Exome-wide association study of blood lipids in 1,158,017 individuals from diverse populations

Abstract

Rare coding alleles have crucial roles in the molecular diagnosis of genetic diseases. However, the systematic identification of these alleles has been challenging due to their scarcity in the general population. Here we discovered and characterized rare coding alleles contributing to genetic dyslipidemia, a principal risk for coronary artery disease (CAD), among 1,158,017 multi-ancestral individuals. Testing 2,997,401 rare coding variants, we identified 800 exome-wide significant associations (176 predicted loss of function (pLoF) and 624 missense variants). Associated alleles are enriched in functional variant classes, show significant additive and recessive associations, exhibit similar effects across populations and resolve pathogenicity for variants of unknown significance. Furthermore, we identified five lipid-associated genes associated with CAD. Among them, silencing RORC represents a potential therapeutic target for lowering low-density lipoprotein cholesterol. This study provides resources and insights for understanding causal mechanisms, quantifying the expressivity of rare coding alleles and identifying new drug targets for dyslipidemia across diverse populations.

Multi-ancestry exome-wide analyses in the Million Veteran Program, UK Biobank and All of Us identify rare coding variants across 209 genes associated with blood lipid traits, highlighting new candidate drug targets across diverse populations.

Main

Family-based discovery and characterization of rare coding alleles causative of familial hypercholesterolemia (FH) have yielded important insights for coronary artery disease (CAD), the leading cause of premature mortality among adults^1,2. While FH is associated with a heightened risk for early-onset CAD, early intervention using lipid-lowering medications can considerably mitigate this risk, suppressing cumulative exposure to continuously high levels of low-density lipoprotein cholesterol (LDLC)^3,4. However, FH remains substantially underdiagnosed and undertreated^4–7. This highlights the need for increased efforts to systematically catalog and characterize pathogenic variants associated with FH for more accurate genetic diagnosis. In addition, like other Mendelian conditions, population-based genetic analyses have often shown that expressivity (continuous effects on lipid levels) and penetrance (likelihood of CAD) may not be sufficiently high for some previously implicated pathogenic variants relative to initial descriptions in family-based studies^8–13. As rare Mendelian alleles are increasingly returned to asymptomatic individuals through screening or secondary reporting^14–16, allele-specific prognosis is increasingly important.

Furthermore, clinically curated variants are enriched among individuals who are genetically similar to European reference populations, reflecting biases in accumulated knowledge and data. In contrast, variants associated with non-European reference populations are more likely to be reclassified, susceptible to population-related biased filters¹⁷ and underdiagnosed due to limited data availability¹⁸.

To address these challenges, we assembled a large-scale, finely imputed/sequenced dataset encompassing lipid measures from over a million individuals combining Million Veteran Program (MVP)¹⁹, UK Biobank (UKB)²⁰ and All of Us Research Program (AOU) cohorts, which included more than 230,000 individuals who are genetically similar to non-European reference populations, which are also historically under-represented in genomic research. This diverse dataset allowed us to identify and characterize rare coding variant associations with blood lipids (total cholesterol (TC), LDLC, high-density lipoprotein cholesterol (HDLC) and triglycerides (TG)) and to validate their generalizability across populations. The summary of the estimated effects provides a resource for further functional assessment and clinical utility.

Results

Study population

We generated a large-scale clinical genetic dataset by imputing MVP (634,535 individuals) to TOPMed imputation reference panel (version r2)²¹, which includes 308,107,085 variants from 97,256 individuals representing diverse populations. Combined with whole-exome sequence (WES) data in UKB (431,178 individuals)²² and whole-genome sequence (WGS) data in AOU (92,304 individuals)²³, we generated a cohort of 1,158,017 individuals, including 238,243 (20.57%) from non-European populations (Fig. 1a and Supplementary Table 1). The large-scale imputation reference panel, including diverse populations, allowed to impute rare variants with high accuracy comparable to sequenced data (Extended Data Fig. 1a–c and Supplementary Note ‘Imputation accuracy of rare variants in the Million Veteran Program’).

Fig. 1. Exome-wide association study for blood lipids in over 1 million individuals.

a,b, Overview of the study. The number of individuals included in the analysis by study (a) and by population (b). c, Correlation between the number of individuals and identified variants in the target region. The horizontal axis shows the number of individuals in each population by study. The vertical axis shows the number of variants identified in the corresponding population. The size of point is proportional to the number of individuals. d, Distribution of effect sizes for exome-wide significant associations is shown. Each dot represents a variant–trait pair with significant association in this study (Methods). All four blood lipids are plotted. The horizontal axis indicates the MAF, while the vertical axis displays the effect size for each allele from the regression model (β), with the unit of effect size normalized to the s.d. of blood lipids. The lines represent the statistical power of 80% at sample sizes of 1 million (dark gray), 500,000 (medium gray) and 100,000 (light gray) individuals. e, The bar charts show the number of variants associated with different phenotypes and variant classes. f, The bar charts show the number of genes associated with blood lipid phenotypes. g, Direction of the effects for associated variants. Variants positively associated with the blood lipids are displayed on the positive side of the vertical axis. The height of each bar represents the number of variants in that category. Bar colors indicate variant classes, with blue for missense variants and red for pLoF variants.

Extended Data Fig. 1. The imputation quality, allelic diversity, coverage, and power in the study.

a, Imputation accuracy in MVP whole-genome imputation data by TOPMed imputation reference panel. Each dot indicates mean R² (squared Pearson’s correlation coefficient), TPR, and FPR by population and MAC/MAF bins. TPR and FPR were computed by comparing dichotomized hard-called dosage (imputed data) and dichotomized sequenced genotype (WGS data; Supplementary Note—Imputation accuracy of rare variants in the Million Veteran Program). b,c, Shared and unique variants across MVP, UKB, and AOU for pLoF (b) and missense (c) variants. The central matrices define the variant sharing status between MVP, UKB, and AOU. The top panel quantifies the variants within the groups defined in the central matrices. The bottom panel summarizes the count of variants in each study. d, Variant coverage. The relative proportions of SNVs identified in this study are shown as a fraction of all possible SNVs within the target transcripts. e, Simulated power curves for different sample sizes. The horizontal axis indicates minor allele frequency, and the vertical axis indicates effect size. The dark blue line indicates 80% power curve at 1 million sample size, the intermediate curve indicates 500,000 sample size, and the gray curve indicates 100,000 sample size, respectively. f, Gene-based power estimation. The color of the bar charts indicates the highest power of the coding variant in the gene. The top panel shows pLoF variants, and the bottom panel shows missense variants. β indicates simulated effect size. TPR, true positive rate; FPR, false positive rate; MAC, minor allele count; MAF, minor allele frequency; WGS, whole-genome sequence; MVP, Million Veteran Program; UKB, UK Biobank; TOPMed, Trans-Omics for Precision Medicine; AFR, African-like population; AMR, Admixed-American-like population; ASN, Asian-like population; EAS, East-Asian-like population; EUR, European-like population; HIS, Hispanic-like population; SAS, South-Asian-like population; pLoF, predicted loss-of-function.

Variant identification

We curated the variants with minor allele count (MAC) ≥5 detected in ±50 bp of exome target regions used in UKB-WES (Methods). Annotation using 19,603 protein-coding transcripts identified 214,000 predicted loss of function (pLoF; stop gain, frameshift insertion/deletion and canonical splice site) and 2,766,489 missense variants (missense single nucleotide variant (SNV) and in-frame insertion/deletion; Supplementary Table 2). These variants covered 1.73% of all possible pLoF SNVs and 3.81% of all possible missense SNVs (Extended Data Fig. 1d, Supplementary Table 3 and Supplementary Note ‘Variant coverage’).

In addition, using the SpliceAI algorithm²⁴, we detected 23,523 putatively cryptic splice variants (variants associated with donor/acceptor gain, donor/acceptor loss in distant position from canonical splice site with Delta Score (DS) of >0.8; Methods). We reclassified these cryptic splice variants as pLoF and included them in association analyses. In total, we identified 237,523 pLoF variants (214,000 canonical pLoF + 23,523 cryptic splice) and 2,759,878 missense variants (2,766,489 canonical missense − 6,611 cryptic splice) in this study (Supplementary Table 4). The MVP and AOU study populations, including diverse populations, effectively increase the diversity of variants included in this study (Extended Data Fig. 1b,c).

We identified at least one testable pLoF in 89.20% (17,486/19,603) of assessed transcripts and a missense variant in 95.30% (18,682/19,603). Among them, 63.37% (12,422/19,603) and 94.68% (18,560/19,603) transcripts had at least one pLoF or missense variant with 80% statistical power to detect the effect size of one s.d. of phenotypes per allele (Extended Data Fig. 1e,f, Supplementary Table 5 and Supplementary Note ‘Power calculation’).

Association analysis

We tested linear associations of the imputed/sequenced genotypes of rare (5 ≤ MAC and the highest minor allele frequency (MAF) across studied populations (MAF_POPMAX) < 1%) pLoF variants or missense variants with blood lipids (TC, LDLC, HDLC and TG) using an additive model stratified by population groups (four from MVP, five from UKB and five from AOU; Fig. 1a–c and Supplementary Table 1; Methods) followed by fixed effects meta-analysis including the 14 population groups. We observed a linear relationship between the number of variants tested and the number of participants across studies and population groups, irrespective of genotyping strategy (Fig. 1c). In total, we tested 10,656,739 variant–phenotype combinations in the additive model. The highest lambda genomic control in four tested traits was 1.025 for HDLC, indicating suitable calibration (Extended Data Fig. 2a). In addition, we conducted recessive model analysis for 233,971 variant–phenotype combinations with 5≤ minor homozygote counts and minor homozygote frequency <1%. Exome-wide significance was defined as P < 4.5 × 10⁻⁹ (0.05/(10,656,739 + 233,971)).

Extended Data Fig. 2. Exome-wide association analysis over a million individuals.

a, Quantile-quantile plots. Top, quantile-quantile plots for four tested lipid traits. Each dot indicates a tested variant. The horizontal axes show the expected negative log₁₀ P-values, and the vertical axes show the observed negative log₁₀ P-values. Colors indicate variant annotation. Dotted lines show expected distribution. Bottom panel focused variants with expected P < 0.01. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. b, 184 exome-wide significant loci. The horizontal axis shows genomic coordinates, and the vertical axis shows the negative log₁₀ P-values. The red triangles indicate pLoF variants, and blue indicate missense variants. The upward triangles indicate trait-increasing associations, and downward triangles indicate trait-decreasing associations. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. c, Penetrant association of APOB p.M3438X. The curve indicates LDLC distribution of the European-like population in the UKB (n = 409,046). The red triangles indicate LDLC level of the carriers of APOB p.M3438X. d, Replication evidence in the independent study for associated variants. Dots indicate rare coding genetic variants that are significantly associated with blood lipids in this study. The horizontal axes display its effect sizes from this study (discovery, n_MAX = 1,057,837), while the vertical axes present its effect sizes from the previous exome-array study²⁵ (replication, n_MAX = 358,251). The error bars represent the 95% confidence intervals in each study. The color of each dot indicates the variant status: (i) dark red for variants with 80% statistical power to detect association and a minor allele count of ≥50 in the replication study; (ii) light red for variants with 80% statistical power to detect association and a minor allele count of <50 in the replication study; and (iii) gray for variants without 80% statistical power. TC, total cholesterol; LDLC, low-density lipoprotein cholesterol; HDLC, high-density lipoprotein cholesterol; TG, triglycerides; GC; genomic control; pLoF, predicted loss-of-function; chr, chromosome.

We identified 800 additive exome-wide significant (EWS) associations in 184 loci (202 associations in 45 loci for TC, 235 associations in 48 loci for LDLC, 222 associations in 47 loci for HDLC and 141 associations in 44 loci for TG; Fig. 1d, Extended Data Fig. 2b and Supplementary Table 6). Among the 209 genes identified by the additive model, 40 genes harbored 176 individual pLoF variant associations, and 193 genes harbored 624 individual missense variant associations (Fig. 1e,f), often with multiple associations per gene (Fig. 1g). Among these, ten gene–trait pairs (GYS2 and TC, LDLC and HDLC, STS and HDLC, SH3TC1 and TC, ETV6 and TC, PCSK6 and TC, PCSK9 and HDLC, POR and HDLC, and PTPRB and TG) were not overlapping with previously reported loci associated with blood lipids. Furthermore, recessive modeling identified 109 EWS associations across 53 genes (Supplementary Table 7).

We observed significant enrichment of EWS variants in pLoF or missense variants compared to synonymous/noncoding variants (odds ratio (OR)_EWS/non-EWS = 6.33, 95% confidence interval (CI) = 5.02–7.91, P = 6.0 × 10⁻⁴¹ for pLoF variants, and 2.27 (1.98–2.60), P = 8.9 × 10⁻³² for missense variants). One of the strongest signals was the APOB pLoF variant (p.M3438X), which reduced LDLC levels by 3.14 s.d. per allele (mean LDLC was 57.9 mg dl⁻¹ for 5 carriers and 145 mg dl⁻¹ for 409,041 noncarriers in European-like population (EUR) in UKB; Extended Data Fig. 2c).

We compared our findings with prior exome-array-based rare variant analyses²⁵ and observed concordant results across studies (Extended Data Fig. 2d and Supplementary Note Replication analysis). Using linkage disequilibrium-independent rare coding variants associated with EWS, we estimated the phenotype variance explained (PVE) for each trait. Collectively, rare coding variations contributed to additional 2.03–3.75% PVE in blood lipids compared to 15.8–22.1% PVE by common variants (Extended Data Fig. 3, Supplementary Table 8 and Supplementary Note ‘Phenotype variance explained by rare coding variants’).

Extended Data Fig. 3. Contribution of rare coding variants to trait variance.

a, Phenotypic variance explained (PVE) by common and rare variants. The height of the bar chart indicates the PVE by GWAS lead variant (gray) and the sum of rare coding variants in the locus (red). PVE is computed by the formula 2f(1 − f)β², where f is the allele frequency and β is the effect size. b, PVE by individual variants. Gray dots indicate common³⁹ and red dots indicate rare (this study) variants. Numbers in the plot indicate the number of variants included in the analysis. Boxes show the interquartile range (IQR) of PVE; the centerline indicates the median. Whiskers extend to the most extreme values within 1.5× IQR of the quartiles. c, Trait variance by rare coding variant and common genetic signals. The horizontal axis indicates PVE by lead variant in the GWAS loci. The vertical axis indicates the sum of PVEs by rare coding variants in the locus. d, The cumulative contribution of lead and rare coding variants for trait variance. PVE by each rare variant in representative genes. Lead variant in the locus is in gray, the sum of PVEs by pLoF in red and missense in dark blue. e, Cumulative PVE_rare by genes. The height of the bars indicates the cumulative PVE from rare variants, ordered from genes with the highest to the lowest per-gene PVE. Genes comprising the minimal set that accounts for 75% of the total PVE_Rare in the phenotype are highlighted in yellow. PVE, phenotypic variance explained; GWAS, genome-wide association study; TC, total cholesterol; high-density lipoprotein cholesterol; LDLC, low-density lipoprotein cholesterol; TG, triglycerides; pLoF, predicted loss-of-function.

Variant function predicts phenotype expressivity

To gain further insights into genetic associations and variant functions, we used existing in silico methods for predicting variant functionality. For pLoF variants, we used the LOFTEE²⁶ plugin in VEP²⁷ and identified 163,643 ‘high-confidence’ (HC) pLoF variants (87.7% pLoF variants). For missense variants, we applied 29 in silico deleterious prediction algorithms²⁸, from which we derived an ensembled Missense Score (MiS; Methods) and grouped them into bins ([0, 0.5], (0.5, 0.7], (0.7, 0.9], and (0.9, 1]), where deleteriousness increases with increasing value (Supplementary Tables 9 and 10). We observed strong linear relationships across variant deleteriousness, lower allele frequencies and phenotype association (Fig. 2a and Supplementary Table 11). Notably, HC pLoF and deleterious missense variants with MiS (0.9, 1.0) and (0.7, 0.9) exhibited similarly constrained low MAF (median = 0.0023%, 0.0021% and 0.0024%, respectively) and were more likely to be EWS (OR_EWS/non-EWS = 7.24 (95% CI = 5.66–91.8) and P = 5.2 × 10⁻⁴⁰ for pLoF; OR = 11.61 (7.02–18.15) and P = 6.2 × 10⁻¹⁵ for MiS (0.9, 1.0); OR = 5.02 (3.87–6.43) and P = 1.9 × 10⁻²⁶ for MiS (0.7, 0.9)). Furthermore, the cryptic splice variants exhibited a similar level of constraint (median MAF = 0.0028%) and were equally enriched for EWS variants (OR_EWS/non-EWS = 5.96 (95% CI = 2.71–11.42), P = 3.1 × 10⁻⁵). As an illustration of the importance of functional context, we tested 53 pLoF variants observed in APOB, 19 of which showed EWS. However, these EWS variants were markedly depleted in the last exon, and pLoF variants located there were predicted to be ‘low-confidence’ pLoF (Fig. 2b). This finding is consistent with prior evidence that protein-truncating events near the 3′ end of a gene body have minimal impact on gene function^26,29.

Fig. 2. Different expressivity of rare coding variants by variant classes.

a, Variant deleteriousness, constraints and statistical associations. The images represent variant classes as pLoF (red), missense (blue) and synonymous/noncoding (gray, used as reference). The ranges associated with the blue points depict the MiS for missense variants. We computed the MiS for missense SNVs by using 29 in silico deleteriousness prediction algorithms. The score was calculated as the number of deleterious predictions divided by the number of available algorithms for each variant, with values ranging from 0 to 1 (Methods). Based on the MiSs, missense variants were grouped into bins. pLoF variants were grouped by LOFTEE predictions. The horizontal axis indicates the median MAF for each variant class, while the vertical axis shows the ORs of EWS to non-EWS variants in reference to synonymous/noncoding variants. ORs were estimated by Fisher’s exact test. Circle size corresponds to the number of variants achieving exome-wide significance in each variant class. The dashed curve is the estimated line, and the shaded area is its 95% CI. b, Penetrance of pLoF variants in APOB. Gray rectangles represent the APOB gene model. Dots correspond to pLoF variants tested in this study, with dot size denoting effect allele frequency, and color indicating variant class (HC pLoF, LC pLoF and cryptic splice). The x axis represents genomic coordinates on hg38, while the y axis shows z values (β divided by s.e.) for LDLC associations calculated using a linear model implemented by REGENIE (Methods). Dots in the lower part indicate significant associations with negative effect sizes for LDLC. Exome-wide significance is indicated by a dotted line. No variants tested showed significant positive associations with LDLC. c, Different distributions of MiSs (see above) observed in hypermorphic and hypomorphic variants. The box plot displays the distribution of MiSs for missense variants within genes that have at least one EWS association by pLoF. A hypomorphic variant is defined as having the same directional association with EWS pLoF association. The P values were calculated by a two-sided Wilcoxon rank-sum test. The P values were not adjusted for multiple testing correction. Conversely, a hypermorphic variant is defined as having an opposite directional association to EWS. Boxes show the IQR of MiSs, and the centerline indicates the median. Whiskers extend to the most extreme values within 1.5× IQR of the quartiles. LC, low confidence; IQR, interquartile range.

Distinguishing hypomorphic and hypermorphic variants

Multiple EWS pLoF associations allowed to assess the effect directions of genetic deficiency in 23 gene–phenotype pairs (Fig. 1g). These included 128 pLoF variant–phenotype pairs, and all exhibited consistent effect directions except for one intronic variant with high cryptic splice potential (DS for donor gain of 0.82) in CETP that decreased HDLC levels (rs182237338; Supplementary Note ‘A cryptic splice variant in CETP with discordant effect direction with canonical pLoF variant’). A total of 87% (239/275) missense variants showed concordant effect directions with pLoF variants in the same genes (hypomorphic variants). However, 36 associations in ten genes were found to have opposite effect direction to pLoF variants (hypermorphic variants; Supplementary Table 6). Some previously discovered hypermorphic variants included PCSK9 (p.R469W, p.R496W³⁰) and APOB (p.R3527Q), but most are newly discovered. One such example is LDLR p.S849L that showed a strong negative association with LDLC (β = −1.07 (s.e. = 0.087), P = 3.6 × 10⁻³⁴), indicating gain of function. Another example is APOB p.G4395S, which has only been observed in African-like populations (AFR) in our study and other databases³¹ and consistently shows a positive association with LDLC (β_AOU-AFR = 0.024 (s.e. = 0.303), β_MVP-AFR = 0.433 (0.080), β_UKB-AFR = 0.775 (0.253)). While MiS was an important factor in predicting hypomorphic associations, it did not predict hypermorphic associations (Fig. 2c).

Cryptic splicing variants as new candidates for loss of function

For all identified variants in this study, we predicted the variant’s potential for splice site disruption/creation using SpliceAI²⁴ and derived a DS—a numeric score ranging from 0 to 1 (Extended Data Fig. 4 and Supplementary Note ‘Cryptic splice annotation’). The score distribution was sparse, and only 0.598% (58,402/9,399,797) variants had high DS (>0.8); 43.5% variants with high DS were not located in the canonical splice sites (Extended Data Fig. 4a). We observed a strong enrichment of cryptic splice variants disrupting the donor structure (donor loss) in the splice donor 5th base (Extended Data Fig. 4b), which are not typically considered as pLoF in the current practice. One representative example was rs200831171—a splice donor 5th base variant of APOA5 and associated with higher TG concentrations. This intronic variant has high donor loss potential (DS_{Donor Loss} = 0.97) and was associated with increased TG levels with the largest effect size (β = 1.10 (s.e. = 0.079), P = 7.0 × 10⁻⁴⁴) among six EWS coding variants in APOA5 associated with TG (Extended Data Fig. 4c). While the splice donor 5th base variant is known for its functional impact, variants at this same position with high DS were significantly enriched for associations with blood lipids compared with those with low DS (OR = 9.88 (95% CI = 1.13–118.4), P = 0.0186). Including this variant, we identified 15 EWS cryptic splice variants (Supplementary Table 6). Overall, cryptic splicing variants showed equivalent effect sizes with pLoF variants (median β_{cryptic splice} = 1.092, interquartile range = 0.601–1.118, normalized to pLoF as 1, P = 0.71 by Wilcoxon rank-sum test; Extended Data Fig. 4d), and larger effect sizes than missense variants (median β_missense = 0.408 (0.136–0.701), P = 7.0 × 10⁻⁴).

Extended Data Fig. 4. Cryptic splice variants affect human blood lipids.

a, Distribution of cryptic splice variants across canonical variant classes. The bar graphs illustrate the proportion of cryptic splice variants within the canonical annotations, with the colors of the bars indicating the delta score (DS). b, Distribution of cryptic splice variants around exon-intron boundary. The histogram shows the positions of cryptic splicing variants (DS > 0.8) in relation to the exon-intron boundary. Exons are represented by blue rectangles. c, Strong expressivity of APOA5 cryptic splice variant. Each dot represents the variant’s effect size estimated by the linear model. The error bar indicates 95% confidence interval of effect size. The unit of effect size is the standard error of blood triglycerides. Red dots indicate pLoF variants, and blue dots indicate missense variants. The number of samples ranges from 444,840 to 1,093,214. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. d, Strong expressivity of cryptic splice variants. The horizontal axis shows the normalized effect sizes for pLoF, pLoF (cryptic splice) and missense variants. The analysis was restricted to genes harboring pLoF, cryptic splice, and missense variants. Boxes show the interquartile range (IQR) of normalized effect sizes, and the centerline indicates the median. Whiskers extend to the most extreme values within 1.5× IQR of the quartiles, and points beyond are plotted as outliers. Two-sided P-values were obtained from the Wilcoxon rank-sum test and were not adjusted for multiple comparisons. UTR, untranslated region; pLoF, predicted loss-of-function; TG, triglycerides.

New rare variant association outside of established lipid loci

We identified associations for 11 variants residing outside established lipid loci (Extended Data Fig. 5 and Supplementary Table 6). One example is a rare missense variant GYS2 p.Y636H (MAF = 0.0431%), which showed significant associations with decreased TC, LDLC and HDLC (β_TC = −0.24, P_TC = 2.3 × 10⁻¹⁵; β_LDLC = −0.19, P_LDLC = 4.0 × 10⁻¹⁰ and β_HDLC = −0.22, P_HDLC = 7.1 × 10⁻¹⁴). GYS2 encodes glycogen synthetase 2, is expressed in the liver³² and is a causal gene for glycogen storage diseases. Another example is the STS gene on the X chromosome. A rare missense variant (p.H439R) in this gene was associated with decreased HDLC. STS encodes steroid sulfatase, which is directly involved in steroid metabolism. Other new loci identified by this study include SH3TC1 (TC), ETV6 (TC), PCSK6 (TC), PCSK9 (HDLC), POR (HDLC) and PTPRB (TG).

Extended Data Fig. 5. New loci driven by rare coding variants.

a–j, The horizontal axes represent genomic coordinates, while the vertical axes denote the negative log₁₀ P-values. Red dots illustrate the association of rare coding variants in genes with significant variants. In contrast, blue dots show the association of rare coding variants in genes without significant variants. Gray dots represent common variant associations from a previous study³⁹. The dashed line in the top panel indicates the exome-wide significance threshold (P < 4.5 × 10⁻⁹). The bottom panel illustrates the coding genes within the locus; genes harboring significant variants are highlighted in red, and others are in blue. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. TC, total cholesterol; LDLC, low-density lipoprotein cholesterol; HDLC, high-density lipoprotein cholesterol; TG, triglycerides.

New insights into causal genes within established lipid loci

Lead variants in genome-wide association studies (GWASs) are typically common and noncoding, making the causal gene unclear. Rare variant association studies more directly interrogate perturbations of gene products, providing greater confidence in causal gene inference. One such example is 1q21.1, an established HDLC GWAS locus comprising 21 genes (Extended Data Fig. 6a). A rare pLoF variant in only PDZK1 at 1q21.1 was associated with increased HDLC levels (MAF = 0.016%, β = 0.30, P = 1.1 × 10⁻¹⁰), strongly implicating PDZK1 as the causal gene at this locus. The PDZK1 gene product is known to interact with the HDLC-related gene SCARB1 (ref. ³³).

Extended Data Fig. 6. Implicated putative causal genes in the established lipid-associated loci.

a–d, The horizontal axes represent genomic coordinates, while the vertical axes denote the negative log₁₀ P-values for PDZK1 (a), SREBF1 (b), AR (c), and CREB3L1 (d). Red dots illustrate the association of rare coding variants in genes with significant variants. In contrast, blue dots show the association of rare coding variants in genes without significant variants. Gray dots represent common variant associations from a previous study³⁹. The dashed line in the top panel indicates the exome-wide significance threshold (P < 4.5 × 10⁻⁹). The bottom panel illustrates the coding genes within the locus; genes harboring significant variants are highlighted in red, and others are in blue. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. LDLC, low-density lipoprotein cholesterol; HDLC, high-density lipoprotein cholesterol; TG, triglycerides; MVP, Million Veteran Program; UKB, UK Biobank; AFR, African-like population; EUR, European-like population; HIS, Hispanic-like population.

Another example is SREBF1, which encodes a master regulator for lipogenesis³⁴. rs114001633 is a rare missense variant in SREBF1 that is associated with higher TC levels (MAF = 0.74%, β = 0.0442, P = 2.8 × 10⁻⁹) and is 372-kb upstream from the index GWAS noncoding variant (rs3088233; Extended Data Fig. 6b). In this region, we observed two additional signals (rs35372447 and rs9909417) that reach genome-wide significance and that are mutually independent (r² ≤ 0.031)³⁵. Although SREBF1 is near rs35372447 (220-kb downstream), rs35372447 is located between PEMT and RAI1. Nevertheless, rs3088233 and rs35372447 are both strong expression quantitative trait loci for SREBF1 (ref. ³⁶), supporting a causal mechanism involving SREBF1 in this complex locus.

A further example is CELSR2 on chromosome 1. Previously, SORT1 was identified as a putative causal gene in this locus based on functional validation of the lead noncoding variant, which is a SORT1 expression quantitative trait locus³⁷. However, in this study, we did not see EWS coding variants in SORT1. Instead, we observed four independent missense variants in CELSR2 (p.Q126K, p.R2253H, p.V2287I, p.P2807A), each with highly significant associations with LDLC. From these results, in addition to SORT1, it may be speculated that CELSR2 could potentially be involved in lipid metabolism in this region, as suggested by a recent mouse study³⁸.

Other examples included a missense association in the androgen receptor (AR) p.Q799E on X chromosome with HDLC (Extended Data Fig. 6c) and CREB3L1 on chromosome 11 with TG (Extended Data Fig. 6d). Of 150 putative effector genes with EWS coding variants in known GWAS loci, 87 (58%) were the nearest genes of GWAS lead variants³⁹ (Supplementary Table 12).

By systematic conditioning analysis and introducing rare coding alleles as covariates, we confirmed independence of rare coding associations and common genetic associations (Extended Data Fig. 7 and Supplementary Note ‘Independent association of rare coding variants from common variants for lipids’). Reflecting functional relevance, we observed stronger enrichment of genes harboring EWS coding variants than the nearest genes to the common variant GWAS signals in relevant pathways (Extended Data Fig. 8, Supplementary Table 13 and Supplementary Note ‘Enhanced gene set enrichment by rare coding associations for lipids’). We also observed high but not full concordance across genes harboring EWS coding variants and those identified through common variant-based gene prioritization⁴⁰.

Extended Data Fig. 7. Independence of common genetic signals and rare genetic signals.

a, Each dot indicates common genetic variant (minor allele frequency ≥ 1%) associated with blood lipids within the loci identified by rare genetic associations in this study. We compare nonconditioned and conditioned statistics in this figure to assess the independence of common genetic signals and rare genetic signals. In conditioned analysis, we introduced all the associated rare variant genotypes as covariates in the linear regression model (Methods; Supplementary Note—Independent association of rare coding variants from common variants for lipids). The horizontal axes show the negative log₁₀ P-values without conditioning, and the vertical axes show them with conditioning by exome-wide significant rare variant genotypes. The P-values were calculated by linear regression model with two-sided test. The P-values were not adjusted for multiple testing correction. b, The number of common genetic signals affected by rare genetic signals is summarized in the bar chart. The bar chart indicates the number of common genetic signals, and the color classifies the signals based on the P-values of common genetic signals after conditioning by rare genetic signals. Two-sided P-values were obtained from a linear regression model and were not adjusted for multiple comparisons. GWS, genome-wide significant; MVP, Million Veteran Program; UKB, UK Biobank; AFR, African-like population; AMR, Admixed-American-like population; ASN, Asian-like population; EAS, East-Asian-like population; EUR, European-like population; HIS, Hispanic-like population; SAS, South-Asian-like population; TC, total cholesterol; LDLC, low-density lipoprotein cholesterol; HDLC, high-density lipoprotein cholesterol; TG, triglycerides.

Extended Data Fig. 8. Enhanced enrichment of associated genes in putative causal pathways.

a, Pathway enrichment by common and rare genetic signals. Venn diagram showing significantly enriched pathways for gene sets based on common and rare variant associations. The gene set for common variants was defined by the nearest genes to the lead common variant (Closest genes)³⁹, while the gene set for rare variants was defined by the genes harboring exome-wide significant associations in this study (Coding evidence). b, Pairwise comparison of odds ratios for gene sets (n = 96) associated with both common and rare variants. The vertical axis shows the relative odds ratio (OR_coding/OR_closest). Two-sided P-value was obtained from the paired Wilcoxon rank-sum test. Box plot shows the median value as the centerline; box boundaries show the first and third quartiles and whiskers extending 1.5× the interquartile range. c, Pathway enrichment analysis was performed on genes harboring rare coding variants associated with lipids (Coding evidence) and on genes closest to common variant associations with blood lipids (Closest genes). The top five enriched pathways for each trait are displayed. The horizontal axis denotes the odds ratio, with yellow bars indicating the odds ratios for the gene set with rare variants and blue bars for the gene set with common variants. GO, gene ontology; TC, total cholesterol; HDLC, high-density lipoprotein cholesterol; LDLC, low-density lipoprotein cholesterol; TG, triglycerides.

Population-enriched coding associations and shared effect sizes

Inclusion of the diverse populations enabled testing for associations with ancestry-enriched alleles. By intrapopulation meta-analysis, we identified 655 signals in EUR, 124 in AFR and 45 in Admixed-American-like populations (AMR; Fig. 3a). Most of these signals are population-specific (631/655 associations were specific for EUR, 105/124 for AFR and 18/45 for AMR), and, overall, we identified 130 lipid-associated alleles that were only significant in non-EUR populations. These alleles are exclusively or dominantly found in AFR/AMR (Fig. 3b).

Fig. 3. Shared allelic effects across diverse populations.

a, Upset plot shows the combinations of populations that observed EWS signals through intrapopulation meta-analysis. The bar chart at the top quantifies the number of EWS associations across various combinations of populations. Each bar represents the total number of associations observed for specific combinations of populations, as indicated by the connected points in the central matrix. The central matrix shows the population combinations involved in each set of associations, where filled squares indicate the populations included in a particular combination. Right: the horizontal bar chart shows the number of associations observed within each population individually. b, Allele frequency comparison for non-EUR specific signals. Each point represents an EWS association that is significant only in non-EUR groups (left, AFR; right, AMR/HIS). The vertical axes show the MAF in EUR, while the horizontal axes show the MAF in AFR or AMR/HIS. Gray points indicate variants that were not tested in the EUR group due to low allele frequencies. c, Observed effect sizes across populations. Each point indicates EWS variant–trait pair. The horizontal axes show the effect sizes in the EUR population. The vertical axes show the effect sizes in AFR and AMR/HIS populations. Dots indicate estimated effect sizes and error bars show 95% CIs. r² indicates the squared Pearson’s correlation coefficient of effect sizes. Sample size by population (min–max across analyses)—n_AFR = 120,442–146,876; n_AMR = 12,259–66,387; n_EUR = 374,479–911,705. d, Consistent effect size of PCSK9 p.C679X (stop-gain) variant across multiple populations. The squares indicate estimated effect sizes of PCSK9 p.C679X on blood LDLC level in the studied population. The error bars show its 95% CI. The size of squares is proportional to alternate allele frequency (AAF). Two-sided P values were obtained from a linear regression model and were not adjusted for multiple comparisons.

While we observed substantial differences in variant frequencies, we found highly similar effect sizes across genetically dissimilar groups (r² = 0.9; Fig. 3c) for EWS variants. One example is a stop-gain variant in PCSK9 (p.C679X; Fig. 3d), which is dominant in AFR (MAF_AOU-AFR = 0.951%, MAF_AOU-AMR < 0.211%, MAF_MVP-AFR = 0.828%, MAF_MVP-EUR = 0.009%, MAF_MVP-HIS < 0.039%, MAF_UKB-AFR = 0.978%), but included consistently large effects on LDLC levels (median β = −1.036 (range = −1.140 to −0.874)) across populations. Furthermore, this variant was significantly associated with high HDLC levels, consistent with recent findings from experimental biology^41–43 and human trials^44,45.

As demonstrated with polygenic risk scores, estimates from large population studies are expected to be valuable resources for assessing individual risk. To explore the feasibility of a rare variant-based risk score, we estimated the carrier frequencies of these alleles in the study populations. The prevalence of lipid-associated variants was 67.5% in MVP and 74.0% in UKB overall, but there were substantial differences among genetically similar groups, with the highest in EUR (MVP = 75.7%, UKB = 75.2%) and the lowest in Asian-like population (ASN) or East-Asian-like population (EAS; MVP = 15.4%, UKB = 5.5%), likely due to differences in the size of the discovery analysis (Extended Data Fig. 9 and Supplementary Table 14).

Extended Data Fig. 9. Limited discovery of non-European lipid-associated alleles.

This figure shows the proportion of individuals with lipid-associated alleles identified in this study. The colors of bar charts indicate allele counts of lipid-associated alleles possessed by individuals. The percentages in the bars show the proportion of individuals without lipid-associated alleles in the population. MVP, Million Veteran Program; UKB, UK Biobank; AFR, African-like population; AMR, Admixed-American-like population; HIS, Hispanic-like population; ASN, Asian-like population; EAS, East-Asian-like population; EUR, European-like population; SAS, South-Asian-like population.

Insights from recessive modeling

We identified 109 variant–trait pairs with significant associations in the recessive model (P_recessive < 4.5 × 10⁻⁹; Fig. 4 and Supplementary Table 7). Among these associations, we observed 38 recessive effect sizes that significantly deviated from the additive assumption (P_deviation < 0.05/109; Methods). One example is ANGPTL4 p.E40K for TG with larger effect sizes in the recessive model (β_TG-recessive = −0.818) compared to the additive expectation (2 × β_TG-additive = −0.540). Another example is TM6SF2 p.L156P, which showed >3 times higher effect size for LDLC in the recessive model (β_{LDLC-recessive} = −0.942, P_{LDLC-recessive} = 1.1 × 10⁻³²) compared to the additive expectation (2 × β_{LDLC-additive} = −0.307). Heterozygosity for this variant has been linked to hepatic TG accumulation and impaired very low-density lipoproteins (a hepatic precursor of LDL) intracellular trafficking. Another example was observed for HBB p. E7V (rs334), the causal variant for sickle cell anemia, with TC and LDLC. While the additive associations were weak for these traits (P_TC-additive = 0.0005 and P_{LDLC-additive} = 0.017), the recessive associations showed the largest effect sizes (β_TC-recessive = −1.26, P_TC-recessive = 2.9 × 10⁻¹⁹, P_TC-deviation = 9.9 × 10⁻¹⁵ and β_{LDLC-recessive} = −1.12, P_{LDLC-recessive} = 8.3 × 10⁻¹², P_{LDLC-deviation} = 2.0 × 10⁻⁹). Purely recessive associations were also observed for ABHD15 pLoF and lower TG (β_TG-recessive = −0.586, P_TG-recessive = 5.8 × 10⁻¹¹). In the heterozygote state, the association was not observed (β_TG-additive = 0.01, P_TG-additive = 0.12). ABHD15 is known to interact with PDE3B and influence insulin signaling⁴⁶. We did not detect strong recessive associations or strong deviation from additive assumptions among assessed variants in the previously known recessive FH genes (LDLRAP1 (refs. ^47,48), ABCG5, ABCG8, LIPA and CYP7A1), while the numbers of tested variants were limited (22 variants in total).

Fig. 4. Recessive alleles associated with blood lipids.

a, Comparison of effect sizes between additive and recessive models. The horizontal axis displays the effect size as estimated by linear model under additive assumption, while the vertical axis shows the effect size estimated under recessive assumption (Methods). Dots indicate estimated effect sizes for genetic variants and error bars show 95% CIs. Dashed lines represent the predictions of recessive effect sizes based on the additive model estimates (y = 2x) and estimates that are twice as large (y = 4x) as those from the additive model. The color of the dots indicates the variant class. If β_recessive did not significantly deviate from β_additive, the dots were colored gray. Sample sizes for each variant are reported in Supplementary Table 7. b, Effect size from population-wise or meta-analysis estimates for variants with the largest deviations in recessive estimates from the predicted effect sizes based on additive model estimates. Gray dots represent additive effect sizes, while dark blue dots correspond to recessive effect sizes calculated by linear model. Error bars show 95% CIs. Sample size by cohort and population (min–max across analyses)—n_AOU:AFR = 12,721; n_AOU:EUR = 53,851–54,601; n_MVP:EUR = 444,840–447,007; n_MVP:HIS = 50,920; n_UKB:AFR = 7,410; n_UKB:EUR = 375,137–409,472.

Pathogenicity reassessment of FH variants

Curated pathogenic variants have a crucial role in the molecular diagnosis of FH. To contribute to this essential resource, we re-evaluated curated variants using our population-scale genomic dataset. By intersecting 9,620 FH-related variants reported in ClinVar database⁴⁹ with 1,520 tested variants in this study, we identified 79 pathogenic/likely pathogenic (P/LP) variants, 721 benign/likely benign (B/LB) variants and 483 variants of uncertain significance (VUS) in PCSK9, APOB and LDLR, respectively. The B/LB variants showed a higher allele frequency compared to other classes (Fig. 5a and Supplementary Note ‘Assessment for clinical hypercholesterolemia associated variant’). More than half of the P/LP variants (46/79) are associated with higher LDLC levels (P < 4.3 × 10⁻⁵ (0.05/1,520); Fig. 5b) with median β = 1.56 s.d._LDLC per allele (range = 0.51–2.61). Notably, despite fixed clinical categories of pathogenicity, expressivity varied and was overlapping (Fig. 5c).

Fig. 5. Reevaluation of clinically curated pathogenic variants for FH.

a, Variant allele frequencies of FH-related ClinVar variants observed in the study. Boxes show the IQR of MAFs, and the centerline indicates the median. Whiskers extend to the most extreme values within 1.5× IQR of the quartiles. b, Phenotype associations of FH-related ClinVar variants. The height of the bar indicates total number of variants in the category, and the blue color indicates the proportion of the variants significantly associated with clinical LDLC levels in this study. Statistical significance was determined using the Bonferroni adjustment. c, Distribution of the effect sizes for ClinVar FH-associated variants determined in this study. Each point represents a variant in PCSK9, APOB or LDLR. The color of each point indicates the gene where the variant resides, and the shape indicates its associated disease status in ClinVar. Squares represent variants not registered in ClinVar at the time of the study (31 March 2025). Filled circles and triangles indicate variants annotated only as ‘hypercholesterolemia’, whereas open circles and triangles indicate variants annotated as both ‘hypercholesterolemia’ and ‘hypocholesterolemia’. Triangles specifically indicate variants reclassified due to their large effect size on LDLC. The dashed, vertical line indicates median effect size for established pathogenic variants. Triangles indicate VUS with large effect sizes, as well as pathogenic variants with a negative effect size on clinical LDLC levels.

We identified seven variants across the B/LB/VUS categories with equivalent effect sizes (median β = 1.66 (1.43–1.89); Supplementary Table 15) to P/LP variants, including two missense variants in PCSK9 (p.E40K and p.E197K), one in APOB (p.K3524T) and four in LDLR (p.H327Y, p.R440G, p.L456P and p.A705P). Among these, LDLR p.H327Y is enriched in South-Asian-like population (SAS; MAF < 0.215%, β = 1.75 (s.e. = 0.32), P = 5.4 × 10⁻⁸), but the pathogenicity of this variant was inconclusive in ClinVar. However, its highly significant association with a larger effect on LDLC than the median effect size of established P/LP variants, supports a pathogenic role for this variant in FH. Another variant, LDLR p.L456P, was enriched in AFR (MAF < 0.0165%, β = 1.66 (s.e. = 0.25), P = 6.4 × 10⁻¹¹).

Clinical outcomes of lipid-associated alleles

To connect the lipid-associated alleles and clinical outcomes, we tested for 800 lipid-associated alleles identified in this study with prevalent/incident CAD. We used a logistic regression framework to test for significant associations between the lipid-associated variants and the occurrence of CAD (Methods). We also observed positive associations of TC, LDLC and TG with CAD risk (Fig. 6 and Supplementary Table 16), including several strong associations for known FH pathogenic variants in LDLC (p.C197Y, p.C184Y, splicing variant). On the other hand, deleterious variants in SCARB1 have been linked to both higher HDLC levels and increased CAD risk. Several established lipid-associated genes (PCSK9, APOB, NPC1L1, ANGPTL3/ANGPTL4, APOC3 and LDLR) were associated with lower LDLC/TG and decreased CAD risk with nominal significance (P_CAD < 0.05). Overall, we identified five genes significantly associated with CAD (FDR_CAD < 0.05), namely, RORC, CFAP65, GTF2E2, PLCB3 and ZNF559-177. Among these genes, RORC is particularly intriguing considering prior studies. In a mouse model, suppression of RORγ resulted in improved metabolic phenotypes such as glucose tolerance⁵⁰. Furthermore, in vitro and in vivo studies suggested beneficial effects of silencing RORC on the development of atherosclerotic disease^51–53, consistent with the protective effect of pLoF in RORC for CAD observed in this study.

Fig. 6. CAD risks in blood lipid-associated alleles.

a–d, Scatterplots indicate effect size in lipids on the horizontal axes: TC (a), LDLC (b), HDLC (c), and TG (d) and log(OR) for CAD on the vertical axes. Nominally associated (P < 0.05) variants with CAD are highlighted in red and the sizes of the points indicate MAF. The associated gene names are highlighted in the corner of quadrant, and the number of associations is indicated. Two-sided P values were obtained from a logistic regression model and were not adjusted for multiple comparisons.

Discussion

In this study, we conducted the largest rare variant association study of blood lipids to date. The substantial sample size enabled the analysis of single rare variants as opposed to more conventional aggregation of rare variants into a statistical unit for burden testing. This analysis not only advances new mechanistic insights but also improves the clinical interpretation of Mendelian dyslipidemia genotypes beyond the current clinical classification schema. Overall, this study demonstrated the capability of population-based analyses to identify rare coding alleles with both mechanistic and clinical implications.

Notably, our study expands allelic diversity by including large cohorts from non-EURs, resulting in the discovery of 130 alleles that are exclusively or dominantly observed in the non-EURs. We typically observed consistent and highly similar effect sizes across populations despite differences in allele frequencies. The transferability of associated rare coding alleles may reflect the causality of these alleles and is consistent with our observations from the systematic evaluation of rare variant burden testing across various traits⁵⁴.

In addition to insights specific to blood lipids, our study provides several observations that may be generalizable. Specifically, our expansive rare variant association study of highly heritable phenotypes enabled qualitative/quantitative assessment of the variant characteristics underlying significant associations. In our study, associated variants are significantly enriched in functional variant classes (HC pLoF or deleterious missense variants), highlighting the importance of further effort for precise classification of variant functionality. We further implemented machine-learning-based splice site prediction²⁴, and successfully reclassified previously underestimated variant class, consistent with our recent publication⁵⁵ in which we demonstrated its clinical associations and experimental validation⁵⁶. The cryptic splice variants showed constraint patterns and enrichment of associated variants equivalent to those of canonical pLoF variants.

The broad range of allele-specific analyses also allowed us to infer the effects of variants on gene function. First, we observed highly concordant effect directions of pLoF alleles (>99%) on the same gene proxied by phenotypic expression, confirming findings from previous studies. Also, most (87%) missense variants showed concordant effect directions with pLoFs; however, the remainder (13%) showed opposite effects, indicating hypermorphic characteristics. In silico deleterious prediction is not effective to capture these hypermorphic alleles, and these variants might be missed by variant filters for gene-based testing despite their empirical functional significance. Increasingly available large-scale genetic analysis across diverse phenotypes and populations, focusing on rare coding variants, may expand the list of hypermorphic alleles and inform models to better detect this phenomenon across genes and domains. Overall, these observations reinforce the current strategy of variant-aggregating burden testing under the assumption of consistent effect directions, with in silico deleterious prediction usefully excluding hypermorphic missense variants. However, the presence of alleles with opposing effects highlights room for improvement in aggregation testing. More accurate inference of the functional direction of variants, and incorporating these characteristics into models, may further enhance the sensitivity and accuracy of gene-based aggregation testing.

We also conducted recessive modeling and identified multiple strong associations. Intriguingly, the recessive associations are robustly shared across studies and populations. Aligned with previous studies focusing on binary traits⁵⁷, some rare alleles have a prominent recessive effect not captured by standard additive modeling, suggesting a contribution to the missing heritability. However, due to the rarity of homozygous conditions for single variants, the assessment was only conducted for a very limited number of variants. Further detailed approaches are required to truly assess the impact of recessive inheritance.

In addition, using estimated effect size and statistical significance driven by population-scale association analysis, we re-assessed a curated database considered as a gold standard for clinical genetic diagnosis. We confirmed the accuracy of most variant annotations aligned with a previous study⁵⁸ and further provided evidence toward potential reclassification of the pathogenicity of other variants. We found two notable candidate variants that are enriched in non-European populations that may represent new FH-related variants, consistent with prior reports^18,59,60, suggesting lower annotation rates for putative monogenic variants enriched in non-European populations compared to European populations. Notably, we observed a range of expressivity for pathogenic alleles that was associated with clinical outcomes.

One important limitation of our study is that it primarily focuses on rare coding variants, while the potential functional impact of rare noncoding variants has not been sufficiently addressed. The observed enrichment of coding variants supports their functional importance and causality. Nonetheless, further integrative analyses using large-scale WGS data⁶¹ will be needed to more fully explore the interplay between coding and noncoding variants.

In conclusion, we conducted a rare variant-focused genetic study for blood lipids involving over a million individuals, yielding hundreds of rare alleles associated with blood lipids and improved mechanistic understanding of rare variant associations. Our study suggests that population-scale rare variant analysis is now adequately powered for heritable phenotypes, allowing for the classification of rare pathogenic alleles and providing new insights into variant expressivity and penetrance, toward improved diagnosis and more quantitative prognosis.

Methods

Ethics oversight

This study received ethics approval from the Veterans Affairs (VA) Central Institutional Review Board (protocol 16-06). The study protocols were approved by the Mass General Brigham Institutional Review Board (protocols 2016P002395 and 2021P002228). The analysis for UKB was performed under application 7089.

Blood lipids phenotyping

UKB is a volunteer cohort²⁰ of approximately 500,000 residents aged 40 to 69 years living in the United Kingdom, recruited in 2006–2010. In the UKB, blood lipids were measured using blood samples collected at enrollment. We adjusted TC and LDLC levels by dividing by 0.7 for individuals prescribed lipid-lowering medication at enrollment as previously described³⁹.

The MVP is a national hospital-based cohort initiated in 2011 by the United States Department of VA. Recruitment was conducted in the VA-affiliated hospitals across the United States^19,62. In the MVP, lipid phenotypes were derived from longitudinal lipid measurements over time. For TC, LDLC and TG, we used the highest value recorded, while for HDLC we selected the lowest value as previously described⁶³.

The AOU is a U.S.-based population cohort that began enrollment in 2018 under the National Institutes of Health. Participants were enrolled through a network of more than 340 recruitment sites. In AOU, lipid phenotypes were derived similarly to those in MVP.

CAD phenotyping

In the UKB and AOU, we ascertained CAD cases based on at least one of the following criteria: (1) any International Classification of Diseases (ICD) code in the in-hospital record or death registry (I21–I25 in ICD10; 410–414 in ICD9), or (2) any procedure code for coronary revascularization (K40–K45, K49, K50 and K75 in Office of Population Censuses and Surveys’ Classification of Surgical Operations version 4 (OPCS4), 33510–33523, 33533–33536, 92920–92950 in Current Procedural Terminology, Fourth Edition). In the MVP, we used a previously established CAD definition⁶⁴. ICD9, ICD10 and Current Procedural Terminology codes, along with self-report, were used to determine CAD cases and controls. Qualifying codes were those pertaining to acute myocardial infarction (inpatient only), stable ischemic heart disease (inpatient or outpatient) and coronary revascularization (inpatient and outpatient). Cases were individuals who had at least two qualifying codes on different dates within a 12-month period. Controls were individuals who carried no codes and who did not self-report a history of CAD.

Quality control for microarray genotyping in UKB

We conducted sample quality control as discussed further. Among 488,175 individuals, we removed samples with aneuploidy (n = 651), sex–gender mismatch implying phenotypic quality issues (n = 378), higher heterozygosity or missing outlier (n = 968), leading to a total of 1,811 (0.4%) samples removed. A total of 486,364 quality control passed individuals remained (9,454 AFR; 2,413 AMR; 2,582 EAS; 461,352 EUR and 10,563 SAS).

Population ascertainment in UKB

Using reference population data from the 1000 Genomes Project and microarray genotypes in UKB using the Affymetrix UK BiLEVE Axiom and UKB Axiom arrays, we established genetically determined population ascertainment. First, we extracted quality-controlled variants from 1000 Genomes data (nonpalindromic SNV, MAF > 1%; a population-specific Hardy–Weinberg equilibrium (HWE), P > 1 × 10⁻⁶). Next, we extracted the intersection of the quality-controlled 1000 Genomes data and the study population. Using intersected variants, we pruned variants on the 1000 Genomes data using PLINK2 (ref. ⁶⁵) software (3 June 2022 release) with --indep-pairwise option (window size of 50, sliding window size of 10, r² < 0.2). This yielded 224,993 variants. Using pruned variants, we calculated the SNV weights for genetic principal components (PCs). Then, we projected study participants to the genetic PC space. Using the 1000 Genomes reference population annotation, we trained the k-nearest neighbor model using class R package (version 7.3). Then, we split study cohort into the five genetic populations (AFR, AMR, EAS, EUR and SAS) and conducted association analyses separately to minimize potential effects of heterogeneity.

Quality control for WES in UKB

We curated genotypes from the 450k release using Deep Variant⁶⁶. For genotype-level quality control, we first used Hail’s ‘split_multi_hts’ function to divide multiallelic sites. We then filtered out low-quality genotypes based on the following criteria: (1) genotyping quality ≤20, (2) genotype depth (DP) either ≤10 or >200, (3) for heterozygous genotypes—(DP_reference + DP_alternate)/(DP_total) > 0.9 and DP_alternate/DP_total > 0.2 and (4) for alternate homozygous genotypes—DP_alternate/DP_total > 0.9.

These processes retained 26,645,535 variants in the 454,756 sequenced samples. We excluded (1) 6,131,710 variants with missing rate >10%; (2) 47,441 variants with HWE P < 1 × 10⁻¹⁵ and (3) 364,207 variants located within low-complexity regions⁶⁷. Cumulatively, we excluded 6,289,813 variants, resulting in 20,355,722 retained variants.

Among 454,756 individuals whole-exome sequenced, we identified 452,929 individuals with overlapping array genotyped data. For these individuals, we conducted sample-level quality control. First, we calculated array-exome genotype discordance rate and F statistics in nonpseudo-autosomal region X chromosome variants to detect potential sample or phenotypic swapping in exome data. For this analysis, we used pruned and stringent variant quality control criteria (missingness < 1%, MAF > 0.1%). We identified 27 potential sex-swapping (27 females with F statistics >0.6, 0 male with F statistics <0.6) and 0 discordant genotypes between exome and array data (nonreference homozygote concordance rate of <0.8). We calculated array-exome discordance using pruned, nonpalindromic, high-quality exome data and the corresponding array data. Next, we removed samples with a high missing rate (>10%, n = 12). Then, we filtered samples using the following autosomal quality control outlier metrics (outside mean ± 8 s.d.): heterozygous/homozygous rate (n = 761), transition/transversion rate (n = 0), SNV-to-insertion/deletion ratio (n = 2), number of singletons (n = 283). In total, 1,052 (0.2%) samples were removed, resulting in 451,877 samples in the final dataset. After removing these samples, we also removed 226,083 monomorphic variants, retaining 20,129,639 variants across the 451,877 samples in the final dataset.

Quality control for microarray genotyping, population ascertainment and imputation in MVP

Genotyping was performed using the custom Axiom array (MVP1.0), and variant and sample quality control were previously described in detail¹⁹. We used the latest release (release 4) data for this analysis. Release 4 data included array genotypes and genetic dosage imputed to the TOPMed imputation (version r2) reference panel²¹. The quality-controlled sample size was 657,242. We grouped participants into four population groups (AFR, ASN, EUR and HIS) following the harmonized ancestry and race/ethnicity (HARE) algorithm previously established in MVP⁶⁸.

Quality control for WGS in AOU

We curated genotypes from the jointly called WGS call set (version 7) provided by AOU²³. We split multiallelic sites to biallelic variants using Hail’s ‘split_multi_hts’ function; then, low-quality genotypes flagged as FAIL in FT field were set as missing. The genotypes were exported as BGEN files and converted to PGEN files for quality control procedures and downstream analysis. We filtered variants (1) flagged in the FILTER column in original VDS, (2) located in the low-complexity region, (3) low call rate (<90%), (4) monomorphic and (5) population-specific HWE P < 1 × 10⁻¹⁵. Finally, we excluded flagged individuals (n = 549) and those with a genotype missing rate >1% (n=396).

Exome-wide association analysis

We used the association analysis framework implemented in REGENIE software (version 3.1.3)⁶⁹. We used array-based autosomal genotypes for step 1, excluding variants with MAF < 1%, HWE P < 1 × 10⁻¹⁵, call rate < 98%, and located in the Major Histocompatibility Complex region (chromosome 6: 23–37 Mb). We pruned variants using PLINK2 (ref. ⁶⁵) software (3 June 2022 release) with --indep-pairwise option (window size of 1000, sliding window size of 100, r² < 0.9) by genetic population in each cohort. The association model was adjusted by age, age², sex and the first ten genetic PCs, and blood lipids measurements were inverse rank normalized. For CAD analysis, we used firth logistic regression implemented in REGENIE software. We tested all quality-controlled genotypes in UKB-WES. For the MVP whole-genome-imputed dataset and AOU WGS dataset, we restricted the analysis to the exome sequence targeting file used for UKB-WES (https://biobank.ndph.ox.ac.uk/ukb/refer.cgi?id=3801) with 50 bp flanking on both sides of the target region. After generating summary statistics for each cohort (MVP, AOU and UKB) and each population (AFR, HIS, ASN, EUR in MVP and AFR, AMR, EAS, EUR, SAS in UKB and AOU), we meta-analyzed the results using GWAMA⁷⁰ software (version 2.2.2).

The summary statistics were obtained by meta-analysis using GWAMA⁷⁰ fixed-effect model. We computed effect size and P values for all variants in the exome region, irrespective of variant annotation. While we used summary statistics for synonymous/noncoding variants as reference to contextualize coding associations, the statistical significance of these variants was not considered throughout the study. In this study, we primarily applied additive model for the association analysis. In addition to the additive model, we performed association analyses modeling recessive effects. For the additive model, we restricted the analysis to variants with 5≤ MAC and MAF_POPMAX of <1% before meta-analysis. For the recessive model, we also restricted the analysis to variants with 5≤ estimated minor homozygote counts (number of participants × MAF²) and estimated minor homozygote frequency (MAF²) <1%. To test whether the recessive effect is significantly larger than expected from the additive model, we computed a z score and corresponding P values as follows:

$$z=\frac{2\times {\beta }_{\mathrm{additive}}-{\beta }_{\mathrm{recessive}}}{\sqrt{{\left(2\times \mathrm{s}.\mathrm{e}{.}_{\mathrm{additive}}\right)}^2+{\left(\mathrm{s}.\mathrm{e}{.}_{\mathrm{recessive}}\right)}^2}}$$

Among the 109 recessive associations that surpassed the EWS threshold (P < 4.5 × 10⁻⁹), we tested the recessive effect against the additive effect and found 39 cases with a significant deviation (P < 0.05/109). To define the new genetic loci identified in our study, we assessed 500-kb intervals on each side of significant variants and merged overlapping regions. We then assessed the overlap with regions reported by a recent large-scale GWAS meta-analyses for blood lipids³⁹ and prior reports in the GWAS Catalog⁷¹. Regions identified in this study that did not overlap with previously reported loci were classified as new loci.

Per the MVP and AOU reporting guidelines, we masked the MAF of variants with a MAC ≤40 in the summary statistics. This measure is implemented to prevent the potential identification of individuals participating in the study. The β coefficients and s.e. are shown in the original data, as we received an exception to the All of Us Data and Statistics Dissemination Policy from the All of Us Resource Access Board. The directions of the effects in the Supplementary Tables 6, 7, 15 and 16 and in the published summary statistics, indicate the variant effect in alphabetical order—AOU_AFR, AOU_AMR, AOU_EAS, AOU_EUR, AOU_SAS, MVP_AFR, MVP_ASN, MVP_EUR, MVP_HIS, UKB_AFR, UKB_AMR, UKB_EAS, UKB_EUR and UKB_SAS. We reported unadjusted P values without correction for multiple testing throughout the manuscript.

Variant annotation

We used a single transcript for each gene based on Gencode (version 41)⁷² canonical and coding transcript (coding transcript set, n = 19,603; Supplementary Data) for all annotations. First, we annotated tested variants (variants within ±50 bp from target region in the UKB-WES) with the VEP²⁷ software (version 107, aligned with Gencode (version 41)) and selected annotations on the coding transcript set. If we found variants overlapping in more than two transcripts in the coding transcripts set, we selected higher functional consequence. If the consequences were equivalent, we selected the annotation on the longest transcript. We selected pLoF (‘Impact High’) or missense (‘Impact Moderate’) variants as coding variants. Next, we ran SpliceAI²⁴ for all the variants tested for the coding transcript set with default parameters. SpliceAI returns DS, which represent potential for cryptic splicing (Supplementary Note ‘Power calculation’). We treated non-pLoF variants with DS >0.8 as cryptic splice variants and reclassified them as pLoF.

MiS

For further classification of missense variants, we applied ensemble prediction using 29 in silico prediction models to assess the deleteriousness of missense SNVs as we previously established^54,73. Using precomputed in silico predictions from the dbNSFP²⁸ database (version 4.2), we annotated all the missense variants using the dbNSFP plugin for VEP. We binarized the predictions into ‘deleterious’ or ‘tolerant’ using an algorithm-specific threshold. The summaries of the prediction models used in this study are provided in Supplementary Tables 9 and 10. We computed the MiS, ranging from 0 to 1, by dividing the number of deleterious predictions by the total number of available algorithms for the corresponding variant.

Pathway enrichment analysis

We compared pathway enrichments among genes identified by rare coding variant associations in this study and those identified by common variant association studies. The analysis was restricted to genes included in the coding transcript set (n = 19,603). For genes supported by rare coding variant associations in this study, we chose those with the smallest P value within their respective loci. For genes supported by common variant association studies, we selected the closest gene to the lead variant identified in a recent GWAS³⁹. For the selected gene sets, enrichment was tested using the enrichR (version 3.0) R package⁷⁴, considering the following pathway sets: Reactome_2022, KEGG_2021_Human, GO_Biological_Process_2023, GO_Cellular_Component_2023, GO_Molecular_Function_2023, ChEA_2022, ENCODE_TF_ChIP-seq_2015, ENCODE and ChEA Consensus TFs from ChIP-X, and Enrichr Submissions TF-Gene Cooccurrence. Enrichment was considered significant if the P value was lower than the threshold adjusted by the Holm method.

Replication analysis

For replication, we used data from a previous large-scale exome-array study⁷⁵. All variants were updated to the hg38 reference genome using the LiftoverVcf function in GATK⁷⁶. We then combined the updated summary statistics from the previous study with those from our current study. In total, we identified 387 combinations of variants and phenotypes that matched between both studies. The concordance of effect sizes and statistical significance was assessed. A directional concordance was noted if the effect direction was the same in both the replication dataset and our study.

Variance explained

Per variant explained variance (Var) was computed by the following formula³⁹, where f represents the allele frequency and β represents the effect size:

$$\mathrm{Var}=2f\left(1-f\right){\beta }^2$$

We calculated the variance explained by common variants using the index variant from the latest GWAS for common variants (511 variants for TC, 442 variants for LDLC, 562 variants for HDLC, 480 variants for TG)³⁹. To eliminate linked variants in EWS rare variants, we used the ‘clump’ function in PLINK1.9 (ref. ⁶⁵). With the MVP-imputed genotypes and UKB-WES, we clumped EWS variants using an r² threshold of 0.01. By this process, variants in linkage disequilibrium (r² ≥ 0.01) with any variants that had smaller P values were excluded. In addition, EWS variants that do not present in the MVP or UKB were omitted from the analysis. As a result, 172, 195, 182 and 121 variants from MVP and 179, 197, 185 and 128 variants from UKB were retained for TC, LDLC, HDLC and TG, respectively.

Conditioning analysis

To evaluate the independence of genetic signals derived from rare coding variants and common variants, we executed a conditional analysis where rare coding variants were incorporated as covariates. This analysis was performed in addition to using the standard covariates applied in our primary analyses, which included sex, age, the age² and the first ten genetic PCs. For conditioning purposes, we used genotype data for EWS rare coding variants as covariates. In the MVP, we introduced 185, 207, 203 and 131 rare coding variants as covariates for TC, LDLC, HDLC and TG, respectively. Similarly, in the UKB, 197, 224, 209 and 140 variants were introduced to the model for TC, LDLC, HDLC and TG, respectively. By comparing the β and P values obtained from the analyses conducted with or without these genotype covariates, we aimed to ascertain the extent to which signals from rare variants are dependent on or independent from those associated with common variants. This approach leveraged the ‘condition’ function available in the REGENIE software package. Conditioning was done in both step 1 and step 2.

Pathogenic variant reclassification

We curated pathogenic alleles for FH, a well-known monogenic condition linked to severe hypercholesterolemia and premature CAD, from the ClinVar⁴⁹ database, downloading bulk data on 31 March 2025. We first extracted genetic regions corresponding to the PCSK9, APOB and LDLR genes from the VCF file. Using the same pipeline as for the tested variants, we annotated these variants and excluded pLoF variants for PCSK9 and APOB due to their known reduction of LDLC levels. We then classified variants as necessary with conflicting interpretations by majority vote. We calculated the difference in evidence (number of pathogenic + likely pathogenic − (benign + likely benign + uncertain significance)); if the score was greater than 0, the variants were considered P/LP; otherwise, they were considered benign. This process resulted in a categorized list of variants in three classes—(1) P/LP, (2) VUS and (3) B/LB. We intersected these variants with those in APOB/LDLR/PCSK9 tested in this study and defined the pathogenic effect size by taking the median of positive effect sizes from known pathogenic variants. To identify a subset of VUS to be reclassified as P/LP, we ranked the variants by their effect sizes and grouped them accordingly, ensuring that the median effect size of this group was larger than that of the known P/LP variants. Throughout this process, we included all the variants in ClinVar dataset irrespective of the review status.

External data

For replication, we obtained summary statistics from previous exome-array-based study²⁵. We lifted summary statistics from hg19 coordinates to hg38 using the LiftOverVcf function in the GATK software. We removed insertions/deletions due to ambiguousness of alleles (n = 24) and failed in lifting (n = 71). In total, we successfully lifted >99.96% (292,322/292,417) variants in the data. For common variant integration analysis, we obtained summary statistics from the latest GWAS³⁹. We used summary statistics from trans-population meta-analysis (with_BF_meta-analysis_AFR_EAS_EUR_HIS_SAS_*_INV_ALL_with_N_1.gz, for autosomes and meta-analysis_chrX_AFR_EAS_EUR_HIS_SAS_*_INV_ALL_with_N_1.gz for X chromosome). All summary statistics were downloaded from the Global Lipids Genetics Consortium website (http://www.lipidgenetics.org), and more than 99.82% of variants were successfully lifted.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41588-026-02613-y.

Author contributions

S.K., P.T.E., Y.V.S., P.W.F.W. and P.N. conceptualized the study. S.K., Z.Y., D.K., J.E.H., S.L.C., J.A.L. and K.C. curated phenotype data. S.K., S.H.C., S.J.J., M.S.S., D.K., J.E.H., J.S.D. and P.S.T. curated genotype data. S.K., Z.Y., S.H.C., S.J.J., M.S.S., S.K.Z., M.N.T. and A.R. analyzed data. S.K., J.E.H., S.K.Z., M.N.T., A.R., J.S.D., C.S., I.S., S.M.D., K.-M.C., T.L.A., D.J.R., G.M.P., P.T.E., Y.V.S., P.W.F.W. and P.N. interpreted data. S.K., Y.V.S., P.W.F.W. and P.N. prepared the initial draft of the paper. S.K., Z.Y., S.J.J., M.S.S., D.K., J.S.D., C.S., I.S., S.M.D., K.M.C., T.L.A., D.J.R., G.M.P., P.T.E., Y.V.S., P.W.F.W. and P.N. provided critical review and edits for the paper. W.H., P.S.T., K.C., P.T.E., Y.V.S., P.W.F.W. and P.N. supervised the project. A.B., K.L., W.H., P.S.T., K.C., P.T.E., Y.V.S. and P.W.F.W. managed the project administration. K.C., P.T.E., Y.V.S., P.W.F.W. and P.N. obtained funding for the project.

Peer review

Peer review information

Nature Genetics thanks Liam Brunham, Nathan Stitziel, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Data availability

Summary statistics are publicly available through the dbGaP (https://dbgap.ncbi.nlm.nih.gov/) or VA CIPHER (https://phenomics.va.ornl.gov/web/cipher/partner/mvp) websites under accession phs001672. Individual-level data from MVP and UKB are available upon application to the respective custodians (MVP, https://www.mvp.va.gov/pwa/discover-mvp-data; UKB, https://www.ukbiobank.ac.uk/use-our-data/apply-for-access/).

Code availability

The analysis codes and supplemental data are available via Zenodo at 10.5281/zenodo.11092802 (ref. ⁷⁷). The docker/singularity images used in the analysis are publicly available through docker hub (https://hub.docker.com/u/skoyamamd).

Competing interests

D.K. is a scientific advisor and reports consulting fees from Bitterroot Bio unrelated to the present work. P.T.E. receives sponsored research support from Bayer AG, Bristol Myers Squibb, Pfizer and Novo Nordisk; he has also served on advisory boards or consulted for Bayer AG. P.N. reports research grants from Allelica, Amgen, Apple, Boston Scientific, Cleerly, Genentech/Roche, Ionis, Novartis and Silence Therapeutics; personal fees from AIRNA, Allelica, Apple, AstraZeneca, Bain Capital, Blackstone Life Sciences, Bristol Myers Squibb, Creative Education Concepts, CRISPR Therapeutics, Eli Lilly, Esperion Therapeutics, Foresite Capital, Foresite Labs, Genentech/Roche, GV, HeartFlow, Incyte, Magnet Biomedicine, Merck, Novartis, Novo Nordisk, TenSixteen Bio and Tourmaline Bio; equity in Bolt, Candela, Mercury, MyOme, Parameter Health, Preciseli and TenSixteen Bio; royalties from Recora for intensive cardiac rehabilitation; and spousal employment at Vertex Pharmaceuticals, all unrelated to the present work. The remaining authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41588-026-02613-y.

Supplementary information

The online version contains supplementary material available at 10.1038/s41588-026-02613-y.