneficial gene expression profiles in adipose tissue, with higher transcript levels of adipose master regulator Pparg and its target genes, as well as higher transcript levels of lipolysis and oxidative phosphorylation genes⁴⁷ (Extended Data Fig. 8a–d). The adaptations in adipose tissue appeared to be secondary to changes in hepatic energy metabolism, given that our AAV8-gRNA targeting approach did not perturb the Flcn, Fnip1 and Fnip2 genes in non-hepatic tissues, including adipose (Extended Data Fig. 8a and Supplementary Fig. 29). Notably, hepatic Inhbe transcript levels were reduced by around 52–55% after Flcn or Fnip1 + Fnip2 inhibition (Extended Data Fig. 8e). Activin E, encoded by Inhbe, is a hepatokine that inhibits adipose lipolysis and regulates adipose depots through the ALK7 receptor^9,47, and similar levels of Inhbe reduction have previously been shown to elicit weight loss, fat loss and increased lipid oxidation in obese mice⁴⁸. Thus, the favourable adipose gene expression signatures observed with hepatic FNIP–FLCN pathway inhibition may be partly mediated by a reduction in hepatic activin E expression.

To validate that the energy metabolism changes observed in our in vivo experiments were due to specific gene targeting in hepatocytes as opposed to other liver cells, we used GalNAc-conjugated Flcn siRNAs and demonstrated that selective hepatocyte Flcn mRNA and protein knockdown was associated with a phenotype analogous to that observed in the AAV8-gRNA studies (Extended Data Fig. 9).

We further performed distinct series of AAV8-gRNAs experiments investigating liver Flcn or Fnip1 + Fnip2 inhibition at thermoneutrality or in a chow diet model (Supplementary Figs. 30 and 31). In experiments at thermoneutrality, we observed similar protective phenotypes in the treated mice as those observed in the ambient temperature conditions (Supplementary Fig. 30), indicating that enhanced energy metabolism, rather than cold stress, confers metabolic benefits on Flcn or Fnip1 + Fnip2 hepatic inhibition. In chow diet experiments, we observed that Flcn or Fnip1 + Fnip2 inhibition protected against fat mass gain and lowered liver triglycerides, although, as expected, the phenotype was less pronounced than that observed in the HFHFD challenge model (Supplementary Fig. 31).

Finally, to evaluate the long-term consequences of hepatic silencing, we performed a 30-week Flcn AAV-gRNA experiment in the setting of HFHFD, which demonstrated reduced liver lipids and markedly lower ALT levels, indicative of favourable metabolism and liver protection (Extended Data Fig. 10a–e). This was further supported by gene expression profiling demonstrating that long-term Flcn inhibition prevented the induction of fibrosis and liver damage/repair genes that was observed in control animals on an HFHFD (Extended Data Fig. 10f).

Given the enrichment for liver-expressed genes in our discovery analysis, we used liver RNA-sequencing (RNA-seq) and expression QTL (eQTL) data from a bariatric surgery cohort of 1,946 people for deeper mining of causative genes.

Of 1,617 independent common variant associations with TG:HDL ratio in our analysis, 151 (9%) at 150 genes co-localized with a sentinel liver eQTL (posterior probability > 0.9; Supplementary Table 26). Of the 150 genes implicated by the co-localized eQTLs, 9 (15%) were among the 59 exome-wide significant genes (a 21-fold enrichment compared with what is expected by chance; P_enrichment = 1.4 × 10⁻⁹), while an additional 9 genes met a more liberal Bonferroni correction for 150 liver eQTL-implicated genes (gene-burden P value greater than 1.04 × 10⁻⁷, but smaller than 3.3 × 10⁻⁴) (Supplementary Table 26).

Among the 18 genes with both gene-burden and common variant liver eQTL evidence, 15 showed directional concordance, whereby the liver-expression-lowering allele was associated with the TG:HDL ratio in the same direction as the rare deleterious coding variants (observed percentage, 83%; expected percentage under null hypothesis, 50%; P_binomial = 0.0075) (Supplementary Table 27), indicative of true signal enrichment.

Among these genes was hepsin (HPN), which encodes a hepatocyte-specific transmembrane serine protease that is known to curb energy expenditure and prevent adipocyte browning in mice, in part due to regulation of HGF/MET signalling^49,50 (Supplementary Figs. 32 and 33a,b). An eQTL for higher liver expression of hepsin co-localized with an association peak for higher TG:HDL ratio (rs45512696) (Fig. 5a,b, Supplementary Fig. 34 and Supplementary Table 26). Conversely, ultra-rare pLOF variants in HPN showed large-effect associations with lower levels of TG:HDL ratio, lower triglycerides, lower LDL-C and lower circulating albumin (Fig. 5c and Supplementary Fig. 35).

On the basis of these orthogonal lines of human genetics evidence, we hypothesized that reducing liver expression of hepsin might lead to lower atherogenic lipid levels and tested this hypothesis in mice. AAV8-delivered short-hairpin RNAs (AAV8-shRNAs) targeting Hpn achieved 87% mRNA knockdown in the liver and were associated with lower plasma cholesterol and triglycerides, but had no impact on body weight (Fig. 5d–h and Supplementary Fig. 33c,d). Furthermore, consistent with human genetics associations, Hpn knockdown led to reduced circulating albumin and total protein levels (Fig. 5i and Supplementary Fig. 33e).

Our large epidemiological and genomic study of the TG:HDL ratio provides substantial evidence advancing understanding of the genetic basis of lipid and energy metabolism in humans.

Important insights from this analysis include (1) the validation of the TG:HDL ratio as a biomarker of body energy state and risk of cardiometabolic disease in the general population and across ancestries; (2) the identification of several dozen associations with the TG:HDL ratio for rare coding variants, greatly expanding the catalogue of genes and molecular players in lipid and energy metabolism in humans; (3) the description of ultra-rare coding variants in the FNIP1 pathway, showing large-effect protective associations with a broad range of cardiometabolic traits and diseases (atherogenic lipids, glycaemia, liver fat and damage, and cardiometabolic disease outcomes); and (4) the demonstration of convergence of rare coding and common variant associations, showing a central role of liver genes in TG:HDL-ratio biology, including the identification of rare variants in hepsin conferring a favourable lipidemic profile.

Our genetic findings demonstrate the importance of the FNIP1 pathway in human energy metabolism and in the risk of cardiometabolic diseases in the general population. Previous work^35,41,46 and the experimental data presented here suggest that the protective effects of FNIP1 variants may be due to enhanced lipid breakdown, mitochondrial fatty acid oxidation and energy expenditure in metabolically active tissues. It is therefore possible that therapeutic modulation of FNIP1 might be beneficial in a range of cardiometabolic disease settings through enhanced fat breakdown and energy expenditure.

Given the precedent for the development of hepatocyte targeted siRNA therapeutics in humans^{29,30,43,44,45}, we specifically explored the possibility that liver inhibition of this pathway may induce a favourable metabolic profile. In human hepatocytes, silencing of FNIP1 induced lipid breakdown and lysosomal genes and, across various mouse models, hepatic silencing of Flcn and of Fnip1 + Fnip2 induced a similar expression profile in liver and adipose tissues, with beneficial impacts on weight gain, circulating lipids, liver fat, insulin sensitivity and glycaemic control.

Conceptually, selective inhibition in hepatocytes may also help to avoid potential adverse phenotypes of systemic loss-of-function of this pathway, such as those observed in humans with heterozygous FLCN variants, homozygous FNIP1 variants or, to a milder degree, heterozygous FNIP1 variants. Liver safety is another important consideration and previous experiments in mouse models of metabolic liver disease have shown benefits for liver-specific inhibition of Flcn^41,42. However, a recent study has shown susceptibility to liver injury and carcinogenesis for knockout of Flcn in hepatocytes and cholangiocytes in mice⁵¹. In our experiments, liver knockdown was associated with a protective liver phenotype, particularly in the HFHFD setting. In humans, liver cancer has not been reported as a clinical feature of homozygous FNIP1 deficiency^{31,32,33,34,52,53} or Birt–Hogg–Dubé syndrome³⁶ and, in our large dataset, FNIP1 variants were associated with lower ALT, lower liver fat and protection from metabolic liver disease consistent with a favourable liver phenotype. Neither FNIP1 nor FLCN heterozygous variants showed statistically significant associations with hepatic cancers. Besides the liver, previous literature on mouse models has suggested a role for the FNIP1–folliculin pathway on skeletal muscle and adipose tissue metabolism^{54,55,56,57,58,59}. These findings are broadly consistent with results from our human genetics study and it is therefore possible that muscle- or adipose-specific inhibition of this pathway may also be beneficial to human metabolism.

Taken together, our data demonstrate the importance of the FNIP1 pathway in human energy metabolism. Future work across experimental settings will ultimately clarify the full phenotypic landscape, as well as the potential for efficacy and safety of modulating this pathway in humans.

In conclusion, our human genetic analysis of the energy-state biomarker TG:HDL ratio in over 1 million people provides an illustration of the power of mega-scale exome sequencing and rare coding variant analysis to help map biologically important pathways in human disease, including the identification of potential therapeutic targets for common cardiometabolic diseases that collectively still represent the most common cause of death worldwide.

Study population

The genetic discovery analysis of TG:HDL ratio encompassed multiple cohorts with a cumulative sample size of 1,032,116 participants. Study participants were recruited from various population-based and hospital-based health systems across three continents (Fig. 1 and Supplementary Table 1). These cohorts included: the UK Biobank population-based study (UKB, n = 409,602)⁶⁰; the Geisinger Health System MyCode cohort (GHS, n = 143,129)⁶¹; the Mexico City Prospective Study (MCPS, n = 136,725), a study of residents in two districts of Mexico City (Coyoacán and Iztapalapa)⁶² with a genetically admixed population with components of Indigenous American, European and African ancestry as described in detail previously⁶³; the Mayo Clinic Project Generation health system-based cohort (MAYO-RGC, n = 80,941)⁶⁴; the population-based BangladEsh Longitudinal Investigation of Emerging Vascular and nonvascular Events (BELIEVE, n = 69,663)⁶⁵; the Colorado Center for Personalized Medicine biobank, a health system-based cohort (CCPM, n = 56,173)⁶⁶; the UCLA ATLAS Precision Health Biobank, a health system-based cohort (ATLAS, n = 45,533)⁶⁷; the Mount Sinai BioMe BioBank, a health system-based cohort (BioMe, n = 45,440)⁶⁸; the University of Pennsylvania Penn Medicine BioBank, a health system-based cohort (PMBB, n = 25,785)⁶⁹; the population-based Dallas Biobank Study at UT Southwestern Medical Center (DBS, n = 13,979)⁷⁰; and the population-based Malmö Diet and Cancer Study (MDCS, n = 5,146)⁷¹. Analyses of other cardiometabolic phenotypes included participants from the aforementioned cohorts and also 4,755 participants from the Indiana University School of Medicine (Indiana-CLDB) study and 53,325 participants from the South Asia Biobank⁷². All studies received approval from the relevant ethics committees as listed below, and the participants gave their informed consent to take part in the research. UKB cohort: ethical approval for the UKB was obtained from the North West Centre for Research Ethics Committee (11/NW/0382). The work described here was approved by the UKB under application number 26041. GHS cohort: The MyCode Community Health Initiative was approved by the Geisinger institutional review board (IRB; study 2006-0258). MCPS cohort: the study was approved by scientific and ethics committees within the Mexican National Council of Science and Technology (0595 P-M), the Mexican Ministry of Health and the Central Oxford Research Ethics Committee (C99.260). MAYO-RGC cohort: the study was approved by the Mayo Clinic IRB under protocol number 09-007763. BELIEVE cohort: BELIEVE has received approvals from the relevant institutional review boards of the Bangladesh Medical Research Council, the National Heart Foundation Hospital and Research Institute, icddr,b and Bangabandhu Sheikh Mujib Medical University (BMRC/NREC/2013- 2016/390; BMRC/NREC/2016-2019/243; BSMMU/2019/1184; BSMMU/2019/1185; PR-18051; HBREC.2019.09). CCPM cohort: biospecimens and associated data used in this study were obtained from the biobank at the Colorado Center for Personalized Medicine (CCPM) at the University of Colorado Anschutz Medical Campus (CU AMC). All samples and data were collected under IRB approved protocol (15-0461). ATLAS cohort: ATLAS is an approved study by the University of California Los Angeles (UCLA) IRB (UCLA IRB 17-001013). BioMe/Mount Sinai Million Health Discoveries Program: The Icahn School of Medicine at Mount Sinai’s IRB, Program for the Protection of Human Subjects (PPHS), approved the BioMe Biobank and Mount Sinai Million Health Discoveries Program (PPHS IRB 11-01139 and 21-01743). PMBB cohort: the PMBB is approved by the University of Pennsylvania IRB under protocol number 813913. DBS cohort: The University of Texas Southwestern (UTSW) Biobank was reviewed and approved by the UTSW IRB under the study STU 022011-116: “Genetic Variation in a Multi-Ethnic Population”. MDCS cohort: IRB approval for MDCS was obtained from the Ethics committee of Lund University (LU 51-90) and the Regional Board of Ethics in Lund (Dnr 2016/479). Indiana-CLDB: the study has been reviewed by Indiana University IRB and approved under the protocol number 1105005445. South Asia Biobank: the study received ethical approval from the institutional Research Ethics Committee (18IC4698).

Clinical phenotypes

Clinical biomarkers

Measurements of blood lipids, glycaemic markers, transaminases, CRP and blood pressure were obtained either through direct clinical assessments or electronic health record (EHR) extraction, depending on the cohort type. For participants in population-based cohort studies, these measurements were conducted during standardized visits to designated assessment centres. By contrast, for participants from health-system-based cohorts, data were extracted retrospectively from EHR datasets and participant-specific median values were used. Blood lipid levels were primarily assessed using routine clinical chemistry methods across most cohorts. In the MCPS and BELIEVE studies, lipid measurements were derived from nuclear magnetic resonance metabolomics profiling, performed using the Nightingale Health platform⁷³. Blood lipids were corrected for medication use by dividing by a class-specific correction factor obtained from previously reported clinical trials^{19,74,75,76,77,78}. The TG:HDL ratio was calculated using triglyceride and HDL cholesterol values obtained from the same blood draw at baseline in epidemiological cohort studies (5 cohorts and 635,115 participants). In EHR-based studies (6 cohorts and 397,001 participants), where multiple measurements were available, median values across clinical encounters were calculated for TG and HDL separately and then the ratio of the median values was calculated. The use of median values across encounters is a standard approach in EHR-based genetic research to define a patient’s typical lipid profile^79,80,81. Importantly, this avoids the exclusion of participants with valid but non-concurrent lipid measurements, thereby preserving sample size and discovery power. A sensitivity analysis was conducted restricting the TG:HDL ratio to values derived from TG and HDL measured concurrently from the same blood draw. A stratified analysis by fasting status was also conducted in the UKB, comparing TG:HDL ratios derived from fasting samples (defined as at least 6 h since the last meal) with those from non-fasting samples (defined as at most 2 h since the last meal). Blood pressure levels were corrected by adding 15 mmHg to systolic and 10 mmHg to diastolic blood pressure in those individuals receiving anti-hypertensive treatment as previously reported⁸².

Liver biopsy

The bariatric surgery cohort at Geisinger Health System (GHS) comprises 3,599 participants of European ancestry who underwent bariatric surgery procedures⁶¹. Intraoperative wedge liver biopsies were systematically obtained 10 cm left of the falciform ligament before any manipulation of the liver or stomach. Biopsies were sectioned, with the principal portion allocated for clinical histopathology—fixed in 10% neutral buffered formalin and stained with haematoxylin and eosin for general assessment and Masson’s trichrome for fibrosis evaluation. The remaining tissue was preserved in a research biobank, either stabilized in the RNAlater tissue collection system (Thermo Fisher Scientific) or snap-frozen in liquid nitrogen. Histopathological evaluation was conducted by an experienced pathologist and independently reviewed by a second pathologist. Scoring followed the Nonalcoholic Steatohepatitis Clinical Research Network (NASH CRN) system⁸³, with steatosis graded as 0 (<5% parenchymal involvement), 1 (5 to <34%), 2 (34 to <67%) or 3 (>67%); lobular inflammation as 0 (none), 1 (mild, <2 foci per ×200 field), 2 (moderate, 2–4 foci per ×200 field) or 3 (severe, >4 foci per ×200 field); hepatocyte ballooning as 0 (none), 1 (few ballooned cells), or 2 (many/prominent ballooned cells); and fibrosis staged from 0 (none) to 4 (cirrhosis) according to established criteria. The MASLD activity score was generated by summing of the steatosis, lobular inflammation and ballooning scores.

Imaging phenotypes

MRI

MRI images were obtained from the UKB imaging cohort. For quantification of liver fat, we used data from a two-dimensional abdominal MRI scan that included the liver; a subset of individuals was scanned using a Dixon gradient echo protocol, while those imaged from 2016 onward were scanned using the IDEAL (iterative decomposition of water and fat with echo asymmetry and least-squares estimation) protocol. The resulting data have an in-plane pixel size of 2.5 × 2.5 mm and a slice thickness of 6 mm.

Whole-body MRI scans, spanning from the neck to the knees, were obtained through a two-point Dixon⁸⁴ spoiled gradient-echo (FLASH) T1-weighted protocol with echo times of 2.39 ms (out of phase) and 4.77 ms (in phase), a repetition time of 6.69 ms and an excitation flip angle of 10 degrees⁸⁵. Image acquisition was split across six distinct anatomical stages, each representing a separate 3D volume, with an in-plane spatial resolution of 2.23 mm × 2.23 mm and an out-of-plane resolution ranging from 3.00 mm to 4.50 mm according to the specific stage.

All images were captured using Siemens MAGNETOM clinical MRI scanners (Siemens Healthineers).

MRI-derived phenotypes

To assess total body adipose tissue volumes, we used an automatic segmentation model developed to delineate major adipose depots from Dixon MRI data into visceral and subcutaneous compartments⁸⁶; visceral fat was further subdivided into mediastinal and abdominal components, and subcutaneous fat into the upper-body, abdominal and gluteofemoral components. The volume of fat within these regions was computed by summing the fat fractions for voxels within the segmented regions. The visceral-to-gluteofemoral fat ratio—a marker associated with poor metabolic outcomes⁹—was computed as the abdominal visceral fat volume divided by the subcutaneous gluteofemoral fat volume; the abdominal-to-gluteofemoral subcutaneous fat ratio was similarly computed using the respective total volumes. All data processing and phenotype extraction was conducted using the UKB Research Analysis Platform.

The liver fat percentage, measured as proton density liver fat fraction (PDFF), represents the proportion of fat content in the liver. A deterministic image processing algorithm was used to segment the liver on MRI images, using a multithresholding approach to exclude vessels. The methodology was validated using a publicly available phantom dataset containing vials with varying fat concentrations. PDFF within the segmented region was determined as the ratio of fat signal to the total fat and water signals^16,87. To account for any systematic differences between PDFF estimates from gradient echo and IDEAL, the mean values were shifted to be equal for the two modalities, and an indicator variable encoding the imaging modality was included as an additional covariate for genetic analyses to account for any residual differences.

BIA

BIA for body composition was performed at the baseline using the Tanita BC 418ma Body Fat Analyzer. Fat and fat-free mass estimates derived from BIA were obtained from UKB fields 23100 and 23101, respectively; these are defined such that fat-free mass is equal to total body weight minus fat mass. The body fat percentage was calculated using the BIA-derived fat mass and baseline weight (UKB field 23098).

DEXA

DEXA measures were obtained using the GE-Lunar iDXA instrument. DEXA-derived measures of lean mass and android/gynoid fat mass were obtained from the UKB (UKB fields 23280, 23245 and 23262 respectively).

Disease definitions

Type 2 diabetes

Individuals diagnosed with type 2 diabetes were ascertained according to established methodologies⁹. Case identification was based on the presence of (1) at least one inpatient or two outpatient EHR entries with type 2 diabetes diagnosis codes (ICD-10: E11, O24.1; or ICD-9 equivalents), or documentation as a cause of death; (2) glycaemic biomarker measurements (HbA1c, random or fasting glucose) consistent with the diabetic range⁸⁸ or algorithmically defined type 2 diabetes status derived from self-reported medical history and medication use⁸⁹. Participants were excluded from cases if they met criteria for type 1 diabetes, indicated by ICD-10 codes E10, O24.0 or algorithmic classification based on self-reported and medication data⁸⁹. Control participants were those not fulfilling type 2 diabetes criteria, with further exclusions applied for any EHR evidence of diabetes (any type), family history of diabetes or glycaemic biomarkers within the prediabetic range.

Liver disease

Cases of liver disease, encompassing MASLD and liver cirrhosis, were identified according to previously reported criteria¹⁶. Case inclusion required fulfilment of at least one of the following: (1) documentation of liver disease in EHRs from at least one inpatient encounter, two or more outpatient encounters or as a recorded cause of death; (2) self-reported diagnosis at the baseline; or (3) documented history of surgical or medical interventions related to liver disease. Individuals not meeting these criteria were classified as controls. Exclusion from the control group was applied if any of the following were present: (1) diagnosis of a liver disease not qualifying as a case; (2) outpatient encounter for the liver disease of interest; (3) elevated ALT levels (>25 IU l⁻¹ in women, >33 IU l⁻¹ in men); or (4) diagnosis of ascites attributed to liver pathology.

CAD

Cases of CAD were identified according to previously described criteria⁹. Inclusion as a case required fulfilment of at least one of the following: (1) documentation of CAD and/or myocardial infarction in EHRs from at least one inpatient encounter, two or more outpatient visits or as a recorded cause of death; (2) self-reported diagnosis of CAD or myocardial infarction at the baseline; or (3) documented history of surgical or medical interventions for CAD, including coronary artery bypass grafting or percutaneous coronary intervention. Individuals not meeting case criteria were classified as controls. Control samples were further excluded if a family history of CAD was identified, on the basis of EHR or self-reported information at the study baseline.

Epidemiological analysis methods

The associations between the TG:HDL ratio and various phenotypes—including body weight, fat distribution, glycaemic indices, hepatic steatosis, liver injury markers and histopathological features—were estimated using correlation or linear regression models. TG:HDL ratio levels were stratified in percentiles. Cox proportional hazards models were used to estimate the time-to-event relationship between the TG:HDL ratio and the incidence of type 2 diabetes, myocardial infarction, MASLD and liver cirrhosis in at-risk individuals, that is, without the outcome at or before baseline.

DNA preparation and exome sequencing

As previously outlined⁹⁰, genomic DNA (gDNA) libraries were constructed by enzymatically fragmenting high-molecular-mass DNA to achieve an average fragment length of 200 bp. To facilitate multiplexed exome capture and sequencing, unique 10 bp asymmetric barcodes were incorporated into the DNA fragments of individual samples during library amplification. Equimolar quantities of these barcoded libraries were combined and subjected to exome capture using either a modified xGen Exome Research Panel probe library (Integrated DNA Technology) or a modified version of the human comprehensive exome panel (Twist Bioscience). After PCR amplification and quantification, the enriched libraries were multiplexed and sequenced on Illumina platforms, generating 75 bp paired-end reads. Sequencing was performed on the Illumina HiSeq 2500 and NovaSeq 6000 instruments, using S2 or S4 flow cells.

Read mapping and variant calling

Sequencing reads in FASTQ format were generated from Illumina image data using bcl2fastq (v.2.19.0) (Illumina). In accordance with the original quality functional equivalent (OQFE) protocol⁹¹, reads were aligned to the GRCh38 reference genome using BWA MEM (v.0.7.15)⁹² in an alt-aware configuration. Duplicate reads were identified and flagged, and additional per-read annotations were incorporated. Variant detection was performed using a Parabricks-accelerated implementation of DeepVariant (v.0.10), using a custom model to call single-nucleotide variants and short insertions and deletions, resulting in per-sample genome variant call format (VCF) files⁹³. These individual VCF files were subsequently merged using GLnexus (v.1.4.3)⁹⁴ to produce a joint-genotyped, multisample cohort-level VCF (pVCF). For downstream analyses, PLINK (v.1.9)⁹⁵ was used to convert the pVCF file into PLINK-compatible files.

Exome sequencing quality control

For each cohort, we performed exome sequencing quality control with the following parameters: sample missingness <10%, variant missingness <10%, ±10 bp buffer window around target regions, a low heterozygosity HWE P > 1 × 10⁻¹⁰⁰ and an excess heterozygosity HWE P > 1 × 10⁻³⁰.

Variant annotation

Variant annotation was performed using Variant Effect Predictor (VEP, v.100.4)⁹⁶, using Ensembl release 100 human protein-coding transcript models⁹⁷. For all analyses, a single functional consequence was assigned to each variant on the basis of its annotation on the canonical transcript. The canonical transcript for each gene was determined using a combination of MANE⁹⁸, APPRIS⁹⁹ and Ensembl canonical tags, following previously described procedures⁹⁸.

Nonsense-mediated decay (NMD) predictions were performed using the NMD plugin in VEP⁹⁶. A variant was predicted to escape NMD if it satisfies any of the four canonical rules: (1) the variant is located in the last exon of the transcript; (2) the variant is located within 50 bases upstream of the penultimate (second to last) exon; (3) the variant is located within the first 100 coding bases of the transcript; and (4) the transcript is intronless, containing only a single exon. A variant is predicted to escape NMD if it satisfies any of these four rules^100,101.

Genetic association analyses

Genetic association analyses were conducted in each cohort using the linear whole-genome regression approach implemented in REGENIE (v.3.4 or older)¹⁰². Two primary analyses were performed across autosomes (chromosomes 1–22) and chromosome X in the discovery cohort: a genome-wide association study (GWAS) of common TOPMed- imputed variants¹⁰³ (AAF ≥ 1%) and a gene-based association analysis of rare coding variants. In the first step of REGENIE (trait prediction based on genetic data), directly genotyped variants were included if they had an AAF ≥ 1%, missingness < 10%, Hardy–Weinberg equilibrium P > 10⁻¹⁵ and passed linkage disequilibrium (LD) pruning (window size, 1,000 variants; sliding window, 100 variants; r² threshold, 0.1). The association model in the second step of REGENIE incorporated the following covariates: age, age², sex, age × sex, age² × sex, the first 10 principal components (PCs) based on common variants (derived from a set of LD-pruned array variants: 1,000-variant windows, 50-variant step size, r² threshold 0.1), the first 20 PCs based on rare variants (the same pruning parameters) and cohort-specific sequencing batch covariates. The second step of REGENIE also included as a covariate a genome-wide leave-one-chromosome-out polygenic score generated in step 1, which adjusts for relatedness and residual population stratification¹⁰². For each variant or gene burden being tested in step 2, the polygenic score used leaves out the chromosome of the variant or the gene burden being tested (https://rgcgithub.github.io/regenie/)¹⁰². The main analysis was a pooled multiancestry analysis (that is, not stratified by ancestry). We also separately performed sensitivity analyses stratified by ancestry (that is, European, admixed American, South Asian, African and East Asian ancestries). For the non-pseudoautosomal regions of chromosome X, a dosage compensation model was used: homozygous reference males were coded as 0, hemizygous males as 2 and heterozygous males set as missing.

To ensure independence between rare and common variants signals in the same chromosome, exome-wide gene-burden analyses were additionally adjusted for common variant signals (minor allele frequency ≥ 1%) identified by fine-mapping in the chromosome of the gene-burden being tested, a procedure that is orthogonal to REGENIE’s leave-one-chromosome-out polygenic score approach and ensures control for local linkage disequilibrium between common and rare variants. We also performed a sensitivity analysis in which the adjustment was performed using common variants identified by ancestry-specific fine-mapping. For the 59 independent gene-burden signals identified in the main analysis, we also performed ancestry-, fasting-state- and sex-stratified gene-burden analyses. To assess whether the inclusion of age and sex covariates introduced bias through differential contributions to TG versus HDL cholesterol estimates, for the 59 independent gene-burden signals, we conducted quality control analyses by removing each covariate group one at a time (that is, age, age², sex, age × sex, and age² × sex) from the genetic association models, finding correlations between effect estimates across the 59 genes r > 0.99 for each leave-one-covariate-out analysis for TG, HDL and TG:HDL ratio, consistent with minimal co-variate impact and consistent influences across TG and HDL.

Common variant analysis

A GWAS of common variants (minor allele frequency ≥ 1%) was conducted for the TG:HDL ratio. Association analyses were performed within each cohort by fitting linear regression models in REGENIE. Subsequently, cohort-specific results were combined using fixed-effect inverse variance-weighted meta-analysis. After meta-analysis, SuSiE (v.0.12.35)¹⁰⁴ was used to identify the most probable causal variants for each independent signal surpassing the genome-wide significance threshold of P < 5 × 10⁻⁸. To define fine-mapping regions, we first identified variants with P < 1×10⁻⁶ and separated them into regions at least 100 kb apart. Only regions containing at least one genome-wide significant variant (P < 5 × 10⁻⁸) were retained. Each region was then extended to the nearest recombination hotspot, with extensions ranging from 25 kb to 250 kb. To ensure adequate coverage for fine-mapping, regions smaller than 100 kb were symmetrically expanded to meet the minimum length requirement. Finally, a 100 kb buffer was applied around each region and regions with overlapping buffers were merged into a single region to form the final loci for causal variant identification. For each locus, an exact LD matrix was computed on the basis of the individuals included in the association analysis using the pooled estimate of the covariance across cohorts. This was calculated using all cohorts in the meta-analysis and the exact set of individuals used for analysis in each cohort. Specifically, the ith cohort’s covariance matrix was weighted by its degrees of freedom (n_i − 1), values across all cohorts were then summed and the sum was divided by the total degrees of freedom (N − k) where N is the overall sample size and k is the number of cohorts in the meta-analysis. The generated pooled covariance matrix was then used for fine-mapping using SuSiE (v.0.12.35)¹⁰⁴. SuSiE generates credible sets of common variants at each locus, representing a high likelihood of causality for the observed signal. Within these credible sets, variants are assigned posterior inclusion probabilities (PIP), summing to unity, with higher PIPs indicating a greater probability of causal association. For the main analysis, we set the L parameter of maximum theoretical independent signals per region to 30. For signal within each region, we identified the minimum set of variants capturing 95% of the cumulative PIP. This set represents the 95% credible set, and the variant with the highest PIP at each locus was designated as the sentinel variant. Sentinel variants with frequency ≥1% were adjusted for in the gene-based association analysis as described above.

Rare variant analysis

Variants predicted to induce frameshifts, premature stop codons or disrupt canonical splice donor or acceptor sites were classified as pLOF. Missense variants were further annotated using six computational pathogenicity prediction tools: ESM1vp¹⁰⁵, SIFT¹⁰⁶, PolyPhen2+HumDiv¹⁰⁷, PolyPhen2+HumVar¹⁰⁷, MutationTaster¹⁰⁸ and LRT¹⁰⁹. Variant inclusion for gene-burden testing was determined by AAF, pLOF status (including stop gained, frameshift, splice donor and splice acceptor variants) and, for missense variants, the number of algorithms predicting a deleterious effect. Burden tests were conducted across all combinations of four AAF thresholds (maximum AAF across ancestries)—singleton, 0.0001, 0.001 and 0.01—and six variant class groupings: (1) pLOF; (2) pLOF plus deleterious missense in 1 out of 5 algorithms (non-ESM1vp); (3) pLOF plus deleterious missense in 5 out of 5 algorithms (non-ESM1vp); (4) pLOF plus deleterious missense according to ESM1vp; (5) pLOF plus deleterious missense in 5 out of 5 algorithms and according to ESM1vp; (6) pLOF plus any missense. For each gene, only the canonical transcript was considered for variant annotation and gene-burden analysis. For each gene, this approach yielded a total of 24 gene-burden exposures, the P values of which were subsequently aggregated using ACAT to derive the overall BURDEN-ACAT P value. BURDEN-ACAT¹¹⁰ uses the ACAT¹¹¹ method to integrate results from 24 distinct burden tests into a single aggregated P value. P < 1.04 × 10⁻⁷ (Bonferroni corrected for 20,000 genes and 24 different gene-burden exposures per gene) was used as the exome-wide significance threshold in the exome-wide gene-burden discovery analysis, consistent with previous literature^8,16. A within-chromosome conditional analysis was also performed adjusting each gene-burden signal for other gene-burden signals on the same chromosome.

Tissue-enrichment analysis

Expression tissue enrichment for identified genes was calculated using gene expression values from the V8 data freeze from GTEx, as previously described⁸. After RNA-seq quality-control steps described in GTEx, the median expression level across individuals was calculated for each tissue and gene. For each tissue, z_tissue scores were calculated by standardizing TMM-normalized gene-expression values for each gene using the median and median absolute deviation of all gene expression values within that tissue. Within a gene, z_gene values were generated using the same standardization approach, across tissues, on z_tissue values. This second layer of standardization ensures that comparisons for a gene across tissues are easier to make and interpret. For each gene, expression enhanced tissues are defined as those where z_gene scores are at least 6 s.d. greater than the median z_gene scores observed for that gene across tissues, while expression enriched tissues are defined as those that meet the enhanced definition and have the largest positive deviation. To quantify per tissue enrichment of gene-burden associations from the TG:HDL ratio exome-wide analyses, we first aggregated all of the P values across all burden tests using the Cauchy combination test¹¹¹ per gene. The OR is based on assessing the relationship between the outcome indicator and the tissue specific indicator variable of the observed expression enriched status per gene. The outcome indicator variable is 1 if the per gene aggregated P ≤ 1.04 × 10⁻⁷ and 0 otherwise. The tissue specific indicator variable is set to 1 for a given gene and tissue if that tissue is identified as an ‘expression enriched tissue’ for that gene (as defined above); otherwise, it is 0. The enrichment OR is then calculated using the Firth-corrected approach¹¹² to account for cells with zero counts after excluding all primary sexual and reproductive tissues. Each tissue enrichment OR can therefore be interpreted as the odds of seeing a TG:HDL ratio analyses P ≤ 1.04 × 10⁻⁷ comparing genes with expression-enriched status in the tissue to those that are not expression enriched in the tissue. For the common-variant tissue enrichment analysis based on autosomal protein-coding genes, we defined the outcome for each tissue as 1/0 if the gene was identified as expression enriched in the tissue. The exposure variable was set to 1 if the gene was identified in the variant to gene prioritization approach and 0 otherwise (see the ‘Variant to gene prioritization’ section below). To minimize the potential confounding, gene length was included as a covariate. The enrichment OR was then estimated using the Firth-corrected logistic model¹¹² to account for cells with zero counts after excluding all primary sexual and reproductive tissues.

Liver RNA-seq

We conducted liver RNA-seq analysis of samples obtained from 1,946 participants enrolled in the Geisinger Health System (GHS) cohort, all of whom underwent perioperative wedge liver biopsies during bariatric surgery.

RNA preparation and sequencing

Total RNA was extracted and processed using the NEBNext Poly(A) mRNA Magnetic Isolation Module and the NEBNext Ultra II Directional RNA Library Prep Kit for Illumina (New England Biolabs), followed by amplification with Kapa HiFi polymerase (Roche) and custom barcoded primers (IDT). Sequencing was performed on the Illumina NovaSeq 6000 platform using S2 flow cells, generating paired-end 75 bp reads. The mean sequencing depth was 72 million reads per sample (median, 68 million), with 93% of samples yielding at least 50 million reads and 99% exceeding 45 million reads, indicating high coverage.

RNA data quality control

RNA-seq data processing was performed according to protocols similar to the GTEx v8 analysis pipeline (https://gtexportal.org/home/documentationPage#staticTextAnalysisMethods). Reads were aligned to the GRCh38/hg38 human reference genome using STAR (v.2.5.3a)¹¹³. Optical duplicate reads were marked using Picard v.2.9.0 with OPTICAL_DUPLICATE_PIXEL_DISTANCE set to 15,000. Gene quantification utilized GENCODE Release 32 annotation, collapsed to a single transcript model per gene. Gene-level expression was quantified with RNA-SeQC (v.1.1.9)¹¹⁴, applying the following read filters: (1) uniquely mapped reads; (2) properly paired reads; (3) alignment distance ≤ 6; and (4) reads fully contained within exon boundaries. For tissue-specific quality control, gene expression values (read counts) were normalized across samples using the TMM method^115,116. Genes were retained for downstream analyses if they met expression thresholds of ≥0.1 TPM in ≥20% of samples and ≥6 unnormalized reads in ≥20% of samples.

eQTL analysis

eQTL analyses followed the GTEx workflow. Covariates included age, sex, the first four PCs derived from common genetic variants and the first 100 PCs from gene expression data to account for potential batch effects. Gene expression levels were transformed using a rank-inverse normal distribution prior to analysis.

eQTL to TG:HDL ratio colocalization

We performed an integrative analysis of TG:HDL ratio fine-mapped sentinel variants and liver eQTLs in the GHS bariatric surgery cohort. By cross-linking 1,617 fine-mapped sentinel variants associated with the TG:HDL ratio to liver eQTL sentinel variants, and including proxies in strong LD (r² > 0.8), we systematically evaluated the potential for shared causal variants. Bayesian statistical co-localization was conducted within 500 kb flanking regions around each matched variant, estimating the posterior probability of a shared causal signal between liver eQTLs and the TG:HDL ratio using the coloc software (v.5.2.3) implemented in R¹¹⁷.

Proteomics and pQTL analysis

For UKB, we used summary-level data from the UKB Plasma Proteomics Project (PPP)¹¹⁸. For GHS, serum samples were obtained for 9,941 individuals and proteomic NPX values were generated by Olink based on the Olink Explore 3072 assay and using the same protocol as used for the UKB PPP¹¹⁸. Analytes were measured across eight different disease specific panels (cardiometabolic I/II, inflammation I/II, neurology I/II and oncology I/II). The samples were processed according to the manufacturer’s protocol with custom automation at the Regeneron Genetics Center. The samples were sequenced on 10B flow cells on the Illumina NovaSeq X platform. The following quality-control steps were applied: PC outliers were removed, putative sample swaps were excluded and data were normalized using NPX Intensity Normalization, which subtracts plate medians per assay. For pQTL analyses, protein levels were rank-inverse normal transformed, then adjusted for age (defined as age at serum collection), sex, age², age × sex and age² × sex. The final pQTL results comprised inverse-variance-weighted meta-analysed data from the UKB PPP project and pQTL data generated from individuals in the GHS cohort.

Variant to gene prioritization

In the TG:HDL-ratio common-variant GWAS, putative causal genes were prioritized with the following criteria: (1) colocalization with GTEx eQTL fine-mapped peaks in any tissue (defined as r² ≥ 0.8); (2) colocalization with pQTL fine-mapped peaks (r² ≥ 0.8); or (3) physical proximity to the nearest gene.

Pathway-enrichment analysis

We used DEPICT (v.1)²⁰ gene-level pathway z-scores as an input. The 59 target genes (mapped to Ensembl IDs) were intersected with the DEPICT gene set, and pathway enrichment was computed through a Stouffer score: z_p = ∑_i z_i,p/√n, with one-sided P values (P = Φ(−z_p)) for enrichment. Multiple-testing control by Bonferroni correction was performed for the total number of pathways tested, yielding P < 5.9 × 10⁻⁶. For reporting, we summarized mean pathway z scores and listed the top contributing genes per pathway (highest positive z scores among the 59 genes).

Mouse models and procedures

C57BL/6NTac mice were from Taconic. Cas9-Ready (v.2.5, 2673) mice, expressing Cas9 transgene under the CAG promoter, were previously reported¹¹⁹. Male mice were housed under a 12 h–12 h light–dark cycle at 22 ± 1 °C or 30 °C (for experiments at thermoneutral conditions), under humidity-controlled conditions (30–70% humidity) in static cages (≤5 mice per cage) with free access to food and water and fed either control chow diet (PicoLab Rodent Diet 20, LabDiet 5053) or HFHFD (Research Diets, D09100310). Mouse experiments were performed on age-matched and strain-matched pairs (littermates). Mice were between 6 and 11 weeks of age at the beginning of the studies. Animal sample size (n) was chosen on the basis of previous experience with similar experiments, experimental feasibility, availability of samples and the number necessary to obtain definitive, significant results. No statistical methods were used to predetermine sample sizes. In all studies, mice were allocated into experimental groups on the basis of matching body weight, body composition and lipid levels. The goal was to balance weight, body composition and lipid levels at the baseline (similar values between groups), to allow evaluation of treatment efficacy after intervention. Investigators were blinded to group allocation during serum chemistry analysis, liver lipid analysis, protein analysis and DNA editing analysis. Blinding was not possible during the in-live part of the mouse studies due to the need to treat mice with different reagents according to experimental designs. All of the animals were monitored for changes in body weight on a weekly basis. No adverse reactions or signs of discomfort were observed during the course of the experiment. Body mass composition was assessed in awake mice using echoMRI. Plasma was collected at various timepoints after AAV administration. The studies involved male mice only, which historically have shown more robust metabolic disease phenotypes compared with female mice^120,121,122. We acknowledge that this experimental design choice may limit the generalizability of mouse model findings to female mice. All animal procedures were conducted in compliance with protocols approved by the Regeneron Pharmaceuticals Institutional Animal Care and Use Committee.

AAV administration

Non-fasted baseline serum chemistry was established on a chow diet, and mice were sorted into treatment groups on the basis of body weight and lipid levels. AAVs encoding shRNAs were of serotype AAV8 designed for liver-targeted mRNA knockdown in mice. AAV-shRNAs were diluted in saline and administered to mice by intravenous injection at a dose of 2.5 × 10¹¹ viral genomes per mouse. Hpn targeting AAV (AAV8-GFP-U6-m-Hpn-shRNA, Vector Biolabs, shAAV-261634) encoded for a 1:1 mixture of 2 shRNAs: shRNA 1, 5′-CCGGGTGGATCTTCAAGGCCATAAACTCGAGTTTATGGCCTTGAAGATCCACTTTTT-3′; shRNA 2, 5′-CCGGCAACGGCACATCGGGCTTCTTCTCGAGAAGAAGCCCGATGTGCCGTTGTTTTT-3′.

Scramble (non-targeting) shRNA (AAV8-GFP-U6-scrmb-shRNA, Vector Biolabs) was used as control AAV, while saline alone was used as a vehicle control.

For studies using gRNAs, AAV8-multi-gRNAs targeting Flcn, Fnip1, Fnip2 or Fnip1 + Fnip2 (combined AAVs) were intravenously administered to recipient mice at 2.5 × 10¹⁰ viral genomes per mouse. Multi-gRNA constructs included five distinct gRNAs targeting the same gene encoded by the same vector, with each gRNA driven by its own U6 promoter. Co-delivery of multiple, distinct gRNAs was previously shown to result in efficient gene perturbations and has been used to dissect molecular pathways in vivo¹²³.

gRNA sequences (name, gRNA sequence, protospacer-adjacent motif): mm_Fnip1_g15, GTAAGTGCTTGTGGATGCAG, GGG; mm_Fnip1_g19, AAGAGGCACTCCTGATCAGG, CGG; mm_Fnip1_g24, GCCAGGAAGAGAACTGAATG, AGG; mm_Fnip1_g25, CTGAATGAGGACAGAGACAG, CGG; mm_Fnip1_g28, AAAGTACCTGAACTCAGTCA, GGG; mm_Fnip2_g2, GCTACAATGGGTAGCTTCTG, TGG; mm_Fnip2_g3, ATTAACCAAGATCCTCAGGC, TGG; mm_Fnip2_g4, GAAGAGCTTGGAGACGGGAA, GGG; mm_Fnip2_g5, TTGTCTGACTTCGGAGCCAG, CGG; mm_Fnip2_g6, TATGTAGTGTATCTTCAGGG, TGG; mm_Flcn_g21, TTACACCAGAGGGTGCTGAA, GGG; mm_Flcn_g25, ATGGTCTGTGGACAACACGG, CGG; mm_Flcn_g26, AGCATGGTCTGTGGACAACA, CGG; mm_Flcn_g28, GTGTCTCACACACTTACCTG, AGG; mm_Flcn_g30, TATGAGTTTGTGGTGACCAG, TGG; mm_Ttr_g26, TTACAGCCACGTCTACAGCA, GGG.

siRNA administration in vivo

Non-fasted baseline serum chemistry was established on chow diet, and mice were sorted into treatment groups on the basis of body weight and body composition. GalNAc-conjugated siRNAs targeting Flcn, or a non-targeting control sequence, were administered through subcutaneous injections at 10 mg per kg every 10 days. GalNAc-conjugated Flcn or control siRNAs were provided by Alnylam Pharmaceuticals.

Amplicon library preparation

gDNA was extracted from liver samples. Target-specific oligos were designed (21–27 bp) to generate a maximum amplicon size of 350 bp with a primer melting temperature (T_m) of 60–65 °C. Barcode adapter sequences were added to the target specific oligo and the full sequence was ordered from Integrated DNA Technologies (IDT). PCR was completed on each gDNA sample. In brief, in each reaction, 4 ng of gDNA was combined with IDT oligos, Q5 polymerase (M0491, New England Biolabs), 10 μM dNTPs, buffer and water according to the manufacturer’s specifications. The amplification products were then diluted 1:100 and used for the PCR barcoding reaction to create the final sequencing library. Each barcoding reaction contained a single amplified target with a forward and reverse primer containing a unique barcode and index. Each plate of PCRs was pooled in equal volumes and then purified in a single tube using AMPure XP reagent (A63881, Beckmann-Coulter) according to the manufacturer’s instructions. The final library concentration was measured using the Qubit fluorometer (Q32866, Invitrogen). Then, 4 nmol of the prepared library was loaded onto the Illumina MiSeq system according to the manufacturer’s instructions using the 2×300 read kit (MS-102-3003, Illumina).

Sequence mapping and characterization

Barcoded samples were demultiplexed to individual reads (FASTQ format). Forward and reverse reads of each FASTQ file were then merged using PEAR (v.0.9.8)¹²⁴. Merged reads were mapped to the Mus musculus genome version 10 (mm10) using Bowtie2 (v.2.5.4)¹²⁵. Each sample was sequenced with a minimum of 20,000 merged reads across the expected guide cleavage location. Finally, characterization of barcoded samples was performed using a custom perl script (available at https://rgc-community.regeneron.com/ on the page dedicated to this manuscript). In brief, all insertions, deletions or base changes within a window of 20 bases upstream and downstream of the expected cut site were considered to be CRISPR-induced modifications. The number of reads containing insertions, deletions or base changes was compared to the number of reads with wild-type sequence to determine the percentage of editing per group.

Liver and lipid analysis

Blood was collected in EDTA tubes and plasma was obtained by centrifugation at 10,000 rpm for 10 min at 4 °C. Circulating total cholesterol, LDL-C, HDL-C, NEFA, albumin, total protein and ALT levels were measured in plasma using the ADVIA Chemistry XPT blood chemistry analyzer (Bayer). Non-HDL-C levels were calculated by subtracting HDL-C from total cholesterol values. To determine liver lipid levels, snap-frozen liver samples were weighed and homogenized in chloroform:methanol (2:1) solution, followed by addition of saline and centrifugation to achieve phase separation. The organic phase (bottom layer) was transferred into a new tube and evaporated with nitrogen gas. The dried lipids were then solubilized with chloroform:Triton X-100 (3:1) solution. Triglyceride and cholesterol content was measured enzymatically (Infinity, Thermo Fisher Scientific) according to the manufacturer’s instructions and normalized to wet tissue weight, as previously described⁴⁷.

Glycaemic control

To assess glycaemic control, an insulin tolerance test was performed. Mice were fasted for 4 h, followed by insulin administration through intraperitoneal injection (HumulinR, Lilly, 0.75 U per kg body weight). The tip of the tail of each mouse was scratched to draw blood. Blood samples were collected at 0, 30, 60, 90 and 120 min, and glucose was measured using the Accu-check blood glucose monitoring system (Roche). For insulin measurements, blood was collected in capillary tubes containing protease inhibitors. Blood was centrifuged at 10,000 rpm for 10 min to separate the plasma and the mouse insulin ELISA kit (Mercodia, 10-1247-01) was used to determine insulin levels.

Quantitative PCR with reverse transcription

Tissue samples were collected in RNAlater (Thermo Fisher Scientific) and frozen. Total RNA was extracted using TRIzol reagent and Direct-zol RNA miniprep kits according to the manufacturer’s instructions (Thermo Fisher Scientific, Zymo Research). gDNA was removed using the DNase I provided in the Direct-zol RNA kit (Zymo Research). mRNA (up to 2 μg) was reverse transcribed into cDNA using SuperScript VILOTM Master Mix (Thermo Fisher Scientific). cDNA was amplified in duplicate reactions containing 5 μl of 2× TaqMan Gene Expression Master Mix (Thermo Fisher Scientific), 0.5 μl TaqMan assay probes (20×, Thermo Fisher Scientific), 3.5 μl nuclease-free H₂O and 1 μl cDNA using the QuantStudio 6 Flex Real-Time PCR System (Thermo Fisher Scientific) and data were analysed using the \({2}^{-\Delta \Delta {C}_{{\rm{t}}}}\) method.

RNA-seq and read mapping

Total RNA was purified from liver (n = 5 mice fed a chow diet for 30 weeks, n = 7 mice fed an HFHFD for 30 weeks (vehicle control group), n = 8 mice treated with Flcn AAV8-gRNA on an HFHFD for 30 weeks). Tissue samples were transferred from RNAlater (Thermo Fisher Scientific, AM7021) to 1–3 ml Trizol reagent (Thermo Fisher Scientific, 15596026). The samples were homogenized on the custom Omni homogenizer (Omni, 51-000-1) at 20,000 rpm for 180 s. The lysates were phase-separated with chloroform and the aqueous phase was purified on the KingFisher Flex (Thermo Fisher Scientific, 5400630) system with the MagMAX-96 for Microarrays Total RNA Isolation Kit (Thermo Fisher Scientific, AM1839) with an additional DNase (Qiagen, 79254) step added between the first and second washes. RNA was quantified on the Lunatic (Unchained Labs, 700-2000) system using a standard Lunatic plate (Unchained Labs, 701-2019), and the integrity was read on the 5300 Fragment Analyzer (Agilent, M5311AA) using an RNA Kit (Agilent, DNF-471-1000) according to the manufacturer’s protocol. mRNA-seq libraries were generated using the KAPA mRNA HyperPrep Kit (Roche Sequencing). Starting material was 500 ng RNA and fragmentation was done at 85 °C for 6 min. cDNA was ligated with 1.5 µM xGen Dual Index UMI Adapters (Integrated DNA Technologies) and amplified using 12 PCR cycles. Sequencing of the resulting libraries was done on NovaSeq 6000 (Illumina) using a 51 cycle, single-end sequencing recipe. Raw sequence data (BCL files) were converted to FASTQ format by Illumina BCL Convert (v.4.3.6). Reads were decoded on the basis of their barcodes, and read quality was evaluated with FastQC (v.0.12.1) (www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were mapped to the mouse genome (GRCm38.REGN) using ArrayStudio software (v.12.5) (OmicSoft) allowing two mismatches. Reads mapped to the exons of a gene were summed at the gene level. Differential gene expression analysis was performed using the DESeq2 (v.1.34.0) package¹²⁶. Benjamini–Hochberg multiple testing correction was applied to the obtained P values. Significant genes were determined using cut-offs of FDR < 0.05 and fold change > 2.

scRNA-seq

Five human single-cell RNA-seq (scRNA-seq) datasets were downloaded from the Gene Expression Omnibus (GEO): GSE136103 (ref. ¹²⁷), GSE168933 (ref. ¹²⁸), GSE115469 (ref. ¹²⁹), GSE158723 (ref. ¹³⁰) and GSE185477 (ref. ¹³¹). Only samples from human liver where scRNA-seq was performed were included from these datasets. For GSE136103 (ref. ¹²⁷), counts tables were used as provided by the authors. For the other four datasets, raw data were processed with 10x Genomics Cell Ranger (v.7). Doublets were detected and removed using Scrublet (v.0.2.3)¹³². Cells with <200 detected genes and/or mitochondrial counts >10% were removed. Counts from all datasets were merged into a single AnnData object and downstream processing was performed using scanpy (v.1.9.1)¹³³. Genes not expressed in any cell in the merged object were filtered out. Counts were normalized to 10,000 counts per cell then log-transformed. The top 5,000 highly variable genes were identified, and principal component analysis was performed using the top 100 PCs. Batch effects across datasets were corrected using the python implementation of Harmony, Harmonypy (v.0.0.10)¹³⁴. A shared nearest-neighbour graph was constructed on the corrected embedding using 20 neighbours and 100 PCs and a uniform manifold approximation and projection (UMAP)¹³⁵ was computed. Leiden clustering was performed at a resolution of 0.1, and clusters were annotated to major cell types using canonical markers. Myeloid and lymphoid compartments were extracted and further subclustered to resolve additional cell types. Mouse liver scRNA data were obtained from a previous study (GSE156052)¹³⁶. Author-provided UMAP coordinates and cell type annotations were used without modification.

Protein detection using liquid chromatography–mass spectrometry

Protein lysates were generated from livers using SDS buffer containing protease inhibitors. After reduction and alkylation, proteins were enzymatically digested into peptides using trypsin/Lys-C. Synthetic stable isotope labelled peptide (¹³C₆,¹⁵N₄-arginine or ¹³C₆,¹⁵N₂-lysine, AQUA QuantProHeavy, Thermo Fisher Scientific) standards for FLCN and FNIP2 were added for validation. Peptides were separated on a C18 column using a 70 min gradient and analysed on the Thermo Scientific Exploris480 Mass Spectrometer using parallel reaction monitoring. Extracted ion chromatograms for target peptides (FLCN, LLEGAPTEDTLVQMEK; FNIP2, GPSSEPVPNR) were generated using Thermo Scientific Freestyle software (v.1.8.65.0) and the area under the curve was used for quantification. Data were normalized to histone H4C1 protein (DNIQGITKPAIR) to account for input differences.

Knockdown in primary human hepatocytes

Cryopreserved primary human hepatocytes were obtained from Thermo Fisher Scientific (donor HU8450, male, HMCPTS). Cells were authenticated by the manufacturer by evaluating phase I enzyme activities, transporter activity and bile canaliculi formation. Cells were not tested for mycoplasma contamination. Cells were thawed and cryopreservation medium was exchanged for plating medium (William’s E medium (A1217601, Thermo Fisher Scientific) supplemented with primary hepatocyte thawing/plating supplements (CM3000, Thermo Fisher Scientific)). Cells were seeded into collagen-precoated 12-well tissue culture plates (356500, Corning) at a density of 600,000 cells per well. After 6 h, when cells had settled and formed a monolayer, the plating medium was exchanged for maintenance medium (William’s E medium supplemented with primary hepatocyte maintenance supplements, serum-free (CM4000, Thermo Fisher Scientific)). Hepatocytes were transfected with 100 nM siRNAs (Silencer Select siRNA: FNIP1 s41304, 4392421; siNeg Silencer siRNA 1, AM4635, Thermo Fisher Scientific) using lipofectamine RNAiMax reagent (13778150, Invitrogen) diluted in OptiMEM medium (51985034, Gibco), according to the manufacturer’s protocol. Cells were collected after 96 h and collected for RNA extraction using the RNeasy Mini Kit (74104, Qiagen), followed by gene expression analysis and immunoblotting.

Immunoblot analysis

Primary human hepatocyte cell pellets were resuspended in 1× RIPA buffer (20-188, EMD Millipore) supplemented with protease inhibitors (PIA32955, Thermo Fisher Scientific). The lysates were centrifuged at 10,000g for 10 min at 4 °C. The supernatant was collected and protein concentrations were determined using the DC assay (5000111, Bio-Rad). Identical amounts (8 μg) of protein were size-fractionated on 4–20% gradient SDS–PAGE gels (5671093, Bio-Rad) under reducing conditions and transferred to PVDF membranes (1704157, Bio-Rad). The membranes were blocked with 5% milk in Tris-buffered saline with 0.1% Tween-20. After blocking, the membranes were probed with primary antibodies (anti-human FNIP1, Cell Signaling, 36892, 1:1,000; anti-HSP90, Cell Signaling, 4877, 1:1,000) and the signal was detected using an enhanced chemiluminescent detection system (WBULS0500, EMD Millipore) and imaged on the Amersham Imager 600 (GE Healthcare Life Sciences). Full scans of the immunoblots are provided in Supplementary Fig. 1.

Statistical analysis and reproducibility

In cell and animal experiments, statistical and graphical data analyses were performed using Microsoft Excel and Prism 10 (GraphPad). Data are expressed as mean ± s.e.m. or mean ± 95% CIs. Mean values were compared using one-way or two-way ANOVA as implemented in GraphPad Prism 10 (GraphPad). P < 0.05 was considered to be significant. All experiments with AAV8-gRNAs (Flcn, Fnip1 and Fnip2) in mice fed high-fat diets for up to 13 weeks, studies in primary human hepatocytes (Fnip1) and studies with AAV-shRNAs (Hpn) were independently performed twice and produced similar results. Studies with long-term (30 weeks) inhibition of Flcn, Fnip1 and Fnip2, as well as studies of mice fed chow, at thermoneutrality or using GalNAc-conjugated siRNAs were performed once, but yielded results that were similar to the AAV8-gRNA studies in mice fed high-fat diets.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

1. Naghavi, M. et al. Global burden of 288 causes of death and life expectancy decomposition in 204 countries and territories and 811 subnational locations, 1990-2021: a systematic analysis for the Global Burden of Disease Study 2021. Lancet 403, 2100–2132 (2024).

   Article  Google Scholar 

2. Bouchard, C. et al. Genetic effect in resting and exercise metabolic rates. Metabolism 38, 364–370 (1989).

   Article  CAS  PubMed  Google Scholar 

3. Bouchard, C. et al. The response to long-term overfeeding in identical twins. N. Engl. J. Med. 322, 1477–1482 (1990).

   Article  CAS  PubMed  Google Scholar 

4. Willer, C. J. et al. Discovery and refinement of loci associated with lipid levels. Nat. Genet. 45, 1274–1283 (2013).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Locke, A. E. et al. Genetic studies of body mass index yield new insights for obesity biology. Nature 518, 197–206 (2015).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

6. Shungin, D. et al. New genetic loci link adipose and insulin biology to body fat distribution. Nature 518, 187–196 (2015).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

7. Lotta, L. A. et al. Integrative genomic analysis implicates limited peripheral adipose storage capacity in the pathogenesis of human insulin resistance. Nat. Genet. 49, 17–26 (2017).

   Article  CAS  PubMed  Google Scholar 

8. Akbari, P. et al. Sequencing of 640,000 exomes identifies GPR75 variants associated with protection from obesity. Science https://doi.org/10.1126/science.abf8683 (2021).

9. Akbari, P. et al. Multiancestry exome sequencing reveals INHBE mutations associated with favorable fat distribution and protection from diabetes. Nat. Commun. 13, 4844 (2022).

   Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

10. Vaduganathan, M., Mensah, G. A., Turco, J. V., Fuster, V. & Roth, G. A. The global burden of cardiovascular diseases and risk: a compass for future health. J. Am. Coll. Cardiol. 80, 2361–2371 (2022).

    Article  PubMed  Google Scholar 

11. Oliveri, A. et al. Comprehensive genetic study of the insulin resistance marker TG:HDL-C in the UK Biobank. Nat. Genet. 56, 212–221 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

12. DeForest, N. et al. Genome-wide discovery and integrative genomic characterization of insulin resistance loci using serum triglycerides to HDL-cholesterol ratio as a proxy. Nat. Commun. 15, 8068 (2024).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

13. Salazar, M. R. et al. Comparison of the abilities of the plasma triglyceride/high-density lipoprotein cholesterol ratio and the metabolic syndrome to identify insulin resistance. Diab. Vasc. Dis. Res. 10, 346–352 (2013).

    Article  PubMed  PubMed Central  Google Scholar 

14. Kosmas, C. E. et al. The triglyceride/high-density lipoprotein cholesterol (TG/HDL-C) ratio as a risk marker for metabolic syndrome and cardiovascular disease. Diagnostics https://doi.org/10.3390/diagnostics13050929 (2023).

15. Abul-Husn, N. S. et al. A protein-truncating HSD17B13 variant and protection from chronic liver disease. N. Engl. J. Med. 378, 1096–1106 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Verweij, N. et al. Germline mutations in CIDEB and protection against liver disease. N. Engl. J. Med. 387, 332–344 (2022).

    Article  CAS  PubMed  Google Scholar 

17. Nag, A. et al. Human genetics uncovers MAP3K15 as an obesity-independent therapeutic target for diabetes. Sci. Adv. 8, eadd5430 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

18. Deaton, A. M. et al. Rare loss of function variants in the hepatokine gene INHBE protect from abdominal obesity. Nat. Commun. 13, 4319 (2022).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

19. Hindy, G. et al. Rare coding variants in 35 genes associate with circulating lipid levels—a multi-ancestry analysis of 170,000 exomes. Am. J. Hum. Genet. 109, 81–96 (2022).

    Article  CAS  PubMed  Google Scholar 

20. Pers, T. H. et al. Biological interpretation of genome-wide association studies using predicted gene functions. Nat. Commun. 6, 5890 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

21. Cohen, J. C., Boerwinkle, E., Mosley, T. H. Jr & Hobbs, H. H. Sequence variations in PCSK9, low LDL, and protection against coronary heart disease. N. Engl. J. Med. 354, 1264–1272 (2006).

    Article  CAS  PubMed  Google Scholar 

22. Musunuru, K. et al. Exome sequencing, ANGPTL3 mutations, and familial combined hypolipidemia. N. Engl. J. Med. 363, 2220–2227 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

23. Dewey, F. E. et al. Genetic and pharmacologic inactivation of ANGPTL3 and cardiovascular disease. N. Engl. J. Med. 377, 211–221 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

24. Pollin, T. I. et al. A null mutation in human APOC3 confers a favorable plasma lipid profile and apparent cardioprotection. Science 322, 1702–1705 (2008).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

25. Crosby, J. et al. Loss-of-function mutations in APOC3, triglycerides, and coronary disease. N. Engl. J. Med. 371, 22–31 (2014).

    Article  PubMed  PubMed Central  Google Scholar 

26. Jorgensen, A. B., Frikke-Schmidt, R., Nordestgaard, B. G. & Tybjaerg-Hansen, A. Loss-of-function mutations in APOC3 and risk of ischemic vascular disease. N. Engl. J. Med. 371, 32–41 (2014).

    Article  PubMed  Google Scholar 

27. Schwartz, G. G. et al. Alirocumab and cardiovascular outcomes after acute coronary syndrome. N. Engl. J. Med. 379, 2097–2107 (2018).

    Article  CAS  PubMed  Google Scholar 

28. Raal, F. J. et al. Evinacumab for homozygous familial hypercholesterolemia. N. Engl. J. Med. 383, 711–720 (2020).

    Article  CAS  PubMed  Google Scholar 

29. Watts, G. F. et al. Plozasiran for managing persistent chylomicronemia and pancreatitis risk. N. Engl. J. Med. 392, 127–137 (2025).

    Article  CAS  PubMed  Google Scholar 

30. Ray, K. K. et al. Two phase 3 trials of inclisiran in patients with elevated LDL cholesterol. N. Engl. J. Med. 382, 1507–1519 (2020).

    Article  CAS  PubMed  Google Scholar 

31. Saettini, F. et al. Absent B cells, agammaglobulinemia, and hypertrophic cardiomyopathy in folliculin-interacting protein 1 deficiency. Blood 137, 493–499 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

32. Moreno-Corona, N., Valagussa, A., Thouenon, R., Fischer, A. & Kracker, S. A case report of folliculin-interacting protein 1 deficiency. J. Clin. Immunol. 43, 1751–1753 (2023).

    Article  PubMed  Google Scholar 

33. Niehues, T. et al. Mutations of the gene FNIP1 associated with a syndromic autosomal recessive immunodeficiency with cardiomyopathy and pre-excitation syndrome. Eur. J. Immunol. 50, 1078–1080 (2020).

    Article  CAS  PubMed  Google Scholar 

34. Roncareggi, S., Iritani, B. M. & Saettini, F. FNIP1 deficiency: pathophysiology and clinical manifestations of a rare syndromic primary immunodeficiency. Curr. Issues Mol. Biol. https://doi.org/10.3390/cimb47040290 (2025).

35. Malik, N. et al. Induction of lysosomal and mitochondrial biogenesis by AMPK phosphorylation of FNIP1. Science 380, eabj5559 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

36. Sattler, E. C. & Steinlein, O. K. in GeneReviews (eds Adam, M. P. et al.) https://www.ncbi.nlm.nih.gov/books/NBK1522/ (Univ. Washington, 1993).

37. Tsun, Z. Y. et al. The folliculin tumor suppressor is a GAP for the RagC/D GTPases that signal amino acid levels to mTORC1. Mol. Cell 52, 495–505 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

38. Lawrence, R. E. et al. Structural mechanism of a Rag GTPase activation checkpoint by the lysosomal folliculin complex. Science 366, 971–977 (2019).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

39. Shen, K. et al. Cryo-EM structure of the human FLCN-FNIP2-Rag-Ragulator complex. Cell 179, 1319–1329 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

40. Napolitano, G. et al. A substrate-specific mTORC1 pathway underlies Birt–Hogg–Dubé syndrome. Nature 585, 597–602 (2020).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

41. Gosis, B. S. et al. Inhibition of nonalcoholic fatty liver disease in mice by selective inhibition of mTORC1. Science 376, eabf8271 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

42. Paquette, M. et al. Loss of hepatic Flcn protects against fibrosis and inflammation by activating autophagy pathways. Sci. Rep. 11, 21268 (2021).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

43. Garrelfs, S. F. et al. Lumasiran, an RNAi therapeutic for primary hyperoxaluria type 1. N. Engl. J. Med. 384, 1216–1226 (2021).

    Article  CAS  PubMed  Google Scholar 

44. Balwani, M. et al. Phase 3 trial of RNAi therapeutic givosiran for acute intermittent porphyria. N. Engl. J. Med. 382, 2289–2301 (2020).

    Article  CAS  PubMed  Google Scholar 

45. Adams, D. et al. Patisiran, an RNAi therapeutic, for hereditary transthyretin amyloidosis. N. Engl. J. Med. 379, 11–21 (2018).

    Article  CAS  PubMed  Google Scholar 

46. Li, J. et al. Myeloid Folliculin balances mTOR activation to maintain innate immunity homeostasis. JCI Insight 5, https://doi.org/10.1172/jci.insight.126939 (2019).

47. Adam, R. C. et al. Activin E-ACVR1C cross talk controls energy storage via suppression of adipose lipolysis in mice. Proc. Natl Acad. Sci. USA 120, e2309967120 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

48. Sugiyama, M. et al. Inhibin betaE (INHBE) is a possible insulin resistance-associated hepatokine identified by comprehensive gene expression analysis in human liver biopsy samples. PLoS ONE 13, e0194798 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

49. Li, S. et al. Hepsin enhances liver metabolism and inhibits adipocyte browning in mice. Proc. Natl Acad. Sci. USA 117, 12359–12367 (2020).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

50. Kirchhofer, D. et al. Hepsin activates pro-hepatocyte growth factor and is inhibited by hepatocyte growth factor activator inhibitor-1B (HAI-1B) and HAI-2. FEBS Lett. 579, 1945–1950 (2005).

    Article  CAS  PubMed  Google Scholar 

51. Custode, B. M. et al. Folliculin depletion results in liver cell damage and cholangiocarcinoma through MiT/TFE activation. Cell Death Differ. 32, 1460–1472 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

52. Ulas, S. et al. Clinical and immunologic features of a patient with homozygous FNIP1 variant. J Pediatr. Hematol. Oncol. 46, e472–e475 (2024).

    Article  CAS  PubMed  Google Scholar 

53. Spivak, I. et al. A novel mutation in FNIP1 associated with a syndromic immunodeficiency and cardiomyopathy. Immunogenetics 77, 2 (2024).

    Article  PubMed  PubMed Central  Google Scholar 

54. Reyes, N. L. et al. Fnip1 regulates skeletal muscle fiber type specification, fatigue resistance, and susceptibility to muscular dystrophy. Proc. Natl Acad. Sci. USA 112, 424–429 (2015).

    Article  CAS  PubMed  ADS  Google Scholar 

55. Xiao, L. et al. AMPK phosphorylation of FNIP1 (S220) controls mitochondrial function and muscle fuel utilization during exercise. Sci. Adv. 10, eadj2752 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

56. Hasumi, H. et al. Regulation of mitochondrial oxidative metabolism by tumor suppressor FLCN. J. Natl Cancer Inst. 104, 1750–1764 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

57. Yan, M. et al. Chronic AMPK activation via loss of FLCN induces functional beige adipose tissue through PGC-1α/ERRα. Genes Dev. 30, 1034–1046 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

58. Wada, S. et al. The tumor suppressor FLCN mediates an alternate mTOR pathway to regulate browning of adipose tissue. Genes Dev. 30, 2551–2564 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

59. Yin, Y. et al. FNIP1 regulates adipocyte browning and systemic glucose homeostasis in mice by shaping intracellular calcium dynamics. J. Exp. Med. https://doi.org/10.1084/jem.20212491 (2022).

60. Sudlow, C. et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. PLoS Med. 12, e1001779 (2015).

    Article  PubMed  PubMed Central  Google Scholar 

61. Carey, D. J. et al. The Geisinger MyCode community health initiative: an electronic health record-linked biobank for precision medicine research. Genet. Med. 18, 906–913 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

62. Tapia-Conyer, R. et al. Cohort profile: the Mexico City Prospective Study. Int. J. Epidemiol. 35, 243–249 (2006).

    Article  PubMed  Google Scholar 

63. Ziyatdinov, A. et al. Genotyping, sequencing and analysis of 140,000 adults from Mexico City. Nature 622, 784–793 (2023).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

64. Olson, J. E. et al. Characteristics and utilisation of the Mayo Clinic Biobank, a clinic-based prospective collection in the USA: cohort profile. BMJ Open 9, e032707 (2019).

    Article  PubMed  PubMed Central  Google Scholar 

65. Chowdhury, R. et al. Cohort profile: the BangladEsh Longitudinal Investigation of Emerging Vascular and nonvascular Events (BELIEVE) cohort study. BMJ Open 15, e088338 (2025).

    Article  PubMed  PubMed Central  Google Scholar 

66. Wiley, L. K. et al. Building a vertically integrated genomic learning health system: the biobank at the Colorado Center for Personalized Medicine. Am. J. Hum. Genet. 111, 11–23 (2024).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

67. Lajonchere, C. et al. An integrated, scalable, electronic video consent process to power precision health research: large, population-based, cohort implementation and scalability study. J. Med. Internet Res. 23, e31121 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

68. Abul-Husn, N. S. & Kenny, E. E. Personalized medicine and the power of electronic health records. Cell 177, 58–69 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

69. Park, J. et al. A genome-first approach to aggregating rare genetic variants in LMNA for association with electronic health record phenotypes. Genet. Med. 22, 102–111 (2020).

    Article  CAS  PubMed  Google Scholar 

70. Kozlitina, J. et al. Exome-wide association study identifies a TM6SF2 variant that confers susceptibility to nonalcoholic fatty liver disease. Nat. Genet. 46, 352–356 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

71. Berglund, G., Elmstahl, S., Janzon, L. & Larsson, S. A. The Malmo Diet and Cancer Study. Design and feasibility. J. Intern. Med. 233, 45–51 (1993).

    Article  CAS  PubMed  Google Scholar 

72. Song, P. et al. Data resource profile: understanding the patterns and determinants of health in South Asians—the South Asia Biobank. Int. J. Epidemiol. 50, 717–718 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

73. Wurtz, P. et al. Quantitative serum nuclear magnetic resonance metabolomics in large-scale epidemiology: a primer on -omic technologies. Am. J. Epidemiol. 186, 1084–1096 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

74. Baigent, C. et al. Efficacy and safety of more intensive lowering of LDL cholesterol: a meta-analysis of data from 170,000 participants in 26 randomised trials. Lancet 376, 1670–1681 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

75. Cannon, C. P. et al. Ezetimibe added to statin therapy after acute coronary syndromes. N. Engl. J. Med. 372, 2387–2397 (2015).

    Article  CAS  PubMed  Google Scholar 

76. Sudhop, T. et al. Inhibition of intestinal cholesterol absorption by ezetimibe in humans. Circulation 106, 1943–1948 (2002).

    Article  CAS  PubMed  Google Scholar 

77. Lloyd-Jones, D. M. et al. 2017 Focused Update of the 2016 ACC Expert Consensus Decision Pathway on the Role of Non-Statin Therapies for LDL-Cholesterol Lowering in the Management of Atherosclerotic Cardiovascular Disease Risk: a report of the American College of Cardiology Task Force on Expert Consensus Decision Pathways. J. Am. Coll. Cardiol. 70, 1785–1822 (2017).

    Article  PubMed  Google Scholar 

78. Birjmohun, R. S., Hutten, B. A., Kastelein, J. J. & Stroes, E. S. Efficacy and safety of high-density lipoprotein cholesterol-increasing compounds: a meta-analysis of randomized controlled trials. J. Am. Coll. Cardiol. 45, 185–197 (2005).

    Article  CAS  PubMed  Google Scholar 

79. Kurki, M. I. et al. FinnGen provides genetic insights from a well-phenotyped isolated population. Nature 613, 508–518 (2023).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

80. Dewey, F. E. et al. Distribution and clinical impact of functional variants in 50,726 whole-exome sequences from the DiscovEHR study. Science https://doi.org/10.1126/science.aaf6814 (2016).

81. Dennis, J. K. et al. Clinical laboratory test-wide association scan of polygenic scores identifies biomarkers of complex disease. Genome Med. 13, 6 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

82. Tobin, M. D., Sheehan, N. A., Scurrah, K. J. & Burton, P. R. Adjusting for treatment effects in studies of quantitative traits: antihypertensive therapy and systolic blood pressure. Stat. Med. 24, 2911–2935 (2005).

    Article  MathSciNet  PubMed  Google Scholar 

83. Kleiner, D. E. et al. Design and validation of a histological scoring system for nonalcoholic fatty liver disease. Hepatology 41, 1313–1321 (2005).

    Article  PubMed  Google Scholar 

84. Dixon, W. T. Simple proton spectroscopic imaging. Radiology 153, 189–194 (1984).

    Article  CAS  PubMed  Google Scholar 

85. Littlejohns, T. J. et al. The UK Biobank imaging enhancement of 100,000 participants: rationale, data collection, management and future directions. Nat. Commun. 11, 2624 (2020).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

86. Herman, J. L. et al. Humans with function-disrupting variants in the myostatin gene (MSTN) have increased skeletal muscle mass and strength, and less adiposity. Nat. Commun. https://doi.org/10.1038/s41467-026-70422-2 (2026).

87. O’Dushlaine, C. et al. Genome-wide association study of liver fat, iron, and extracellular fluid fraction in the UK Biobank. Preprint at medRxiv https://doi.org/10.1101/2021.10.25.21265127 (2021).

88. American Diabetes Association. 2. Classification and Diagnosis of Diabetes: Standards of Medical Care in Diabetes—2021. Diabetes Care 44, S15–S33 (2021).

89. Eastwood, S. V. et al. Algorithms for the capture and adjudication of prevalent and incident diabetes in UK Biobank. PLoS ONE 11, e0162388 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

90. Sun, K. Y. et al. A deep catalogue of protein-coding variation in 983,578 individuals. Nature 631, 583–592 (2024).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

91. Krasheninina, O. et al. Open-source mapping and variant calling for large-scale NGS data from original base-quality scores. Preprint at bioRxiv https://doi.org/10.1101/2020.12.15.356360 (2020).

92. Li, H. & Durbin, R. Fast and accurate short read alignment with Burrows-Wheeler transform. Bioinformatics 25, 1754–1760 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

93. Yun, T. et al. Accurate, scalable cohort variant calls using DeepVariant and GLnexus. Bioinformatics 36, 5582–5589 (2021).

    Article  PubMed  PubMed Central  Google Scholar 

94. Lin, M. F. et al. GLnexus: joint variant calling for large cohort sequencing. Preprint at bioRxiv https://doi.org/10.1101/343970 (2018).

95. Chang, C. C. et al. Second-generation PLINK: rising to the challenge of larger and richer datasets. Gigascience 4, 7 (2015).

    Article  PubMed  PubMed Central  Google Scholar 

96. McLaren, W. et al. The Ensembl variant effect predictor. Genome Biol. 17, 122 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

97. Cunningham, F. et al. Ensembl 2022. Nucleic Acids Res. 50, D988–D995 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

98. Morales, J. et al. A joint NCBI and EMBL-EBI transcript set for clinical genomics and research. Nature 604, 310–315 (2022).

    Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

99. Rodriguez, J. M. et al. APPRIS: annotation of principal and alternative splice isoforms. Nucleic Acids Res. 41, D110–D117 (2013).

    Article  CAS  PubMed  Google Scholar 

100. Coban-Akdemir, Z. et al. Identifying genes whose mutant transcripts cause dominant disease traits by potential gain-of-function alleles. Am. J. Hum. Genet. 103, 171–187 (2018).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

101. Lindeboom, R. G., Supek, F. & Lehner, B. The rules and impact of nonsense-mediated mRNA decay in human cancers. Nat. Genet. 48, 1112–1118 (2016).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

102. Mbatchou, J. et al. Computationally efficient whole-genome regression for quantitative and binary traits. Nat. Genet. 53, 1097–1103 (2021).

     Article  CAS  PubMed  Google Scholar 

103. Taliun, D. et al. Sequencing of 53,831 diverse genomes from the NHLBI TOPMed Program. Nature 590, 290–299 (2021).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

104. Wang, G., Sarkar, A., Carbonetto, P. & Stephens, M. A simple new approach to variable selection in regression, with application to genetic fine mapping. J. R. Stat. Soc. B 82, 1273–1300 (2020).

     Article  MathSciNet  Google Scholar 

105. Gillies, C. E. et al. Variant classification using proteomics-informed large language models increases power of rare variant association studies and enhances target discovery. Genet. Epidemiol. 49, e70023 (2025).

     Article  CAS  PubMed  Google Scholar 

106. Vaser, R., Adusumalli, S., Leng, S. N., Sikic, M. & Ng, P. C. SIFT missense predictions for genomes. Nat. Protoc. 11, 1–9 (2016).

     Article  CAS  PubMed  Google Scholar 

107. Adzhubei, I., Jordan, D. M. & Sunyaev, S. R. Predicting functional effect of human missense mutations using PolyPhen-2. Curr. Protoc. Hum. Genet. https://doi.org/10.1002/0471142905.hg0720s76 (2013).

     Article  PubMed  PubMed Central  Google Scholar 

108. Schwarz, J. M., Rodelsperger, C., Schuelke, M. & Seelow, D. MutationTaster evaluates disease-causing potential of sequence alterations. Nat. Methods 7, 575–576 (2010).

     Article  CAS  PubMed  Google Scholar 

109. Chun, S. & Fay, J. C. Identification of deleterious mutations within three human genomes. Genome Res. 19, 1553–1561 (2009).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

110. Ziyatdinov, A. et al. Joint testing of rare variant burden scores using non-negative least squares. Am J. Hum. Genet. 111, 2139–2149 (2024).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

111. Liu, Y. & Xie, J. Cauchy combination test: a powerful test with analytic p-value calculation under arbitrary dependency structures. J. Am. Stat. Assoc. 115, 393–402 (2020).

     Article  MathSciNet  CAS  PubMed  Google Scholar 

112. Firth, D. Bias reduction of maximum likelihood estimates. Biometrika 80, 27–38 (1993).

     Article  MathSciNet  ADS  Google Scholar 

113. Dobin, A. et al. STAR: ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).

     Article  CAS  PubMed  Google Scholar 

114. DeLuca, D. S. et al. RNA-SeQC: RNA-seq metrics for quality control and process optimization. Bioinformatics 28, 1530–1532 (2012).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

115. Robinson, M. D., McCarthy, D. J. & Smyth, G. K. edgeR: a Bioconductor package for differential expression analysis of digital gene expression data. Bioinformatics 26, 139–140 (2010).

     Article  CAS  PubMed  Google Scholar 

116. Robinson, M. D. & Oshlack, A. A scaling normalization method for differential expression analysis of RNA-seq data. Genome Biol. 11, R25 (2010).

     Article  PubMed  PubMed Central  Google Scholar 

117. Giambartolomei, C. et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 10, e1004383 (2014).

     Article  PubMed  PubMed Central  Google Scholar 

118. Sun, B. B. et al. Plasma proteomic associations with genetics and health in the UK Biobank. Nature 622, 329–338 (2023).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

119. Smagris, E. et al. Divergent role of mitochondrial amidoxime reducing component 1 (MARC1) in human and mouse. PLoS Genet. 20, e1011179 (2024).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

120. Pettersson, U. S., Walden, T. B., Carlsson, P. O., Jansson, L. & Phillipson, M. Female mice are protected against high-fat diet induced metabolic syndrome and increase the regulatory T cell population in adipose tissue. PLoS ONE 7, e46057 (2012).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

121. de Souza, G. O., Wasinski, F. & Donato, J. Jr. Characterization of the metabolic differences between male and female C57BL/6 mice. Life Sci. 301, 120636 (2022).

     Article  PubMed  Google Scholar 

122. Huang, K. P. et al. Sex differences in response to short-term high fat diet in mice. Physiol. Behav. 221, 112894 (2020).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

123. Thurkauf, M. et al. Fast, multiplexable and efficient somatic gene deletions in adult mouse skeletal muscle fibers using AAV-CRISPR/Cas9. Nat. Commun. 14, 6116 (2023).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

124. Zhang, J., Kobert, K., Flouri, T. & Stamatakis, A. PEAR: a fast and accurate Illumina Paired-End reAd mergeR. Bioinformatics 30, 614–620 (2014).

     Article  CAS  PubMed  Google Scholar 

125. Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357–359 (2012).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

126. Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).

     Article  PubMed  PubMed Central  Google Scholar 

127. Ramachandran, P. et al. Resolving the fibrotic niche of human liver cirrhosis at single-cell level. Nature 575, 512–518 (2019).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

128. Buonomo, E. L. et al. Liver stromal cells restrict macrophage maturation and stromal IL-6 limits the differentiation of cirrhosis-linked macrophages. J. Hepatol. 76, 1127–1137 (2022).

     Article  CAS  PubMed  Google Scholar 

129. MacParland, S. A. et al. Single cell RNA sequencing of human liver reveals distinct intrahepatic macrophage populations. Nat. Commun. 9, 4383 (2018).

     Article  PubMed  PubMed Central  ADS  Google Scholar 

130. Payen, V. L. et al. Single-cell RNA sequencing of human liver reveals hepatic stellate cell heterogeneity. JHEP Rep. 3, 100278 (2021).

     Article  PubMed  PubMed Central  Google Scholar 

131. Andrews, T. S. et al. Single-cell, single-nucleus, and spatial RNA sequencing of the human liver identifies cholangiocyte and mesenchymal heterogeneity. Hepatol. Commun. 6, 821–840 (2022).

     Article  CAS  PubMed  Google Scholar 

132. Wolock, S. L., Lopez, R. & Klein, A. M. Scrublet: computational identification of cell doublets in single-cell transcriptomic data. Cell Syst. 8, 281–291 (2019).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

133. Wolf, F. A., Angerer, P. & Theis, F. J. SCANPY: large-scale single-cell gene expression data analysis. Genome Biol. 19, 15 (2018).

     Article  PubMed  PubMed Central  Google Scholar 

134. Korsunsky, I. et al. Fast, sensitive and accurate integration of single-cell data with Harmony. Nat. Methods 16, 1289–1296 (2019).

     Article  CAS  PubMed  PubMed Central  Google Scholar 

135. McInnes, L., Healy, J. & Melville, J. UMAP: uniform manifold approximation and projection for dimension reduction. Preprint at https://arxiv.org/abs/1802.03426 (2020).

136. Su, Q. et al. Single-cell RNA transcriptome landscape of hepatocytes and non-parenchymal cells in healthy and NAFLD mouse liver. iScience 24, 103233 (2021).

     Article  CAS  PubMed  PubMed Central  ADS  Google Scholar 

Download references

Author notes

1. These authors contributed equally: George Hindy, Rene C. Adam

2. These authors jointly supervised this work: Viktoria Gusarova, Luca A. Lotta

Authors and Affiliations

1. Regeneron Genetics Center, Regeneron Pharmaceuticals, Tarrytown, NY, USA

   George Hindy, Olukayode Sosina, David Blair, Joseph Herman, Peter Dornbos, Ernst Mayerhofer, Arthur Gilly, Benjamin Geraghty, Karl Landheer, Liron Ganel, Antoine Baldassari, Chuanyi Zhang, Andrew Deubler, Adolfo Ferrando, Giovanni Coppola, Alan Shuldiner, Katherine Siminovitch, Jason Portnoy, Lyndon Mitnaul, Alison Fenney, Manuel Allen Revez Ferreira, Maya Ghoussaini, Mona Nafde, William Salerno, Lourdes Crane, Christina Beechert, Erin Fuller, Laura M. Cremona, Eugene Kalyuskin, Hang Du, Caitlin Forsythe, Zhenhua Gu, Kristy Guevara, Michael Lattari, Alexander Lopez, Kia Manoochehri, Prathyusha Challa, Manasi Pradhan, Raymond Reynoso, Ricardo Schiavo, Maria Sotiropoulos Padilla, Chenggu Wang, Sarah E. Wolf, Manan Goyal, George Mitra, Sanjay Sreeram, Rouel Lanche, Vrushali Mahajan, Sai Lakshmi Vasireddy, Gisu Eom, Krishna Pawan Punuru, Sujit Gokhale, Benjamin Sultan, Pooja Mule, Mudasar Sarwar, Muhammad Aqeel, Xiaodong Bai, Lance Zhang, Sean O’Keeffe, Razvan Panea, Evan Edelstein, Ayesha Rasool, Evan K. Maxwell, Boris Boutkov, Alexander Gorovits, Ju Guan, Lukas Habegger, Alicia Hawes, Olga Krasheninina, Samantha Zarate, Adam J. Mansfield, Joshua Backman, Kathy Burch, Adrian Campos, Sheila Gaynor, Arkopravo Ghosh, Salvador Romero Martinez, Christopher Gillies, Lauren Gurski, Eric Jorgenson, Tyler Joseph, Michael Kessler, Jack Kosmicki, Priyanka Nakka, Olivier Delaneau, Anthony Marcketta, Joelle Mbatchou, Arden Moscati, Anita Pandit, Jonathan Ross, Carlo Sidore, Eli Stahl, Timothy Thornton, Sailaja Vedantam, Rujin Wang, Kuan-Han Wu, Bin Ye, Blair Zhang, Andrey Ziyatdinov, Yuxin Zou, Jingning Zhang, Kyoko Watanabe, Mira Tang, Frank Wendt, Suying Bao, Kathie Sun, Sean Yu, Aaron Zhang, David Corrigan, Dhruv Shidhaye, Chen Wang, Keyrun Adhikari, Alexander Lachmann, Mark Weiner, Brian Hobbs, Jon Silver, William Palmer, Rita Guerreiro, Amit Joshi, Sarah Graham, Erola Pairo Castineira, Tanima De, Luanluan Sun, Mary Haas, Juan Rodriguez-Flores, Moeen Riaz, Manav Kapoor, Gannie Tzoneva, Momodou W. Jallow, Anna Alkelai, Ariane Ayer, Veera Rajagopal, Sahar Gelfman, Vijay Kumar, Jacqueline Otto, Jose Bras, Silvia Alvarez, Jessie Brown, Hossein Khiabanian, Joana Revez, Kimberly Skead, Valentina Zavala, Jae Soon Sul, Lei Chen, Sam Choi, Amy Damask, Nan Lin, Charles Paulding, Sameer Malhotra, Nadia Rana, Jennifer Rico-Varela, Jaimee Hernandez, Larizbeth Romero, Ashley Paynter, Randi Schwartz, Jody Hankins, Anna Han, Samuel Hart, Ryan Smith, Ann Perez-Beals, Gina Solari, Johannie Rivera-Picart, Michelle Pagan, Sunilbe Siceron, Suganthi Balasubramanian, Marcus B. Jones, Michelle G. LeBlanc, John D. Overton, Jeffrey G. Reid, Goncalo R. Abecasis, Jonathan Marchini, Cristen Willer, Jonas Bovijn, Adam Locke, Aris Baras, Niek Verweij & Luca A. Lotta

2. Regeneron Pharmaceuticals, Tarrytown, NY, USA

   Rene C. Adam, Dwaine Pryce, Joseph Lee, Charleen Hunt, Ivory Mintah, Daphne Sun, Angel Coppola, Kyle Brown, Tram Nguyen, Xin Xia, Andrew J. Murphy, Christos A. Kyratsous, George D. Yancopoulos, Mark W. Sleeman & Viktoria Gusarova

3. Department of Genetics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA, USA

   Daniel J. Rader

4. Department of Clinical Sciences Malmö, Lund University, Malmö, Sweden

   Olle Melander

5. Department of Internal Medicine, Skåne University Hospital, Malmö, Sweden

   Olle Melander

6. Geisinger Obesity Institute, Geisinger Health System, Danville, PA, USA

   Christopher D. Still

7. Faculty of Medicine, National Autonomous University of Mexico, Mexico City, Mexico

   Jaime Berumen, Jesus Alegre-Díaz & Roberto Tapia-Conyer

8. Proyecto OriGen, Tecnológico de Monterrey, Monterrey, Mexico

   Pablo Kuri-Morales

9. Nuffield Department of Population Health, University of Oxford, Oxford, UK

   Jason M. Torres, Jonathan R. Emberson & Rory Collins

10. Geisinger Health System, Danville, PA, USA

    Adam Buchanan, David J. Carey, Christa L. Martin, Michelle Meyer, Kyle Retterer & David Rolston

11. University of Pennsylvania, Philadelphia, PA, USA

    Marylyn D. Ritchie, JoEllen Weaver, Nawar Naseer, Giorgio Sirugo, Afiya Poindexter, Yi-An Ko, Kyle P. Nerz, Jenna Dever, Aidan Harvey, Sydney Linn, Meghan Livingstone, Fred Vadivieso, Stephanie DerOhannessian, Teo Tran, Julia Stephanowski, Salma Santos, Ned Haubein, Joseph Dunn, Anurag Verma, Colleen Morse Kripke, Marjorie Risman, Renae Judy, Colin Wollack, Shefali S. Verma, Scott Damrauer, Yuki Bradford, Scott Dudek & Theodore Drivas

12. University of Colorado Anschutz Medical Campus, Aurora, CO, USA

    Heather D. Anderson, Christina L. Aquilante, Kelsey Arbogast, Ian M. Brooks, Elizabeth E. Burke, Emily M. Casteel, Joanne B. Cole, Curtis R. Coughlin II, Jacob Crawford, Kristy Crooks, Erin Culver, Matthew J. Fisher, Teresa C. Frye, Hunter George, Chris R. Gignoux, Elizabeth K. Gilliland, Casey S. Greene, Emily Hearst, Audrey E. Hendricks, Randi K. Johnson, Shelby Jones, Dave Kao, Gabrielle A. Knortz, Danielle Koffenberger, Santhanagopalan Krishnamoorthy, Lisa Ku, Elizabeth L. Kudron, Rashawnda Lacy, Ethan M. Lange, Joe A. Lesny, Meng Lin, James L. Martin, Adrian M. Stewart, Alaa Radwan, Anna Tanaka, Carolina Sanchez-Wild, Carolyn T. Swartz, Elise L. Shalowitz, Emily B. Todd, Hoda Sherif, Jack Pattee, Jonathan A. Shortt, Katy E. Trinkley, Laura K. Wiley, Neda Rasouli, Nick Rafaels, Nicole L. McDaniel, Nikita Pozdeyev, Sridharan Raghavan & Vendant Vohra

13. Mayo Clinic, Rochester, MN, USA

    James R. Cerhan, Fergus J. Couch, Janet E. Olson, Nicholas B. Larson, Zachary S. Fredericksen, Mine Cicek, Joanna M. Biernacka, Victor M. Karpyak, Prashanthi Vemuri, Vijay K. Ramanan, Mark A. Frye, Sarah A. McLaughlin, Kathryn J. Ruddy, Suzette J. Bielinski, John C. Lieske, W. Michael Hooten, Lisa A. Boardman, Richard B. Kennedy, Andrew D. Badley, Sean C. Dowdy, Jamie N. Bakkum-Gamez, Gretchen E. Glaser, Ping Yang, Celine M. Vachon, Angela Dispenzieri, Samuel O. Antwi, Ann L. Oberg, Kari G. Rabe, Scott H. Kaufmann, Ellen L. Goode, William A. Cliby, J. Eric Ahlskog, James H. Bower, Peter C. Harris, Naveen L. Pereira, Daniel J. Ma, Robert W. Mutter, Jeanette E. Eckel-Passow, Lisa K. Colborn, Andrew J. Danielsen, Jonathan J. Harrington & Jennifer M. Kushwaha

14. Indiana University School of Medicine, Indianapolis, IN, USA

    Naga P. Chalasani, Tae-Hwi L. Schwantes-An & Andrew J. Saykin

15. University of California, Los Angeles, Los Angeles, CA, USA

    Alex Bui, Antonia Petruse, Arash Naeim, Chris Denny, Clara Lajonchere, Clara Magyar, Dan Geschwind, Maryam Ariannejad, Paul Boutros, Paul Spellman, Sarah Dry, Stan Nelson, Albert Duntugan, Bogdan Pasaniuc, Paul Tung, Tim Chang, Yael Berkovich, Ankur Jain, Danielle Martinez, Ghouse Mohammed, Lora Illiev, Michael Broudy, Roni Haas, Shaiful Alam, Taka Yamaguchi & Yash Patel

16. Icahn School of Medicine at Mount Sinai, New York, NY, USA

    Alexander W. Charney

17. University of Texas Southwestern Medical Center, Dallas, TX, USA

    Julia Kozlitina & Jonathan Cohen

18. Nanyang Technological University, Singapore, Singapore

    Marie Loh & John C. Chambers

19. University of Colombo, Colombo, Sri Lanka

    Prasad Katulanda

20. Devki Devi Foundation, New Delhi, India

    Vinitaa Jha

21. University of Kelaniya, Ragama, Sri Lanka

    Anuradhani Kasturiratne

22. Services Institute of Medical Sciences, Lahore, Pakistan

    Khadija Irfan Khawaja

23. BRAC University, Dhaka, Bangladesh

    Malay Kanti Mridha

24. Madras Diabetes Research Foundation, Chennai, India

    Ranjit Mohan Anjana

25. Imperial College London, London, UK

    John C. Chambers

Consortia

Contributions

Conceptualization of the project, study design, interpretation of results and drafting of the manuscript and its revisions: G.H., R.C.A., V.G. and L.A.L. Project coordination: G.H., R.C.A., V.G., L.A.L., M.B.J. and M.G.L. Human genetics analysis: G.H., O.S., D.B., J.H., P.D., E.M., A.G., B.G., K.L., L.G., A. Baldassari, C.Z., S.B., J. Bovijn, A.L., N.V. and L.A.L. In vitro and in vivo experiments: R.C.A., D.P., J.L., C.H., I.M., D.S., A.C., K.B., T.N., X.X. and V.G. Sample collection, data generation, results interpretation and provision of critical information for the manuscript finalization: D.J.R., O.M., C.D.S., J. Berumen, P.K.-M., J.A.-D., J.M.T., J.R.E., R.C. and R.T.-C. Human genetics sequencing and analytical infrastructure: J.D.O., J.G.R., G.R.A., J.M., C.W., A. Baras and L.A.L. Experimental programs infrastructure: A.J.M., C.A.K., G.D.Y. and M.W.S. All of the authors reviewed the manuscript for important intellectual content and approved the final version for submission. V.G. and L.A.L. jointly supervised this work. The contributions of the consortium members were as follows. The members of the Regeneron Genetics Center contributed to exome sequencing, sequence-variant calling, sample handling, genomic data generation, refined phenotype generation, data formatting, processing and quality control, development of analytical infrastructure, review, quality control and visualization of the data and results contributing the manuscript, research program management, coordination and administrative support, review of the manuscript and provision of necessary information for the finalization of the manuscript all of which made this work possible. Members of the GHS-RGC DiscovEHR Collaboration, Penn Medicine Biobank, Colorado Center for Personalized Medicine, Mayo Clinic-RGC Project Generation, Mexico City Prospective Study, Malmö Diet and Cancer Study, Indiana University School of Medicine (Indiana-CLDB), UCLA-RGC ATLAS Collaboration, Mount Sinai Million Health Discoveries Program, UTSW-Dallas Biobank and South Asia Biobank Study performed cohort design, participant consenting and recruitment, phenotype data collection and quality control, sample collection, cohort data generation, quality control and interpretation, review of the manuscript and provision of critical information necessary for the finalization of the manuscript.

Corresponding authors

Correspondence to Viktoria Gusarova or Luca A. Lotta.

Competing interests

Regeneron and Regeneron Genetics Center affiliated authors receive salaries from and own options and/or stock of the company. G.D.Y. is the chief scientific officer and member of the board of directors at Regeneron Pharmaceuticals. A.J.M. is an executive officer of Regeneron Pharmaceuticals. G.H., R.C.A., M.W.S., V.G. and L.A.L. are listed as inventors on pending patent applications relating to FNIP1 and FLCN genetics (US-2025-0129366, WO/2025/049558, PCT/US2025/016730). J.R.E. reports grants to the University of Oxford from AstraZeneca and Regeneron Pharmaceuticals. The other authors declare no competing interests.

Peer review information

Nature thanks David Carling, who co-reviewed with Eneko Pascual-Navarro; Satoshi Koyama; and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions