A statistical framework for powerful multi-trait rare variant analysis in large-scale whole-genome sequencing studies

Abstract

Large-scale whole-genome sequencing (WGS) studies have improved our understanding of the contributions of coding and noncoding rare variants to complex human traits. Leveraging association effect sizes across multiple traits in WGS rare variant association analysis can improve statistical power over single-trait analysis, and also detect pleiotropic genes and regions. Existing multi-trait methods have limited ability to perform rare variant analysis of large-scale WGS data. We propose MultiSTAAR, a statistical framework and computationally-scalable analytical pipeline for functionally-informed multi-trait rare variant analysis in large-scale WGS studies. MultiSTAAR accounts for relatedness, population structure and correlation among phenotypes by jointly analyzing multiple traits, and further empowers rare variant association analysis by incorporating multiple functional annotations. We applied MultiSTAAR to jointly analyze three lipid traits (low-density lipoprotein cholesterol, high-density lipoprotein cholesterol and triglycerides) in 61,861 multi-ethnic samples from the Trans-Omics for Precision Medicine (TOPMed) Program. We discovered new associations with lipid traits missed by single-trait analysis, including rare variants within an enhancer of NIPSNAP3A and an intergenic region on chromosome 1.

Advances in next generation sequencing technologies and the decreasing cost of whole-exome/whole-genome sequencing (WES/WGS) have made it possible to study the genetic underpinnings of rare variants (i.e. minor allele frequency (MAF) < 1%) in complex human traits. Large nationwide consortia and biobanks, such as the National Heart, Lung and Blood Institute (NHLBI)’s Trans-Omics for Precision Medicine (TOPMed) Program¹, the National Human Genome Research Institute’s Genome Sequencing Program (GSP) , the National Institute of Health’s All of Us Research Program², and the UK’s Biobank WGS Program³, are expected to sequence more than a million of individuals in total, at more than 1 billion genetic variants in both coding and noncoding regions of the human genome, while also recording thousands of phenotypes. To mitigate the lack of power of single-variant analyses to identify rare variant associations⁴, variant set tests have been proposed to analyze the joint effects of multiple rare variants ^5-9, where most of the work has focused single trait analysis.

Pleiotropy occurs when genetic variants influence multiple traits¹⁰. There is growing empirical evidence from genome-wide association studies (GWASs) that many variants have pleiotropic effects^11,12. Identifying these effects can provide valuable insights into the genetic architecture of complex traits¹³. As such, it is of increasing interest to identify pleiotropic rare variants by jointly analyzing multiple traits in WGS rare variant association studies (RVASs).

Several existing methods for multi-trait rare variant association analysis, such as MSKAT¹⁴, Multi-SKAT¹⁵ and MTAR¹⁶, have shown that leveraging the cross-phenotype correlation structure can improve the power of multi-trait analyses compared to single-trait analyses when analyzing pleiotropic genes^14-17. However, existing methods do not scale well, and are not feasible when analyzing large-scale WGS studies with hundreds of millions of rare variants in samples exhibiting relatedness and population structure. Furthermore, none of the existing multi-trait rare variant analysis methods leverages functional annotations that predict the biological functionality of variants, resulting in limited interpretability and power loss. While the STAAR method¹⁸ dynamically incorporates multiple variant functional annotations to maximize the power of rare variant association tests, it is designed for single-trait analysis and cannot be directly applied to multiple traits.

To overcome these limitations, we propose the Multi-trait variant-Set Test for Association using Annotation infoRmation (MultiSTAAR), a statistical framework for multi-trait rare variant analyses of large-scale WGS studies and biobanks. It has several features. First, by fitting a null Multivariate Linear Mixed Model (MLMM)¹⁹ for multiple quantitative traits simultaneously, adjusting for ancestry principal components (PCs)²⁰ and using a sparse genetic relatedness matrix (GRM)^21,22, MultiSTAAR scales well but also accounts for relatedness and population structure, as well as correlations among the multiple traits. Second, MultiSTAAR enables the incorporation of multiple variant functional annotations as weights to improve the power of RVASs. Furthermore, we provide MultiSTAAR via a comprehensive pipeline for large-scale WGS studies, that facilitates functionally-informed multi-trait analysis of both coding and noncoding rare variants. Third, MultiSTAAR enables conditional multi-trait analysis to assess rare variant association signals beyond known common and low frequency variants.

In the current study, we conducted extensive simulation studies to demonstrate the validity of MultiSTAAR and to assess the power gain of MultiSTAAR by incorporating multiple relevant variant functional annotations, and its ability in preserving Type I error rates. We then applied MultiSTAAR to perform WGS RVAS of 61,838 ancestrally diverse participants from 20 studies from NHLBI’s TOPMed consortium by jointly analyzing three circulating lipid traits: low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and triglycerides (TG). We show that MultiSTAAR is computationally feasible for large-scale WGS multi-trait rare variant analysis, and in conditional analysis of LDL-C, HDL-C and TG, MultiSTAAR identifies signals that were missed either by the existing multi-trait rare variant analysis methods that overlook variant functional annotations, or by single-trait functionally-informed analysis that ignore correlations between phenotypes.

Results

Overview of the methods

MultiSTAAR is a statistical framework and an analytic pipeline for jointly analyzing multiple traits in large-scale WGS rare variant association studies. There are two main components in the MultiSTAAR framework: (i) fitting null MLMMs using ancestry PCs and sparse GRMs to account for population structure, relatedness and the correlation between phenotypes, and (ii) testing for associations between each aggregated variant set and multiple traits by dynamically incorporating multiple variant functional annotations¹⁸ (Fig. 1a).

Fig. 1 ∣. MultiSTAAR framework and pipeline.

a, MultiSTAAR framework. (i) Fit null Multivariate Linear Mixed Models (MLMMs) using sparse GRM and ancestry PCs to account for population structure, relatedness and the correlation between phenotypes. (ii) Test for associations between each variant set and multiple traits by dynamically incorporating multiple variant functional annotations. b, MultiSTAAR pipeline. (i) Prepare the input data of MultiSTAAR, including genotypes, multiple phenotypes and covariates. (ii) Calculate sparse GRM, ancestry PCs and annotate all variants in the genome. (iii) Perform single variant analysis for common variants. (iv) Define the rare variant analysis units, including gene-centric analysis of five coding functional categories and eight noncoding functional categories and non-gene-centric analysis of sliding windows. (v) Provide result summarization and perform analytical follow-up via conditional analysis.

In WGS RVASs, an important but often underemphasized challenge is selecting biologically-meaningful and functionally-interpretable analysis units, especially for the noncoding genome^23,24. In gene-centric analyses of multiple traits, MultiSTAAR provides five functional categories (masks) to aggregate coding rare variants of each protein-coding gene, as well as an additional eight masks of regulatory regions to aggregate noncoding rare variants. In non-gene-centric analyses of multiple traits, MultiSTAAR performs agnostic genetic region analyses using sliding windows^18,25 (Fig. 1b).

For each rare variant set analyzed, MultiSTAAR first constructs the multi-trait burden, SKAT and ACAT-V test statistics (Methods). For each type of rare variant test, MultiSTAAR calculates multiple candidate P values using different variant functional annotations as weights, following the STAAR framework¹⁸. MultiSTAAR then aggregates the association strength by combining the P values from all annotations using the ACAT method, that provides robustness to correlation between tests⁹, and proposes an omnibus test, MultiSTAAR-O, that leverages the advantages of different type of tests (Methods). Furthermore, MultiSTAAR can test multi-trait rare variants’ associations conditional on a set of known associations (Fig. 1b).

Simulation studies

To evaluate the type I error rates and the power of MultiSTAAR, we performed simulation studies under several configurations. Following the steps described in Data Simulation (Methods), we generated three quantitative traits with a correlation matrix similar to the empirical correlation in the three lipid traits^26-28. We then generated genotypes by simulating 20,000 sequences for 100 different 1 megabase (Mb) regions, each of them were generated to mimic the linkage disequilibrium structure of an African American population by using the calibration coalescent model²⁹. Throughout the simulation studies, we randomly and uniformly selected 5-kilobase (kb) regions from these 1-Mb regions and considered sample sizes of 10,000 for each replicate. The simulation studies focused on aggregating uncommon variants with an MAF < 5%.

Type I error rate evaluations

We performed 10⁸ simulations to evaluate the type I error rates of the multi-trait burden, SKAT, ACAT-V and MultiSTAAR-O tests at α = 10⁻⁴, 10⁻⁵ and 10⁻⁶ (Supplementary Table 1). The results show that, for multi-trait rare variant analysis, all four MultiSTAAR tests controlled the type I error rates at very close to the nominal α levels.

Empirical power simulations

We next assessed the power of MultiSTAAR-O for the analysis of multiple phenotypes under different genetic architectures, while also comparing its power with existing methods. Specifically, we considered four models, in which variants in the signal region (variant-phenotype association regions) were associated with (1) one phenotype only, (2) two positively correlated phenotypes, (3) two negatively correlated phenotypes and (4) all three phenotypes. In addition, we considered different proportions (5%, 15% and 35% on average) of causal variants in the signal region, where causality of variants depended on different sets of annotations, and the effect size directions of causal variants were allowed to vary (Methods). Power was evaluated as the proportions of P values less than α = 10⁻⁷ based on 10⁴ simulations. Overall, MultiSTAAR-O consistently delivered higher power to detect signal regions compared to multi-trait burden, SKAT and ACAT-V tests, through its incorporation of multiple annotations (Extended Data Figs. 2-5, Supplementary Figs. 1-4). This power advantage was also robust to the existence of noninformative annotations.

Application to the TOPMed lipids WGS data

We applied MultiSTAAR to identify rare variant associations with three quantitative lipid traits (LDL-C, HDL-C and TG) through a multi-trait analysis using TOPMed Freeze 8 WGS data, comprising 61,838 individuals from 20 multi-ethnic studies (Supplementary Note). LDL-C values were adjusted for the usage of lipid-lowering medication^26,30 (Methods), and DNA samples were sequenced at >30x target coverage. Sample- and variant-level quality control were performed for each participating study^1,26,30.

Race/ethnicity was measured using a combination of self-reported race/ethnicity and study recruitment information³¹ (Supplementary Note). Of the 61,838 samples, 15,636 (25.3%) were Black or African American, 27,439 (44.4%) were White, 4,461 (7.2%) were Asian or Asian American, 13,138 (21.2%) were Hispanic/Latino American and 1,164 (1.9%) were Samoans. There were 414 million single-nucleotide variants (SNVs) observed overall, with 6.5 million (1.6%) common variants (MAF > 5%), 5.2 million (1.2%) low-frequency variants (1% ≤ MAF ≤ 5%) and 402 million (97.2%) rare variants (MAF < 1%). The study-specific demographics and baseline characteristics are given in Supplementary Table 2.

Gene-centric multi-trait analysis of coding and noncoding rare variants

We applied MultiSTAAR-O on gene-centric multi-trait analysis of coding and noncoding rare variants of genes with lipid traits in TOPMed. For coding variants, rare variants (MAF < 1%) from five coding functional categories (masks) were aggregated, separately, and analyzed using a joint model for LDL-C, HDL-C and TG, including (1) putative loss-of-function (stop gain, stop loss and splice) rare variants, (2) missense rare variants, (3) disruptive missense rare variants, (4) putative loss-of-function and disruptive missense rare variants and (5) synonymous rare variants of each protein-coding gene. The putative loss-of-function, missense and synonymous RVs were defined by GENCODE Variant Effect Predictor (VEP) categories³². The disruptive variants were further defined by MetaSVM³³, which measures the deleteriousness of missense mutations. We incorporated 9 annotation principal components (aPCs)^18,26,34, CADD³⁵, LINSIGHT³⁶, FATHMM-XF³⁷ and MetaSVM³³ (for missense rare variants only) along with the two MAF-based weights⁴ in MultiSTAAR-O (Supplementary Table 3). The overall distribution of MultiSTAAR-O P values was well-calibrated for the multi-trait analysis of coding rare variants (Extended Data Fig. 1b). At a Bonferroni-corrected significance threshold of α = 0.05/(20,000 × 5) = 5.00 × 10⁻⁷ accounting for five different coding masks across protein-coding genes, MultiSTAAR-O identified 51 genome-wide significant associations using unconditional multi-trait analysis (Extended Data Fig. 1a, Supplementary Table 4). After conditioning on previously reported variants associated with LDL-C, HDL-C or TG located within a 1 Mb broader region of each coding mask in the GWAS Catalog and Million Veteran Program (MVP)^26,38,39, 34 out of the 51 associations remained significant at the Bonferroni-corrected threshold of α = 0.05/51 = 9.80 × 10⁻⁴ (Table 1).

Table 1 ∣. TOPMed Gene-centric coding multi-trait analysis results of both unconditional analysis and analysis conditional on known lipids-associated variants.

A total of 61,838 samples from the TOPMed Program were considered in the analysis. Results for the conditionally significant genes (unconditional MultiSTAAR-O P < 5.00 × 10⁻⁷; conditional MultiSTAAR-O P < 9.80 × 10⁻⁴) are presented in the table. MultiSTAAR-O is a two-sided test. Chr. no., chromosome number; Category, functional category; No. of SNVs, number of rare variants (MAF < 1%) of the particular coding functional category in the gene; MultiSTAAR-O, MultiSTAAR-O P value; Variants (adjusted), adjusted variants in the conditional analysis.

Gene	Chr. no.	Category	No. of SNVs	MultiSTAAR-O (Unconditional)	MultiSTAAR-O (Conditional)	Variants (adjusted)
PCSK9	1	Putative loss-of-function	14	1.14E-115	2.66E-08	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
APOB	2	Putative loss-of-function	29	8.04E-28	5.76E-27	rs12478327, rs72654432, rs1042034, rs676210, rs533617, rs17240441, rs34722314, rs563290, rs10692845
ABCA1	9	Putative loss-of-function	28	2.04E-21	5.41E-21	rs2150867, rs33918808, rs112853430, rs4149307, rs9282541, rs1883025, rs1800978
LDLR	19	Putative loss-of-function	19	8.81E-21	7.16E-21	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459
PCSK9	1	Missense	271	8.94E-71	1.29E-10	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
APOB	2	Missense	1407	5.57E-08	4.31E-08	rs12478327, rs72654432, rs1042034, rs676210, rs533617, rs17240441, rs34722314, rs563290, rs10692845
ABCG5	2	Missense	242	5.75E-08	9.81E-08	rs114780578, rs11887534, rs4245791
NPC1L1	7	Missense	477	3.10E-08	1.60E-07	rs217381
LPL	8	Missense	149	9.57E-19	7.14E-04	rs6996383, rs268, rs328, rs3289, rs13702, rs15285, rs78810414, rs28550053, rs12676079, rs55682243
ABCA1	9	Missense	597	3.63E-46	1.75E-33	rs2150867, rs33918808, rs112853430, rs4149307, rs9282541, rs1883025, rs1800978
SCARB1	12	Missense	192	6.77E-15	3.55E-15	rs6488913, rs4765127, rs1716407, rs825456, rs1672875, rs10846744, rs10773112, rs187471874, rs10773119
LIPC	15	Missense	246	2.54E-20	6.66E-15	rs1973688, rs1601935, rs2043082, rs10468017, rs1532085, rs436965, rs35980001, rs1800588, rs2070895, rs113298164
CETP	16	Missense	168	8.84E-14	2.09E-04	rs35571500, rs247617, rs17231506, rs34498052, rs34119551, rs34065661, rs1597000001*, rs7499892, rs5883, rs289719, rs11860407, rs189866004, rs5880
LCAT	16	Missense	107	9.18E-14	3.06E-17	rs111315946, rs150660813, rs4986970, rs35673026, rs1109166, rs548291389
LDLR	19	Missense	342	7.92E-58	2.12E-57	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459
TM6SF2	19	Missense	120	7.06E-08	6.16E-07	rs3761077, rs150641967, rs187429064, rs2074304
PCSK9	1	Putative loss-of-function and disruptive missense	71	1.14E-107	8.22E-17	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
APOB	2	Putative loss-of-function and disruptive missense	75	9.96E-12	9.86E-12	rs12478327, rs72654432, rs1042034, rs676210, rs533617, rs17240441, rs34722314, rs563290, rs10692845
NPC1L1	7	Putative loss-of-function and disruptive missense	303	1.79E-09	8.29E-09	rs217381
ABCA1	9	Putative loss-of-function and disruptive missense	357	7.85E-33	2.66E-33	rs2150867, rs33918808, rs112853430, rs4149307, rs9282541, rs1883025, rs1800978
APOC3	11	Putative loss-of-function and disruptive missense	15	2.86E-126	3.01E-06	rs509728, rs61905072, rs66505542, rs7102314, rs964184, rs75198898, rs142958146, rs2075291, rs3135506, rs651821, rs45611741, rs662799, rs10750097, rs9804646, rs978880643, rs2070669, rs76353203, rs138326449, rs147210663, rs140621530, rs525028, rs141469619, rs188287950, rs202207736
SCARB1	12	Putative loss-of-function and disruptive missense	60	3.49E-17	2.14E-17	rs6488913, rs4765127, rs1716407, rs825456, rs1672875, rs10846744, rs10773112, rs187471874, rs10773119
LIPC	15	Putative loss-of-function and disruptive missense	130	1.01E-19	1.49E-17	rs1973688, rs1601935, rs2043082, rs10468017, rs1532085, rs436965, rs35980001, rs1800588, rs2070895, rs113298164
LCAT	16	Putative loss-of-function and disruptive missense	88	2.38E-16	5.07E-17	rs111315946, rs150660813, rs4986970, rs35673026, rs1109166, rs548291389
LDLR	19	Putative loss-of-function and disruptive missense	221	6.97E-72	1.57E-71	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459
PCSK9	1	Disruptive missense	57	7.03E-19	1.33E-12	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
APOB	2	Disruptive missense	46	5.78E-09	4.48E-09	rs12478327, rs72654432, rs1042034, rs676210, rs533617, rs17240441, rs34722314, rs563290, rs10692845
NPC1L1	7	Disruptive missense	276	3.34E-09	1.57E-08	rs217381
ABCA1	9	Disruptive missense	329	1.17E-22	1.59E-23	rs2150867, rs33918808, rs112853430, rs4149307, rs9282541, rs1883025, rs1800978
APOC3	11	Disruptive missense	6	2.38E-29	3.93E-04	rs509728, rs61905072, rs66505542, rs7102314, rs964184, rs75198898, rs142958146, rs2075291, rs3135506, rs651821, rs45611741, rs662799, rs10750097, rs9804646, rs978880643, rs2070669, rs76353203, rs138326449, rs147210663, rs140621530, rs525028, rs141469619, rs188287950, rs202207736
SCARB1	12	Disruptive missense	51	4.44E-16	2.86E-16	rs6488913, rs4765127, rs1716407, rs825456, rs1672875, rs10846744, rs10773112, rs187471874, rs10773119
LIPC	15	Disruptive missense	112	2.19E-18	2.65E-16	rs1973688, rs1601935, rs2043082, rs10468017, rs1532085, rs436965, rs35980001, rs1800588, rs2070895, rs113298164
LCAT	16	Disruptive missense	84	2.85E-14	6.44E-15	rs111315946, rs150660813, rs4986970, rs35673026, rs1109166, rs548291389
LDLR	19	Disruptive missense	203	2.22E-59	5.13E-59	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459

^*Samoan-specific missense variant.

For non-coding variants, rare variants from eight noncoding masks were analyzed in a similar fashion, including (1) promoter rare variants overlaid with CAGE sites⁴⁰, (2) promoter rare variants overlaid with DHS sites⁴¹, (3) enhancer rare variants overlaid with CAGE sites^42,43, (4) enhancer rare variants overlaid with DHS sites^41,43, (5) untranslated region (UTR) rare variants, (6) upstream region rare variants, (7) downstream region rare variants of each protein-coding gene and (8) rare variants in ncRNA genes²⁴. The promoter rare variants were defined as rare variants in the ±3-kilobase (kb) window of transcription start sites with the overlap of CAGE sites or DHS sites. The enhancer rare variants were defined as RVs in GeneHancer-predicted regions with the overlap of CAGE sites or DHS sites. The UTR, upstream, downstream and ncRNA rare variants were defined by GENCODE VEP categories³². With a well-calibrated overall distribution of MultiSTAAR-O P values (Extended Data Fig. 1d) and at a Bonferroni-corrected significance threshold of α = 0.05/(20,000 × 7) = 3.57 × 10⁻⁷, accounting for seven different noncoding masks across protein-coding genes, MultiSTAAR-O identified 76 genome-wide significant associations using unconditional multi-trait analysis (Extended Data Fig. 1c, Supplementary Table 5). After conditioning on known lipids-associated variants^26,38,39, 6 out of the 76 associations remained significant at the Bonferroni-corrected threshold of α = 0.05/76 = 6.58 × 10⁻⁴ (Table 2). These included promoter CAGE and enhancer CAGE rare variants in APOA1, promoter DHS rare variants in CETP, enhancer CAGE rare variants in SPC24, and enhancer DHS rare variants in NIPSNAP3A and LIPC.

Table 2 ∣. TOPMed Gene-centric noncoding multi-trait analysis results of both unconditional analysis and analysis conditional on known lipids-associated variants.

A total of 61,838 samples from the TOPMed Program were considered in the analysis. Results for the conditionally significant genes (unconditional MultiSTAAR-O P < 3.57 × 10⁻⁷ and conditional MultiSTAAR-O P < 6.58 × 10⁻⁴ for 7 different noncoding masks across protein-coding genes; unconditional MultiSTAAR-O P < 2.50 × 10⁻⁶ and conditional MultiSTAAR-O P < 8.33 × 10⁻³ for ncRNA genes) are presented in the table. MultiSTAAR-O is a two-sided test. Chr. no., chromosome number; Category, functional category; No. of SNVs, number of rare variants (MAF < 1%) of the particular noncoding functional category in the gene; MultiSTAAR-O, MultiSTAAR-O P value; Variants (adjusted), adjusted variants in the conditional analysis; n/a, no variant adjusted in the conditional analysis.

Gene	Chr. no.	Category	No. of SNVs	MultiSTAAR-O (Unconditional)	MultiSTAAR-O (Conditional)	Variants (adjusted)
APOA1	11	Promoter (CAGE)	230	2.33E-07	9.45E-07	rs509728, rs61905072, rs66505542, rs7102314, rs964184, rs75198898, rs142958146, rs2075291, rs3135506, rs651821, rs45611741, rs662799, rs10750097, rs9804646, rs978880643, rs2070669, rs76353203, rs138326449, rs147210663, rs140621530, rs525028, rs141469619, rs188287950, rs202207736
CETP	16	Promoter (DHS)	411	1.21E-12	5.75E-04	rs35571500, rs247617, rs17231506, rs34498052, rs34119551, rs34065661, rs1597000001*, rs7499892, rs5883, rs289719, rs11860407, rs189866004, rs5880
APOA1	11	Enhancer (CAGE)	642	1.88E-24	6.23E-04	rs509728, rs61905072, rs66505542, rs7102314, rs964184, rs75198898, rs142958146, rs2075291, rs3135506, rs651821, rs45611741, rs662799, rs10750097, rs9804646, rs978880643, rs2070669, rs76353203, rs138326449, rs147210663, rs140621530, rs525028, rs141469619, rs188287950, rs202207736
SPC24	19	Enhancer (CAGE)	366	1.33E-08	4.88E-04	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459
NIPSNAP3A	9	Enhancer (DHS)	767	2.63E-08	8.46E-06	rs2150867, rs33918808, rs112853430, rs4149307, rs9282541, rs1883025, rs1800978
LIPC	15	Enhancer (DHS)	3714	4.26E-08	1.25E-04	rs1973688, rs1601935, rs2043082, rs10468017, rs1532085, rs436965, rs35980001, rs1800588, rs2070895, rs113298164
RP11-310H4.2	7	ncRNA	154	1.69E-06	1.69E-06	n/a
MIR4497	12	ncRNA	23	1.37E-06	1.42E-06	rs5800864
RP11-15F12.3	18	ncRNA	64	7.53E-11	7.50E-03	rs77960347, rs117623631, rs9958734, rs7229562, rs8086351, rs10048323, rs8084172

^*Samoan-specific missense variant.

MultiSTAAR-O further identified 6 genome-wide significant associations using unconditional multi-trait analysis at α = 0.05/20,000 = 2.50 × 10⁻⁶ accounting for ncRNA genes (Extended Data Fig. 1e, Supplementary Table 5), with 3 rare variant associations in RP11-15F12.3, RP11-310H4.2 and MIR4497 remained significant at α = 0.05/6 = 8.33 × 10⁻³ after conditioning on known lipids-associated variants^26,38,39 (Table 2).

Notably, among the 9 conditionally significant noncoding rare variants associations with lipid traits, 4 of them were not detected by any of the three single-trait analysis (LDL-C, HDL-C or TG) using unconditional analysis of STAAR-O, including the associations of enhancer DHS rare variants in NIPSNAP3A and LIPC as well as ncRNA rare variants in RP11-310H4.2 and MIR4497 (Supplementary Table 5). These results demonstrate that MultiSTAAR-O can increase power over existing methods, and identify additional trait-associated signals by leveraging cross-phenotype correlations between multiple traits.

Genetic region multi-trait analysis of rare variants

We next applied MultiSTAAR-O to perform genetic region multi-trait analysis to identify rare variants associated with lipid traits in TOPMed. Rare variants residing in 2-kilobase (kb) sliding windows with a 1-kb skip length were aggregated and analyzed using a joint model for LDL-C, HDL-C and TG. We incorporated 12 quantitative annotations, including 9 aPCs, CADD, LINSIGHT, FATHMM-XF along with the two MAF weights in MultiSTAAR-O (Methods). The overall distribution of MultiSTAAR-O P values was well-calibrated for the multi-trait analysis (Fig. 2b). At a Bonferroni-corrected significance threshold of α = 0.05/(2.65 × 10⁶) = 1.89 × 10⁻⁸ accounting for 2.65 million 2-kb sliding windows across the genome, MultiSTAAR-O identified 502 genome-wide significant associations using unconditional multi-trait analysis (Fig. 2a, Supplementary Table 6). By dynamically incorporating multiple functional annotations capturing different aspects of variant function, MultiSTAAR-O detected more significant sliding windows and showed consistently smaller P values for top sliding windows compared with multi-trait analysis using only MAFs as the weight (Fig. 2c). After conditioning on known lipids-associated variants^26,38,39, 7 out of the 502 associations remained significant at the Bonferroni-corrected threshold of α 0.05/502 = 9.96 × 10⁻⁵ (Table 3), including two sliding windows in DOCK7 (chromosome 1: 62,651,447 - 62,653,446 bp; chromosome 1: 62,652,447 - 62,654,446 bp) and an intergenic sliding window (chromosome 1: 145,530,447 - 145,532,446 bp) that were not detected by any of the three single-trait analysis (LDL-C, HDL-C or TG) using STAAR-O (Supplementary Table 6). Notably, all known lipids-associated variants indexed in the previous literature were at least 1-Mb away from the intergenic sliding window.

Fig. 2 ∣. TOPMed Genetic region (2-kb sliding window) unconditional multi-trait analysis results of low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and triglycerides (TG) using TOPMed data.

a, Manhattan plot showing the associations of 2.65 million 2-kb sliding windows versus −log₁₀(P) of MultiSTAAR-O. The horizontal line indicates a genome-wide P value threshold of 1.89 × 10⁻⁸ (n = 61,838). b, Quantile-quantile plot of 2-kb sliding window MultiSTAAR-O P values (n = 61,838). c, Scatterplot of P values for the 2-kb sliding windows comparing MultiSTAAR-O with Burden-MT, SKAT-MT and ACAT-V-MT tests (MT is short for Multi-Trait). Each dot represents a sliding window with x-axis label being the −log₁₀(P) of the conventional multi-trait test and y-axis label being the −log₁₀(P) of MultiSTAAR-O (n = 61,838). Burden-MT, SKAT-MT, ACAT-V-MT and MultiSTAAR-O are two-sided tests. Int*, intergenic sliding window.

Table 3 ∣. TOPMed Genetic region (2-kb sliding window) multi-trait analysis results of both unconditional analysis and analysis conditional on known lipid-associated variants.

A total of 61,838 samples from the TOPMed Program were considered in the analysis. Results for the conditionally significant sliding windows (unconditional MultiSTAAR-O P < 1.89 × 10⁻⁸ and conditional MultiSTAAR-O P < 9.96 × 10⁻⁵) are presented in the table. MultiSTAAR-O is a two-sided test. Chr. no., chromosome number; Start location, start location of the 2-kb sliding window; End location, end location of the 2-kb sliding window; No. of SNVs, number of rare variants (MAF < 1%) in the 2-kb sliding window; MultiSTAAR-O, MultiSTAAR-O P value; Variants (adjusted), adjusted variants in the conditional analysis; n/a, no variant adjusted in the conditional analysis. Physical positions of each window are on build hg38.

Chr. no.	Start location	End location	Gene	No. of SNVs	MultiSTAAR-O (Unconditional)	MultiSTAAR-O (Conditional)	Variants (adjusted)
1	55,051,447	55,053,446	PCSK9	327	7.11E-11	6.60E-08	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
1	55,052,447	55,054,446	PCSK9	320	9.37E-09	9.07E-06	rs12117661, rs2495491, rs11591147, rs67608943, rs72646508, rs693668, rs28362261, rs28362263, rs141502002, rs505151, rs28362286
1	62,651,447	62,653,446	DOCK7	277	5.08E-09	7.56E-10	rs67461605
1	62,652,447	62,654,446	DOCK7	257	4.87E-09	7.24E-10	rs67461605
1	145,530,447	145,532,446	intergenic	233	5.12E-09	5.12E-09	n/a
19	11,104,367	11,106,366	LDLR	336	1.15E-12	8.33E-13	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459
19	11,105,367	11,107,366	LDLR	338	5.97E-14	5.55E-15	rs140753491, rs138294113, rs17242353, rs17242843, rs10422256, rs72658860, rs11669576, rs2738447, rs72658867, rs2738464, rs6511728, rs3760782, rs59168178, rs2278426, rs112942459

^*Samoan-specific missense variant.

Comparison of MultiSTAAR-O with existing multi-trait rare variant tests

Using TOPMed Freeze 8 WGS data, our gene-centric multi-trait analysis of coding rare variants identified 34 conditionally significant associations with lipid traits (Table 1), including NPC1L1 and SCARB1 missense rare variants that were missed by multi-trait burden, SKAT and ACAT-V tests (Supplementary Table 4). Among the 9 and 7 conditionally significant associations detected in gene-centric multi-trait analysis of noncoding rare variants and genetic region multi-trait analysis, MultiSTAAR-O identified 1 and 2 associations, respectively, that were missed by multi-trait burden, SKAT and ACAT-V tests (Supplementary Tables 5-6). These associations included enhancer CAGE rare variants in SPC24 and two sliding windows in LDLR (chromosome 19: 11,104,367 - 11,106,366 bp; chromosome 19: 11,105,367 - 11,107,366 bp).

Computation cost

The computational cost for MultiSTAAR-O to perform WGS multi-trait rare variant analysis of n = 61,838 related TOPMed lipids samples was 2 hours using 250 2.10-GHz computing cores with 12-GB memory for gene-centric coding analysis; or 20 hours using 250 2.10-GHz computing cores with 24-GB memory for gene-centric noncoding analysis; 2 hours using 250 2.10-GHz computing cores with 12-GB memory of ncRNA analysis; and 20 hours using 500 2.10-GHz computing cores with 24-GB memory for sliding window analysis. Runtime for all analyses scales linearly with the sample size²⁴.

Discussion

In this study, we have introduced MultiSTAAR as a general statistical framework and a flexible analytical pipeline for performing functionally-informed multi-trait RVAS in large-scale WGS studies. MultiSTAAR improves power by analyzing multiple traits simultaneously and dynamically incorporating multiple functional annotations, while accounting for relatedness and population structure among study samples.

By jointly analyzing multiple quantitative traits using a multivariate linear mixed model, MultiSTAAR explicitly leverages the correlation among multiple phenotypes to enhance power for detecting additional association signals, outperforming single-trait analyses of the individual phenotypes. MultiSTAAR also enables conditional multi-trait analysis to identify putatively novel rare variant associations independent of a set of known variants. Using TOPMed Freeze 8 WGS data, our gene-centric multi-trait analysis of noncoding rare variants identified 9 conditionally significant associations with lipid traits (Table 2), including 4 noncoding associations that were missed by single-trait analysis using STAAR (Supplementary Table 5). Our genetic region multi-trait analysis of rare variants identified 7 conditionally significant 2-kb sliding windows associated with lipid traits (Table 3), including 3 associations that were missed by single-trait analysis using STAAR (Supplementary Table 6).

By dynamically incorporating multiple annotations capturing diverse aspects of variant biological function in the second step, MultiSTAAR further improves power over existing multi-trait rare variant analysis methods. Our simulation studies demonstrated that MultiSTAAR-O maintained accurate type I error rates while achieving considerable power gains over multi-trait burden, SKAT and ACAT-V tests that do not incorporate functional annotation information (Extended Data Figs. 2-5, Supplementary Figs. 1-4). Notably, the existing ACAT-V method⁹ does not support multi-trait analysis. We extended it to accommodate multi-trait settings and incorporated the multi-trait ACAT-V test into the MultiSTAAR framework (Methods).

Implemented as a flexible analytical pipeline, MultiSTAAR allows for customized input phenotype selection, variant set definition and user-specified annotation weights to facilitate functionally-informed multi-trait analyses. In addition to rare variant association analysis of coding and noncoding regions, MultiSTAAR also provides single-variant multi-trait analysis for common and low-frequency variants under a given MAF or minor allele count (MAC) cutoff (e.g. MAC ≥ 20). Using 61,838 TOPMed lipids samples, it took 8 hours using 250 2.10-GHz computing cores with 12-GB memory for single-variant multi-trait analysis, which is scalable for large WGS/WES datasets. On the other hand, MultiSTAAR could be further extended to allow for dynamic windows with data-adaptive sizes in genetic region analysis^24,44, to properly leverage synthetic surrogates in the presence of partially missing phenotypes⁴⁵, and to incorporate summary statistics for meta-analysis of multiple WGS/WES studies⁴⁶.

In summary, MultiSTAAR provides a powerful statistical framework and a computationally scalable analytical pipeline for large-scale WGS multi-trait analysis with complex study samples. Compared to single-trait analysis, MultiSTAAR offers a notable increase in statistical power when analyzing multiple moderately to highly correlated traits, all while maintaining control over type I error rates across various genetic architectures. As the sample sizes and number of available phenotypes increase in biobank-scale sequencing studies, our proposed method may contribute to a better understanding of the genetic architecture of complex traits by elucidating the role of rare variants with pleiotropic effects.

Methods

Ethics statement

This study relied on analyses of genetic data from TOPMed cohorts. The study has been approved by the TOPMed Publications Committee, TOPMed Lipids Working Group and all the participating cohorts, including Old Order Amish (phs000956.v1.p1), Atherosclerosis Risk in Communities Study (phs001211), Mt Sinai BioMe Biobank (phs001644), Coronary Artery Risk Development in Young Adults (phs001612), Cleveland Family Study (phs000954), Cardiovascular Health Study (phs001368), Diabetes Heart Study (phs001412), Framingham Heart Study (phs000974), Genetic Study of Atherosclerosis Risk (phs001218), Genetic Epidemiology Network of Arteriopathy (phs001345), Genetic Epidemiology Network of Salt Sensitivity (phs001217), Genetics of Lipid Lowering Drugs and Diet Network (phs001359), Hispanic Community Health Study - Study of Latinos (phs001395), Hypertension Genetic Epidemiology Network and Genetic Epidemiology Network of Arteriopathy (phs001293), Jackson Heart Study (phs000964), Multi-Ethnic Study of Atherosclerosis (phs001416), San Antonio Family Heart Study (phs001215), Genome-wide Association Study of Adiposity in Samoans (phs000972), Taiwan Study of Hypertension using Rare Variants (phs001387), and Women’s Health Initiative (phs001237), where the accession numbers are provided in parenthesis. The use of human genetics data from TOPMed cohorts was approved by the Harvard T.H. Chan School of Public Health IRB (IRB13-0353).

Notation and model

Suppose there are $n$ subjects with a total of $M$ variants sequenced across the whole genome. For the i-th subject, let ${\mathit{Y}}_i={(y_{i1},y_{i2},\dots ,y_{iK})}^T$ denote a vector of K quantitative phenotypes; ${\mathit{X}}_i={(x_{i1},x_{i2},\dots ,x_{iq})}^T$ denotes $q$ covariates, such as age, gender and ancestral principal components; ${\mathit{G}}_i={(G_{i1},G_{i2},\dots ,G_{ip})}^T$ denotes the genotype matrix of the $p$ genetic variants in a variant set. Since these K phenotypes may be defined on different measurement scales, we assume that each phenotype has been rescaled to have zero mean and unit variance.

When the data consist of unrelated samples, we consider the following Multivariate Linear Model (MLM)

$${\mathit{Y}}_i=\left[\begin{array}{c}\hfill y_{i1}\hfill \\ \hfill y_{i2}\hfill \\ \hfill ⋮\hfill \\ \hfill y_{ik}\hfill \end{array}\right]=\left[\begin{array}{c}\hfill {\alpha }_{0,1}+{\mathit{X}}_i^T{\alpha }_1+{\mathit{G}}_i^T{\beta }_1\hfill \\ \hfill {\alpha }_{0,2}+{\mathit{X}}_i^T{\alpha }_2+{\mathit{G}}_i^T{\beta }_2\hfill \\ \hfill ⋮\hfill \\ \hfill {\alpha }_{0,K}+{\mathit{X}}_i^T{\alpha }_K+{\mathit{G}}_i^T{\beta }_K\hfill \end{array}\right]+\left[\begin{array}{c}\hfill {\epsilon }_{i1}\hfill \\ \hfill {\epsilon }_{i2}\hfill \\ \hfill ⋮\hfill \\ \hfill {\epsilon }_{iK}\hfill \end{array}\right],$$ (1)

where ${\alpha }_{0,k}$ is an intercept, ${\alpha }_k={({\alpha }_{1,k},{\alpha }_{2,k},\dots ,{\alpha }_{q,k})}^T$ and ${\beta }_k={({\beta }_{1,k},{\beta }_{2,k},\dots ,{\beta }_{p,k})}^T$ are column vectors of regression coefficients for covariates ${\mathit{X}}_i$ and genotype ${\mathit{G}}_i$ in phenotype $k$, respectively. The error terms ${\epsilon }_i={({\epsilon }_{i1},{\epsilon }_{i2},\dots ,{\epsilon }_{iK})}^T$ are independent and identically distributed and follow a multivariate normal distribution with mean a vector of zeros and variance-covariance matrix ${\Sigma }_{K\times K}$, assumed identical for all subjects. For all $n$ subjects, using matrix notation we can write model (1) as

$${\mathit{Y}}_{n\times K}=1_n{\alpha }_0^T+{\mathit{X}}_{n\times q}{\alpha }_{q\times K}+{\mathit{G}}_{n\times p}{\beta }_{p\times K}+{\epsilon }_{n\times K},\#$$ (2)

where $1_n$ is a column vector of 1’s with length $n$, ${\alpha }_0={({\alpha }_{0,1},{\alpha }_{0,2},\dots ,{\alpha }_{0,K})}^T$ is a column vector of regression intercepts, the $k$-th columns of ${\alpha }_{q\times K}$ and ${\beta }_{p\times K}$ are ${\alpha }_k$ and ${\beta }_k$, respectively, and ${\epsilon }_{n\times K}={({\epsilon }_1,{\epsilon }_2,\dots ,{\epsilon }_n)}^T\sim {\text{MatrixNormal}}_{n,K}(0_{n\times K},{\mathbf{I}}_{n\times n},{\Sigma }_{K\times K})$ follows a matrix normal distribution. We calculate the scaled residual for each subject on each phenotype, defined as ${\widehat{\mathit{e}}}_{n\times K}=({\mathit{Y}}_{n\times K}-{\widehat{\mu }}_{n\times K}){\widehat{\Sigma }}_{K\times K}^{-1}$, where ${\widehat{\mu }}_{n\times K}$ (a matrix of fitted values) and ${\widehat{\Sigma }}_{K\times K}$ are estimated under the null MLM ${\mathit{Y}}_{n\times K}=1_n{\alpha }_0^T+{\mathit{X}}_{n\times q}{\alpha }_{q\times K}+{\epsilon }_{n\times K}$, where no variant has any effect on any outcome.

When the data consist of related samples, we consider the following Multivariate Linear Mixed Model (MLMM)^19,47,48

$${\mathit{Y}}_i=\left[\begin{array}{c}\hfill y_{i1}\hfill \\ \hfill y_{i2}\hfill \\ \hfill ⋮\hfill \\ \hfill y_{ik}\hfill \end{array}\right]=\left[\begin{array}{c}\hfill {\alpha }_{0,1}+{\mathit{X}}_i^T{\alpha }_1+{\mathit{G}}_i^T{\beta }_1\hfill \\ \hfill {\alpha }_{0,2}+{\mathit{X}}_i^T{\alpha }_2+{\mathit{G}}_i^T{\beta }_2\hfill \\ \hfill ⋮\hfill \\ \hfill {\alpha }_{0,K}+{\mathit{X}}_i^T{\alpha }_K+{\mathit{G}}_i^T{\beta }_K\hfill \end{array}\right]+\left[\begin{array}{c}\hfill b_{i1}\hfill \\ \hfill b_{i2}\hfill \\ \hfill ⋮\hfill \\ \hfill b_{iK}\hfill \end{array}\right]\left[\begin{array}{c}\hfill b_{i1}\hfill \\ \hfill {\epsilon }_{i2}\hfill \\ \hfill ⋮\hfill \\ \hfill {\epsilon }_{iK}\hfill \end{array}\right],\#$$ (3)

where the random effects $b_{ik}$ account for relatedness and remaining population structure unaccounted by ancestral PCs²⁰. We assume that ${\mathit{b}}_{n\times K}={(b_{ik})}_{n\times K}\sim {\text{MatrixNormal}}_{n,K}(0_{n\times K},{\Phi }_{n\times n},{\Theta }_{K\times K})$ with a variance component matrix ${\Theta }_{K\times K}$ and a sparse genetic relatedness matrix ${\Phi }_{n\times n}$^21,22. For all $n$ subjects, using matrix notation we can rewrite equation (3) as

$${\mathit{Y}}_{n\times K}=1_n{\alpha }_0^T+{\mathit{X}}_{n\times q}{\alpha }_{q\times K}+{\mathit{G}}_{n\times p}{\beta }_{p\times K}+{\mathit{b}}_{n\times K}+{\epsilon }_{n\times K}.\#$$ (4)

We calculate the scaled residual for each subject on each phenotype, defined as ${\widehat{\mathit{e}}}_{n\times K}=({\mathit{Y}}_{n\times K}-{\widehat{\mu }}_{n\times K}){\widehat{\Sigma }}_{K\times K}^{-1}$, where ${\widehat{\mu }}_{n\times K}$ and ${\widehat{\Sigma }}_{K\times K}$ are estimated under the null MLMM ${\mathit{Y}}_{n\times K}=1_n{\alpha }_0^T+{\mathit{X}}_{n\times q}{\alpha }_{q\times K}+{\mathit{b}}_{n\times K}+{\epsilon }_{n\times K}$. Under both MLM and MLMM, our goal is to test for an association between a set of $p$ genetic variants and $K$ quantitative phenotypes, adjusting for covariates and relatedness. This corresponds to testing $H_0:{\beta }_1={\beta }_2=\cdots {\beta }_K=0$.

Multi-trait rare variant association tests using MultiSTAAR

Single-trait score-based aggregation methods^5-9 can be extended to allow for jointly testing the association between rare variants in a variant set and multiple quantitative phenotypes. For a given variant set, let ${\mathit{S}}_{p\times K}={(S_{jk})}_{p\times K}={({\mathit{G}}_{n\times p})}^T{\widehat{\mathit{e}}}_{n\times K}$ denote the matrix of score statistics where $S_{jk}$ is the score statistic for the $j$-th variant on the $k$-th phenotype. For multi-trait burden test using MultiSTAAR (Burden-MT), we consider test statistic

$$Q_{Burden-MT}=\left(\sum _{j=1}^p{w}_j{\mathit{S}}_{j\cdot }\right){\widehat{\mathit{V}}}^{-1}{\left(\sum _{j=1}^p{w}_j{\mathit{S}}_{j\cdot }\right)}^{{}^{{}^{{}^{{}^{{}^T}}}}},$$

where $w_j$ is the weight defined as a function of the MAF for the $j$-th variant^4,18, ${\mathit{S}}_{j\cdot }=(S_{j1},S_{j2},\dots ,S_{jK})$ is the $j$-th row of $\mathit{S}$ and $\widehat{\mathit{V}}$ is the estimated variance-covariance matrix of ${\sum }_{j=1}^p{w}_j{\mathit{S}}_{j\cdot }={\mathit{w}}^T\mathit{S}$. $Q_{Burden-MT}$ asymptotically follows a standard chi-square distribution with $K$ degrees of freedom under the null hypothesis, and its P value can be obtained analytically while accounting for LD between variants and correlation between phenotypes.

For multi-trait SKAT using MultiSTAAR (SKAT-MT), we consider the statistic

$$Q_{SKAT-MT}=\sum _{k=1}^K\sum _{j=1}^p{w}_j^2{S}_{jk}^2.$$

$Q_{SKAT-MT}$ asymptotically follows a mixture of chi-square distributions under the null hypothesis, and its P value can be obtained analytically while accounting for LD between variants and correlation between phenotypes^14,15.

For multi-trait ACAT-V using MultiSTAAR (ACAT-V-MT), we propose test statistic

$$\begin{gathered} Q_{ACAT-V-MT}=\overline{w^2\text{MAF}(1-\text{MAF})}\tan ((0.5-p_0)\pi ) \\ +\sum _{j=1}^{p^{\prime }}{w}_j^2{\text{MAF}}_j(1-{\text{MAF}}_j)\tan \left((0.5-p_j)\pi \right), \end{gathered}$$

where $p^{{}^{{}_{\prime }}}$ is the number of variants with a minor allele count (MAC) greater than 10 and $p_j$ is the multi-trait association P value of individual variant $j$ for those variants with a MAC > 10, whose test statistic is given by the K degrees of freedom multivariate score test

$$Q_j={\mathit{S}}_{j\cdot }{\widehat{\mathit{V}}}_{{\mathit{S}}_{j\cdot }}^{-1}{\mathit{S}}_{j\cdot }^T,$$

where ${\widehat{\mathit{V}}}_{{\mathit{S}}_{j\cdot }}$. is the estimated variance-covariance matrix of ${\mathit{S}}_{j\cdot }$; $p_0$ is the multi-trait burden test P value of extremely rare variants with an MAC ≤ 10 as described above and $\overline{w^2\text{MAF}(1-\text{MAF})}$ is the average of the weights $w_j^2{\text{MAF}}_j(1-{\text{MAF}}_j)$ among the extremely rare variants with an MAC ≤ 10. $Q_{ACAT-V-MT}$ is approximated well by a scaled Cauchy distribution under the null hypothesis, and its P value can be obtained analytically while accounting for LD between variants and correlation between phenotypes^9,49. Note that when K = 1, the multi-trait burden, SKAT, and ACAT-V tests reduce to the original single-trait burden, SKAT and ACAT-V tests.

Suppose we have a collection of $L$ annotations, let $A_{jl}$ denote the $l$-th annotation for the jth variant in the variant set. We define the functionally-informed multi-trait burden, SKAT and ACAT-V test statistics weighted by the $l$-th annotation as follows

$$\begin{gathered} Q{}_{Burden-MT,l,(a_1,a_2})=\left(\sum _{j=1}^p{\widehat{\pi }}_{jl}{w}_{j,(a_1,a_2)}{\mathit{S}}_{j\cdot }\right){\widehat{\mathit{V}}}_{l,(a_1,a_2)}^{-1}{\left(\sum _{j=1}^p{\widehat{\pi }}_{jl}w{}_{j,(a_1,a_2}){\mathit{S}}_{j\cdot }\right)}^T, \\ {Q}_{SKAT-MT,l,(a_1,a_2)}=\sum _{k=1}^K\sum _{j=1}^p{\widehat{\pi }}_{jl}{w}_{j,(a_1,a_2)}^2{S}_{jk}^2, \\ Q{}_{ACAT-V-MT,l,(a_1,a_2}) \\ =\overline{{\widehat{\pi }}_{\cdot l}{w}_{(a_1,a_2)}^2\text{MAF}(1-\text{MAF})}\tan \left((0.5-p_{0,l})\pi \right) \\ +\sum _{j=1}^{M^{\prime }}{\widehat{\pi }}_{jl}{w}_{j,(a_1,a_2)}^2{\text{MAF}}_j(1-{\text{MAF}}_j)\tan \left((0.5-p_j)\pi \right), \end{gathered}$$

where ${\widehat{\pi }}_{jl}=\frac{\text{rank}(A_{jl})}{M}$, $w_{j,(a_1,a_2)}=Beta({\text{MAF}}_j;a_1,a_2)$ with $(a_1,a_2)\in 𝒜=\{(1,25)(1,1)\}$, ${\widehat{\mathit{V}}}_{l,(a_1,a_2)}$ is the estimated variance-covariance matrix of ${\sum }_{j=1}^p{\widehat{\pi }}_{jl}{w}_{j,(a_1,a_2)}{\mathit{S}}_{j\cdot }$. and $\overline{{\widehat{\pi }}_{\cdot l}{w}_{(a_1,a_2)}^2\text{MAF}(1-\text{MAF})}$ is the average of the weights ${\widehat{\pi }}_{jl}{w}_{j,(a_1,a_2)}^2{\text{MAF}}_j(1-{\text{MAF}}_j)$ among the extremely rare variants with MAC ≤ 10. Finally, we define the omnibus MultiSTAAR-O test statistic as

$$\begin{gathered} T_{MultiSTAAR-O}=\frac{1}{3\mid 𝒜\mid }\sum _{(a_1,a_2)\in 𝒜}[T_{MultiSTAAR-B(a_1,a_2})+T_{MultiSTAAR-S(a_1,a_2})\phantom{]} \\ {T}_{MultiSTAAR-A(a_1,a_2)}] \\ =\frac{1}{3\mid 𝒜\mid }\sum _{(a_1,a_2)\in 𝒜}\sum _{l=0}^L[\frac{\tan \{(0.5-p_{Burden-MT,l,(a_1,a_2)})\pi \}}{L+1}\phantom{]} \\ \phantom{[}+\frac{\tan \{(0.5-p_{SKAT-MT,l,(a_1,a_2)})\pi \}}{L=1}+\frac{\tan \{(0.5-p_{ACAT-V-MT,l,(a_1,a_2)})\pi \}}{L=1}], \end{gathered}$$

and the P value of $T_{MultiSTAAR-O}$ can be calculated by

$$p_{MultiSTAAR-O}=\frac{1}{2}-\frac{\{\text{arctan}(T_{MultiSTAAR-O})\}}{\pi }.$$

Data simulation

Type I error rate simulations

We performed simulation studies to evaluate how accurately MultiSTAAR controls the type I error rate. We generated three quantitative traits from a multivariate linear model, conditional on two covariates

$${\mathit{Y}}_i=\left[\begin{array}{c}\hfill Y_{i1}\hfill \\ \hfill Y_{i2}\hfill \\ \hfill Y_{i3}\hfill \end{array}\right]=\left[\begin{array}{c}\hfill 0.5X_{i1}+0.5X_{i2}\hfill \\ \hfill 0.5X_{i1}+0.5X_{i2}\hfill \\ \hfill 0.5X_{i1}+0.5X_{i2}\hfill \end{array}\right]+\left[\begin{array}{c}\hfill {\epsilon }_{i1}\hfill \\ \hfill {\epsilon }_{i2}\hfill \\ \hfill {\epsilon }_{i3}\hfill \end{array}\right],$$

Where $X_{i1}\sim N(0,1),X_{i2}$ Bernoulli(0.5) and

$$\left[\begin{array}{c}\hfill {ϵ}_{i1}\hfill \\ \hfill {ϵ}_{i2}\hfill \\ \hfill {ϵ}_{i3}\hfill \end{array}\right]\sim MVN\left(\left[\begin{array}{c}\hfill 0\hfill \\ \hfill 0\hfill \\ \hfill 0\hfill \end{array}\right],\left[\begin{array}{ccc}\hfill 1.0\hfill & \hfill -0.1\hfill & \hfill 0.2\hfill \\ \hfill -0.1\hfill & \hfill 1.0\hfill & \hfill -0.4\hfill \\ \hfill 0.2\hfill & \hfill -0.4\hfill & \hfill 1.0\hfill \end{array}\right]\right).$$

The correlation matrix of error terms ${\epsilon }_i={({\epsilon }_{i1},{\epsilon }_{i2},{\epsilon }_{i3})}^T$ was chosen to mimic the correlations between three lipid traits LDL-C, HDL-C and TG, estimated from the TOPMed data²⁶. We considered a sample size of 10,000 and generated genotypes by simulating 20,000 sequences for 100 different regions each spanning 1 Mb. The data generation used the calibration coalescent model (COSI)²⁹ with parameters set to mimic the LD structure of African Americans. In each simulation replicate, 10 annotations were generated as A₁, …, A₁₀ all independently and identically distributed as N(0,1) for each variant, and we randomly selected 5-kb regions from these 1-Mb regions for type I error rate simulations. We applied MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O by incorporating MAFs and the 10 annotations together with Burden-MT, SKAT-MT and ACAT-V-MT tests. We repeated the procedure with 10⁸ replicates to examine the type I error rate at levels α = 10⁻⁴, 10⁻⁵. and 10⁻⁶.

Empirical power simulations

Next, we carried out simulation studies under a variety of configurations to assess the the power of MultiSTAAR-O, and how its incorporation of multiple functional annotations affects power compared to the multi-trait burden, SKAT, and ACAT-V tests implemented in MultiSTAAR. In each simulation replicate, we randomly selected 5-kb regions from a 1-Mb region for power evaluations. For each selected 5-kb region, we generated three quantitative traits from a multivariate linear model

$${\mathit{Y}}_i=\left[\begin{array}{c}\hfill Y_{i1}\hfill \\ \hfill Y_{i2}\hfill \\ \hfill Y_{i3}\hfill \end{array}\right]=\left[\begin{array}{c}\hfill 0.5X_{i1}+0.5X_{i2}+{\mathit{G}}_i^T{\beta }_1\hfill \\ \hfill 0.5X_{i1}+0.5X_{i2}+{\mathit{G}}_i^T{\beta }_2\hfill \\ \hfill 0.5X_{i1}+0.5X_{i2}+{\mathit{G}}_i^T{\beta }_3\hfill \end{array}\right]+\left[\begin{array}{c}\hfill {\epsilon }_{i1}\hfill \\ \hfill {\epsilon }_{i2}\hfill \\ \hfill {\epsilon }_{i3}\hfill \end{array}\right],$$

where $X_{1i},X_{2i},{\epsilon }_i$ were defined as in the type I error rate simulations, ${\mathit{G}}_i={(G_{i1},G_{i2},\dots ,G_{ip})}^T$ and ${\beta }_k={({\beta }_{1,k},{\beta }_{2,k},\dots ,{\beta }_{p,k})}^T$ were the genotypes and effect sizes of the $p$ genetic variants in the signal region.

The genetic effect of variant $j$ on phenotype $k$ was defined as ${\beta }_{j,k}=c_j{d}_k{\gamma }_j$ to allow for heterogeneous effect sizes among variants and phenotypes. Specifically, we generated the causal variant indicator $c_j$ according to a logistic model

$$\text{logit}P(c_j=1)={\delta }_0+{\delta }_{l_1}{A}_{j,l_1}+{\delta }_{l_2}{A}_{j,l_2}+{\delta }_{l_3}{A}_{j,l_3}+{\delta }_{l_4}{A}_{j,l_4}+{\delta }_{l_5}{A}_{j,l_5},$$

where $\{l_1,\cdots ,l_5\}\subset \{1,\cdots ,10\}$ were randomly sampled for each region. For different regions, causality of variants depended on different sets of annotations. We set ${\delta }_{l.}=\log (5)$ for all annotations and varied the proportions of causal variants in the signal region by setting δ₀ = logit(0.0015), logit(0.015) and logit(0.18) which corresponds to averaging 5%, 15% and 35% causal variants in the signal region, respectively. We considered four scenarios of phenotypic indicator $d_k$ that reflect different underlying genetic architectures across phenotypes: $(d_1,d_2,d_3)=(1,0,0),(1,0,1),(1,1,0)$ and $(1,1,1)$. These correspond to causal variants in the signal region being associated with (1) one phenotype only, (2) two positively correlated phenotypes, (3) two negatively correlated phenotypes and (4) all three phenotypes. We modeled the absolute effect sizes of causal variants using $\mid {\gamma }_j\mid =c_0\mid {\log }_{10}{MAF}_j\mid$, such that it was a decreasing function of MAF. $c_0$ was set to be 0.13, 0.1, 0.1 and 0.07, respectively, to ensure a decent power of tests under each scenario. We additionally varied the proportions of causal variant effect size directions (signs of $r_j$) by randomly generating 100%, 80%, and 50% variants on average to have positive effects. We applied MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A, and MultiSTAAR-O using MAFs and all 10 annotations together with Burden-MT, SKAT-MT and ACAT-V-MT tests. We repeated the procedure with 10⁴ replicates to examine the power at level α = 10⁻⁷. The sample size was 10,000 across all scenarios.

Lipid Traits

Conventionally measured plasma lipids, including LDL-C, HDL-C, and triglycerides, were included for analysis. LDL-C was either calculated by the Friedewald equation when triglycerides were <400 mg/dl or directly measured. Given the average effect of statins, when statins were present, LDL-C was adjusted by dividing by 0.7. Triglycerides were natural log transformed for analysis. Phenotypes were harmonized by each cohort and deposited into the dbGaP TOPMed Exchange Area.

Multi-trait analysis of lipid levels in the TOPMed WGS data

The TOPMed WGS data consist of multi-ethnic related samples¹. Race/ethnicity was defined using a combination of self-reported race/ethnicity from participant questionnaires and study recruitment information (Supplementary Note)³¹. In this study, we applied MultiSTAAR to perform multi-trait rare variant analysis of three quantitative lipid traits (LDL-C, HDL-C and TG) using 20 study cohorts from the TOPMed Freeze 8 WGS data. LDL-C was adjusted for the presence of medications as before³⁰. For each study, we first fit a linear regression model adjusting for age, age², sex for each race/ethnicity-specific group. In addition, for Old Order Amish (OOA), we also adjusted for APOB p.R3527Q in LDL-C and TC analyses and adjusted for APOC3 p.R19Ter in TG and HDL-C analyses³⁰.

We performed rank-based inverse normal transformation of the residuals of LDL-C, HDL-C and TG within each race/ethnicity-specific group. We then fit a multivariate linear mixed model for the rank normalized residuals, adjusting for 11 ancestral principal components, ethnicity group indicators, and a variance component for empirically derived sparse kinship matrix to account for population structure, relatedness and correlation between phenotypes.

We next applied MultiSTAAR-O to perform multi-trait variant set analyses for rare variants (MAF < 1%) by scanning the genome, including gene-centric analysis of each protein-coding gene using five coding variant functional categories (putative loss-of-function rare variants, missense rare variants, disruptive missense rare variants, putative loss-of-function and disruptive missense rare variants and synonymous rare variants); seven noncoding variant functional categories (promoter rare variants overlaid with CAGE sites, promoter rare variants overlaid with DHS sites, enhancer rare variants overlaid with CAGE sites, enhancer rare variants overlaid with DHS sites, UTR rare variants, upstream region rare variants, downstream region rare variants) and rare variants in ncRNA genes; and genetic region analysis using 2-kb sliding windows across the genome with a 1-kb skip length. The WGS multi-trait rare variant analysis was performed using the R packages MultiSTAAR (version 0.9.7, https://github.com/xihaoli/MultiSTAAR) and STAARpipeline (version 0.9.7, https://github.com/xihaoli/STAARpipeline). The WGS rare variant single-trait analysis of LDL-C, HDL-C and TG was performed using the R package STAARpipeline (version 0.9.7, https://github.com/xihaoli/STAARpipeline). Both multi-trait and single-trait analyses results were summarized and visualized using the R package STAARpipelineSummary (version 0.9.7, https://github.com/xihaoli/STAARpipelineSummary).

Genome build

All genome coordinates are given in NCBI GRCh38/UCSC hg38.

Statistics and reproducibility

Sample size was not predetermined. The multi-trait analysis consists of 20 study cohorts of TOPMed Freeze 8 and had 61,838 samples with lipid traits. We did not use any study design that required randomization or blinding.

Extended Data

Extended Data Fig. 1 ∣. Manhattan plots and Q-Q plots for unconditional gene-centric coding, noncoding and ncRNA analysis of low-density lipoprotein cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C) and triglycerids (TG) using TOPMed data (n = 61,838).

a, Manhattan plots for unconditional gene-centric coding analysis of protein-coding gene. The horizontal line indicates a genome-wide MultiSTAAR-O P value threshold of 5.00 × 10⁻⁷. The significant threshold is defined by multiple comparisons using the Bonferroni correction (0.05/(20,000 × 5) = (5.00 × 10⁻⁷). Different symbols represent the MultiSTAAR-O P value of the protein-coding gene using different functional categories (putative loss-of-function, putative loss-of-function and disruptive missense, missense, disruptive missense, synonymous). b, Quantile-quantile plots for unconditional gene-centric coding analysis of protein-coding gene. Different symbols represent the MultiSTAAR-O P-value of the gene using different functional categories. c, Manhattan plots for unconditional gene-centric noncoding analysis of protein-coding gene. The horizontal line indicates a genome-wide MultiSTAAR-O P value threshold of 3.57 × 10⁻⁷. The significant threshold is defined by multiple comparisons using the Bonferroni correction (0.05/(20,000 × 7) = 3.57 × 10⁻⁷). Different symbols represent the MultiSTAAR-O P value of the protein-coding gene using different functional categories (upstream, downstream, UTR, promoter_CAGE, promoter_DHS, enhancer_CAGE, enhancer_DHS). Promoter_CAGE and promoter_DHS are the promoters with overlap of Cap Analysis of Gene Expression (CAGE) sites and DNase hypersensitivity (DHS) sites for a given gene, respectively. Enhancer_CAGE and enhancer_DHS are the enhancers in GeneHancer predicted regions with the overlap of CAGE sites and DHS sites for a given gene, respectively. d, Quantile-quantile plots for unconditional gene-centric noncoding analysis of protein-coding gene. Different symbols represent the MultiSTAAR-O P-value of the gene using different functional categories. e, Manhattan plots for unconditional gene-centric noncoding analysis of ncRNA gene. The horizontal line indicates a genome-wide MultiSTAAR-O P value threshold of 2.50 × 10⁻⁶. The significant threshold is defined by multiple comparisons using the Bonferroni correction (0.05/20,000 = 2.50 × 10⁻⁶). f, Quantile-quantile plots for unconditional gene-centric noncoding analysis of ncRNA gene. In panels, a, c and e, the chromosome number are indicated by the colors of dots. In all panels, MultiSTAAR-O is a two-sided test.

Extended Data Fig. 2 ∣. Power comparisons of Burden-MT, SKAT-MT, ACAT-V-MT (MT is short for Multi-Trait) and MultiSTAAR methods when variants in the signal region are associated with one phenotype.

Multi-trait Burden, SKAT and ACAT-V tests implemented in MultiSTAAR are denoted by Burden-MT, SKAT-MT and ACAT-V-MT. MultiSTAAR methods incorporating ten functional annotations are denoted by MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O. In each simulation replicate, a 5-kb region was randomly selected as the signal region. Within each signal region, variants were randomly generated to be causal based on the multivariate logistic model and on average there were 5%, 15% or 35% causal variants in the signal region. The effect sizes of causal variants were ${\beta }_j=c_0\mid {\log }_{10}{MAF}_j\mid$, where $c_0$ was set to be 0.13. The barplot of power in the top panel consider settings in which the effect sizes for the causal variants are 100% positive (0% negative), 80% positive (20% negative), and 50% positive (50% negative). The scatterplot of P values in the bottom panel compare MultiSTAAR-O to Burden-MT, SKAT-MT and ACAT-V-MT when 15% of variants in the signal region are causal variants with all positive effect sizes. Power was estimated as the proportion of the P values less than α = 10⁻⁷ based on 10⁴ replicates. Burden-MT, SKAT-MT, ACAT-V-MT, MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O are two-sided tests. Total sample size considered was 10,000.

Extended Data Fig. 3 ∣. Power comparisons of Burden-MT, SKAT-MT, ACAT-V-MT (MT is short for Multi-Trait) and MultiSTAAR methods when variants in the signal region are associated with two positively correlated phenotypes.

In each simulation replicate, a 5-kb region was randomly selected as the signal region. Within each signal region, variants were randomly generated to be causal based on the multivariate logistic model and on average there were 5%, 15% or 35% causal variants in the signal region. The effect sizes of causal variants were ${\beta }_j=c_0\mid {\log }_{10}{MAF}_j\mid$, where $c_0$ was set to be 0.1. The barplot of power in the top panel consider settings in which the effect sizes for the causal variants are 100% positive (0% negative), 80% positive (20% negative), and 50% positive (50% negative). The scatterplot of P values in the bottom panel compare MultiSTAAR-O to Burden-MT, SKAT-MT and ACAT-V-MT when 15% of variants in the signal region are causal variants with all positive effect sizes. Power was estimated as the proportion of the P values less than α = 10⁻⁷ based on 10⁴ replicates. Burden-MT, SKAT-MT, ACAT-V-MT, MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O are two-sided tests. Total sample size considered was 10,000.

Extended Data Fig. 4 ∣. Power comparisons of Burden-MT, SKAT-MT, ACAT-V-MT (MT is short for Multi-Trait) and MultiSTAAR methods when variants in the signal region are associated with two negatively correlated phenotypes.

In each simulation replicate, a 5-kb region was randomly selected as the signal region. Within each signal region, variants were randomly generated to be causal based on the multivariate logistic model and on average there were 5%, 15% or 35% causal variants in the signal region. The effect sizes of causal variants were ${\beta }_j=c_0\mid {\log }_{10}{MAF}_j\mid$, where $c_0$ was set to be 0.1. The barplot of power in the top panel consider settings in which the effect sizes for the causal variants are 100% positive (0% negative), 80% positive (20% negative), and 50% positive (50% negative). The scatterplot of P values in the bottom panel compare MultiSTAAR-O to Burden-MT, SKAT-MT and ACAT-V-MT when 15% of variants in the signal region are causal variants with all positive effect sizes. Power was estimated as the proportion of the P values less than α = 10⁻⁷ based on 10⁴ replicates. Burden-MT, SKAT-MT, ACAT-V-MT, MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O are two-sided tests. Total sample size considered was 10,000.

Extended Data Fig. 5 ∣. Power comparisons of Burden-MT, SKAT-MT, ACAT-V-MT (MT is short for Multi-Trait) and MultiSTAAR methods when variants in the signal region are associated with three phenotypes.

In each simulation replicate, a 5-kb region was randomly selected as the signal region. Within each signal region, variants were randomly generated to be causal based on the multivariate logistic model and on average there were 5%, 15% or 35% causal variants in the signal region. The effect sizes of causal variants were ${\beta }_j=c_0\mid {\log }_{10}{MAF}_j\mid$, where $c_0$ was set to be 0.07. The barplot of power in the top panel consider settings in which the effect sizes for the causal variants are 100% positive (0% negative), 80% positive (20% negative), and 50% positive (50% negative). The scatterplot of P values in the bottom panel compare MultiSTAAR-O to Burden-MT, SKAT-MT and ACAT-V-MT when 15% of variants in the signal region are causal variants with all positive effect sizes. Power was estimated as the proportion of the P values less than α = 10⁻⁷ based on 10⁴ replicates. Burden-MT, SKAT-MT, ACAT-V-MT, MultiSTAAR-B, MultiSTAAR-S, MultiSTAAR-A and MultiSTAAR-O are two-sided tests. Total sample size considered was 10,000.

NHLBI Trans-Omics for Precisition Medicine (TOPMed) Consortium

Namiko Abe⁵⁰, Gonçalo Abecasis⁵¹, Francois Aguet⁵², Christine Albert⁵³, Laura Almasy⁵⁴, Alvaro Alonso⁵⁵, Seth Ament⁵⁶, Peter Anderson⁵⁷, Pramod Anugu⁵⁸, Deborah Applebaum-Bowden⁵⁹, Kristin Ardlie⁵², Dan Arking⁶⁰, Allison Ashley-Koch⁶¹, Stella Aslibekyan⁶², Tim Assimes⁶³, Paul Auer⁶⁴, Dimitrios Avramopoulos⁶⁰, Najib Ayas⁶⁵, Adithya Balasubramanian⁶⁶, John Barnard⁶⁷, Kathleen Barnes⁶⁸, R. Graham Barr⁶⁹, Emily Barron-Casella⁶⁰, Lucas Barwick⁷⁰, Terri Beaty⁶⁰, Gerald Beck⁷¹, Diane Becker⁷², Lewis Becker⁶⁰, Rebecca Beer⁷³, Amber Beitelshees⁵⁶, Emelia Benjamin⁷⁴, Takis Benos⁷⁵, Marcos Bezerra⁷⁶, Larry Bielak⁵¹, Thomas Blackwell⁵¹, Nathan Blue⁷⁷, Russell Bowler⁷⁸, Ulrich Broeckel⁷⁹, Jai Broome⁵⁷, Deborah Brown⁸⁰, Karen Bunting⁵⁰, Esteban Burchard⁸¹, Carlos Bustamante⁸², Erin Buth⁸³, Jonathan Cardwell⁸⁴, Vincent Carey⁸⁵, Julie Carrier⁸⁶, Cara Carty⁸⁷, Richard Casaburi⁸⁸, Juan P Casas Romero⁸⁹, James Casella⁶⁰, Peter Castaldi⁹⁰, Mark Chaffin⁵², Christy Chang⁵⁶, Yi-Cheng Chang⁹¹, Daniel Chasman⁹², Sameer Chavan⁸⁴, Bo-Juen Chen⁵⁰, Wei-Min Chen⁹³, Michael Cho⁸⁵, Seung Hoan Choi⁵², Lee-Ming Chuang⁹⁴, Mina Chung⁹⁵, Ren-Hua Chung⁹⁶, Clary Clish⁹⁷, Suzy Comhair⁹⁸, Matthew Conomos⁸³, Elaine Cornell⁹⁹, Adolfo Correa¹⁰⁰, Carolyn Crandall⁸⁸, James Crapo¹⁰¹, L. Adrienne Cupples¹⁰², Jeffrey Curtis¹⁰³, Brian Custer¹⁰⁴, Coleen Damcott⁵⁶, Dawood Darbar¹⁰⁵, Sean David¹⁰⁶, Colleen Davis⁵⁷, Michelle Daya⁸⁴, Michael DeBaun¹⁰⁷, Dawn DeMeo⁸⁵, Ranjan Deka¹⁰⁸, Scott Devine⁵⁶, Huyen Dinh⁶⁶, Harsha Doddapaneni¹⁰⁹, Qing Duan¹¹⁰, Shannon Dugan-Perez¹¹¹, Ravi Duggirala¹¹², Jon Peter Durda¹¹³, Susan K. Dutcher¹¹⁴, Charles Eaton¹¹⁵, Lynette Ekunwe⁵⁸, Adel El Boueiz¹¹⁶, Patrick Ellinor¹¹⁷, Leslie Emery⁵⁷, Serpil Erzurum¹¹⁸, Charles Farber⁹³, Jesse Farek⁶⁶, Tasha Fingerlin¹¹⁹, Matthew Flickinger⁵¹, Chris Frazar⁵⁷, Mao Fu⁵⁶, Stephanie M. Fullerton⁵⁷, Lucinda Fulton¹²⁰, Stacey Gabriel⁵², Weiniu Gan⁷³, Shanshan Gao⁸⁴, Yan Gao⁵⁸, Margery Gass¹²¹, Heather Geiger¹²², Bruce Gelb¹²³, Mark Geraci¹²⁴, Soren Germer⁵⁰, Robert Gerszten¹²⁵, Auyon Ghosh⁸⁵, Richard Gibbs⁶⁶, Chris Gignoux⁶³, Mark Gladwin¹²⁶, David Glahn¹²⁷, Stephanie Gogarten⁵⁷, Da-Wei Gong⁵⁶, Harald Goring¹²⁸, Sharon Graw¹²⁹, Kathryn J. Gray¹³⁰, Daniel Grine⁸⁴, Colin Gross⁵¹, Yue Guan⁵⁶, Xiuqing Guo¹³¹, Namrata Gupta¹³², Jeff Haessler¹²¹, Michael Hall¹³³, Yi Han⁶⁶, Patrick Hanly¹³⁴, Daniel Harris¹³⁵, Nicola L. Hawley¹³⁶, Ben Heavner⁸³, Susan Heckbert¹³⁷, Ryan Hernandez⁸¹, David Herrington¹³⁸, Craig Hersh¹³⁹, Bertha Hidalgo⁶², James Hixson¹⁴⁰, Brian Hobbs¹⁴¹, John Hokanson⁸⁴, Elliott Hong⁵⁶, Karin Hoth¹⁴², Chao (Agnes) Hsiung¹⁴³, Jianhong Hu⁶⁶, Haley Huston¹⁴⁴, Chii Min Hwu¹⁴⁵, Rebecca Jackson¹⁴⁶, Deepti Jain⁵⁷, Cashell Jaquish¹⁴⁷, Jill Johnsen¹⁴⁸, Andrew Johnson⁷³, Craig Johnson⁵⁷, Rich Johnston⁵⁵, Kimberly Jones⁶⁰, Hyun Min Kang¹⁴⁹, Shannon Kelly¹⁵⁰, Eimear Kenny¹²³, Michael Kessler⁵⁶, Alyna Khan⁵⁷, Ziad Khan⁶⁶, Wonji Kim¹⁵¹, John Kimoff¹⁵², Greg Kinney¹⁵³, Barbara Konkle¹⁵⁴, Holly Kramer¹⁵⁵, Christoph Lange¹⁵⁶, Ethan Lange⁸⁴, Leslie Lange¹⁵⁷, Cathy Laurie⁵⁷, Cecelia Laurie⁵⁷, Meryl LeBoff⁸⁵, Jonathon LeFaive⁵¹, Jiwon Lee⁸⁵, Sandra Lee⁶⁶, Wen-Jane Lee¹⁴⁵, David Levine⁵⁷, Daniel Levy⁷³, Joshua Lewis⁵⁶, Xiaohui Li¹³¹, Yun Li¹¹⁰, Henry Lin¹³¹, Honghuang Lin¹⁵⁸, Simin Liu¹⁵⁹, Yongmei Liu¹⁶⁰, Yu Liu¹⁶¹, Steven Lubitz¹¹⁷, Kathryn Lunetta¹⁶², James Luo⁷³, Ulysses Magalang¹⁶³, Barry Make⁶⁰, Ani Manichaikul⁹³, Alisa Manning¹⁶⁴, JoAnn Manson⁸⁵, Melissa Marton¹²², Susan Mathai⁸⁴, Susanne May⁸³, Patrick McArdle⁵⁶, Merry-Lynn McDonald¹⁶⁵, Sean McFarland¹⁵¹, Stephen McGarvey¹⁶⁶, Daniel McGoldrick¹⁶⁷, Caitlin McHugh⁸³, Becky McNeil¹⁶⁸, Hao Mei⁵⁸, James Meigs¹⁶⁹, Vipin Menon⁶⁶, Luisa Mestroni¹²⁹, Ginger Metcalf⁶⁶, Deborah A Meyers¹⁷⁰, Emmanuel Mignot¹⁷¹, Julie Mikulla⁷³, Nancy Min⁵⁸, Mollie Minear¹⁷², Matt Moll⁹⁰, Zeineen Momin⁶⁶, Courtney Montgomery¹⁷³, Donna Muzny⁶⁶, Josyf C Mychaleckyj⁹³, Girish Nadkarni¹²³, Rakhi Naik⁶⁰, Take Naseri¹⁷⁴, Sergei Nekhai¹⁷⁵, Sarah C. Nelson⁸³, Bonnie Neltner⁸⁴, Caitlin Nessner⁶⁶, Deborah Nickerson¹⁷⁶, Osuji Nkechinyere⁶⁶, Kari North¹¹⁰, Jeff O'Connell¹⁷⁷, Tim O'Connor⁵⁶, Heather Ochs-Balcom¹⁷⁸, Geoffrey Okwuonu⁶⁶, Allan Pack¹⁷⁹, David T. Paik¹⁸⁰, James Pankow¹⁸¹, George Papanicolaou⁷³, Cora Parker¹⁸², Juan Manuel Peralta¹¹², Marco Perez⁶³, James Perry⁵⁶, Ulrike Peters¹⁸³, Lawrence S Phillips⁵⁵, Jacob Pleiness⁵¹, Toni Pollin⁵⁶, Wendy Post¹⁸⁴, Julia Powers Becker¹⁸⁵, Meher Preethi Boorgula⁸⁴, Michael Preuss¹²³, Pankaj Qasba⁷³, Dandi Qiao⁸⁵, Zhaohui Qin⁵⁵, Nicholas Rafaels¹⁸⁶, Mahitha Rajendran⁶⁶, D.C. Rao¹²⁰, Laura Rasmussen-Torvik¹⁸⁷, Aakrosh Ratan⁹³, Robert Reed⁵⁶, Catherine Reeves¹⁸⁸, Elizabeth Regan¹⁸⁹, Muagututi‘a Sefuiva Reupena¹⁹⁰, Rebecca Robillard¹⁹¹, Nicolas Robine¹²², Dan Roden¹⁹², Carolina Roselli⁵², Ingo Ruczinski⁶⁰, Alexi Runnels¹²², Pamela Russell⁸⁴, Sarah Ruuska¹⁹³, Kathleen Ryan⁵⁶, Ester Cerdeira Sabino¹⁹⁴, Danish Saleheen¹⁹⁵, Shabnam Salimi¹⁹⁶, Sejal Salvi⁶⁶, Steven Salzberg⁶⁰, Kevin Sandow¹⁹⁷, Vjay G. Sankaran¹⁹⁸, Jireh Santibanez⁶⁶, Karen Schwander¹²⁰, David Schwartz⁸⁴, Frank Sciurba¹²⁶, Christine Seidman¹⁹⁹, Jonathan Seidman²⁰⁰, Vivien Sheehan²⁰¹, Stephanie L. Sherman²⁰², Amol Shetty⁵⁶, Aniket Shetty⁸⁴, Wayne Hui-Heng Sheu¹⁴⁵, M. Benjamin Shoemaker²⁰³, Brian Silver²⁰⁴, Edwin Silverman⁸⁵, Robert Skomro²⁰⁵, Albert Vernon Smith²⁰⁶, Josh Smith⁵⁷, Nicholas Smith²⁰⁷, Tanja Smith⁵⁰, Sylvia Smoller²⁰⁸, Beverly Snively²⁰⁹, Michael Snyder²¹⁰, Tamar Sofer¹²⁵, Nona Sotoodehnia⁵⁷, Adrienne M. Stilp⁵⁷, Garrett Storm²¹¹, Elizabeth Streeten⁵⁶, Jessica Lasky Su²¹², Yun Ju Sung¹²⁰, Jody Sylvia⁸⁵, Adam Szpiro⁵⁷, Frédéric Sériès²¹³, Daniel Taliun⁵¹, Hua Tang²¹⁰, Margaret Taub⁶⁰, Matthew Taylor¹²⁹, Simeon Taylor⁵⁶, Marilyn Telen⁶¹, Timothy A. Thornton⁵⁷, Machiko Threlkeld²¹⁴, Lesley Tinker²¹⁵, David Tirschwell⁵⁷, Sarah Tishkoff²¹⁶, Catherine Tong²¹⁷, Russell Tracy²¹⁸, Michael Tsai¹⁸¹, Dhananjay Vaidya⁶⁰, David Van Den Berg²¹⁹, Peter VandeHaar⁵¹, Scott Vrieze¹⁸¹, Tarik Walker⁸⁴, Robert Wallace¹⁴², Avram Walts⁸⁴, Fei Fei Wang⁵⁷, Heming Wang²²⁰, Jiongming Wang²²¹, Karol Watson⁸⁸, Jennifer Watt⁶⁶, Daniel E. Weeks²²², Joshua Weinstock¹⁴⁹, Bruce Weir⁵⁷, Scott T Weiss²²³, Lu-Chen Weng¹¹⁷, Jennifer Wessel²²⁴, Cristen Willer¹⁰³, Kayleen Williams⁸³, L. Keoki Williams²²⁵, Scott Williams²²⁶, Carla Wilson⁸⁵, James Wilson²²⁷, Lara Winterkorn¹²², Quenna Wong⁵⁷, Baojun Wu²²⁸, Joseph Wu¹⁸⁰, Huichun Xu⁵⁶, Ivana Yang⁸⁴, Ketian Yu⁵¹, Seyedeh Maryam Zekavat⁵², Yingze Zhang²²⁹, Snow Xueyan Zhao¹⁰¹, Wei Zhao²³⁰, Xiaofeng Zhu²³¹, Elad Ziv²³², Michael Zody⁵⁰, Sebastian Zoellner⁵¹, Mariza de Andrade²³³, Lisa de las Fuentes²³⁴

50 - New York Genome Center, New York, New York, 10013, US; 51 - University of Michigan, Ann Arbor, Michigan, 48109, US; 52 - Broad Institute, Cambridge, Massachusetts, 2142, US; 53 - Cedars Sinai, Boston, Massachusetts, 2114, US; 54 - Children's Hospital of Philadelphia, University of Pennsylvania, Philadelphia, Pennsylvania, 19104, US; 55 - Emory University, Atlanta, Georgia, 30322, US; 56 - University of Maryland, Baltimore, Maryland, 21201, US; 57 - University of Washington, Seattle, Washington, 98195, US; 58 - University of Mississippi, Jackson, Mississippi, 38677, US; 59 - National Institutes of Health, Bethesda, Maryland, 20892, US; 60 - Johns Hopkins University, Baltimore, Maryland, 21218, US; 61 - Duke University, Durham, North Carolina, 27708, US; 62 - University of Alabama, Birmingham, Alabama, 35487, US; 63 - Stanford University, Stanford, California, 94305, US; 64 - Medical College of Wisconsin, Milwaukee, Wisconsin, 53211, US; 65 - Providence Health Care, Medicine, Vancouver, CA; 66 - Baylor College of Medicine Human Genome Sequencing Center, Houston, Texas, 77030, US; 67 - Cleveland Clinic, Cleveland, Ohio, 44195, US; 68 - Tempus, University of Colorado Anschutz Medical Campus, Aurora, Colorado, 80045, US; 69 - Columbia University, New York, New York, 10032, US; 70 - The Emmes Corporation, LTRC, Rockville, Maryland, 20850, US; 71 - Cleveland Clinic, Quantitative Health Sciences, Cleveland, Ohio, 44195, US; 72 - Johns Hopkins University, Medicine, Baltimore, Maryland, 21218, US; 73 - National Heart, Lung, and Blood Institute, National Institutes of Health, Bethesda, Maryland, 20892, US; 74 - Boston University, Massachusetts General Hospital, Boston University School of Medicine, Boston, Massachusetts, 2114, US; 75 - University of Florida, Epidemiology, Gainesville, Florida, 32610, US; 76 - Fundação de Hematologia e Hemoterapia de Pernambuco - Hemope, Recife, 52011-000, BR; 77 - University of Utah, Obstetrics and Gynecology, Salt Lake City, Utah, 84132, US; 78 - National Jewish Health, National Jewish Health, Denver, Colorado, 80206, US; 79 - Medical College of Wisconsin, Pediatrics, Milwaukee, Wisconsin, 53226, US; 80 - University of Texas Health at Houston, Pediatrics, Houston, Texas, 77030, US; 81 - University of California, San Francisco, San Francisco, California, 94143, US; 82 - Stanford University, Biomedical Data Science, Stanford, California, 94305, US; 83 - University of Washington, Biostatistics, Seattle, Washington, 98195, US; 84 - University of Colorado at Denver, Denver, Colorado, 80204, US; 85 - Brigham & Women's Hospital, Boston, Massachusetts, 2115, US; 86 - University of Montreal, US; 87 - Washington State University, Pullman, Washington, 99164, US; 88 - University of California, Los Angeles, Los Angeles, California, 90095, US; 89 - Brigham & Women's Hospital, US; 90 - Brigham & Women's Hospital, Medicine, Boston, Massachusetts, 2115, US; 91 - National Taiwan University, Taipei, 10617, TW; 92 - Brigham & Women's Hospital, Division of Preventive Medicine, Boston, Massachusetts, 2215, US; 93 - University of Virginia, Charlottesville, Virginia, 22903, US; 94 - National Taiwan University, National Taiwan University Hospital, Taipei, 10617, TW; 95 - Cleveland Clinic, Cleveland Clinic, Cleveland, Ohio, 44195, US; 96 - National Health Research Institute Taiwan, Miaoli County, 350, TW; 97 - Broad Institute, Metabolomics Platform, Cambridge, Massachusetts, 2142, US; 98 - Cleveland Clinic, Immunity and Immunology, Cleveland, Ohio, 44195, US; 99 - University of Vermont, Burlington, Vermont, 5405, US; 100 - University of Mississippi, Population Health Science, Jackson, Mississippi, 39216, US; 101 - National Jewish Health, Denver, Colorado, 80206, US; 102 - Boston University, Biostatistics, Boston, Massachusetts, 2115, US; 103 - University of Michigan, Internal Medicine, Ann Arbor, Michigan, 48109, US; 104 - Vitalant Research Institute, San Francisco, California, 94118, US; 105 - University of Illinois at Chicago, Chicago, Illinois, 60607, US; 106 - University of Chicago, Chicago, Illinois, 60637, US; 107 - Vanderbilt University, Nashville, Tennessee, 37235, US; 108 - University of Cincinnati, Cincinnati, Ohio, 45220, US; 109 - Baylor College of Medicine Human Genome Sequencing Center, Houston, Texas, 77030; 110 - University of North Carolina, Chapel Hill, North Carolina, 27599, US; 111 - Baylor College of Medicine Human Genome Sequencing Center, BCM, Houston, Texas, 77030, US; 112 - University of Texas Rio Grande Valley School of Medicine, Edinburg, Texas, 78539, US; 113 - University of Vermont, Pathology and Laboratory Medicine, Burlington, Vermont, 5405, US; 114 - Washington University in St Louis, Genetics, St Louis, Missouri, 63110, US; 115 - Brown University, Providence, Rhode Island, 2912, US; 116 - Harvard University, Channing Division of Network Medicine, Cambridge, Massachusetts, 2138, US; 117 - Massachusetts General Hospital, Boston, Massachusetts, 2114, US; 118 - Cleveland Clinic, Lerner Research Institute, Cleveland, Ohio, 44195, US; 119 - National Jewish Health, Center for Genes, Environment and Health, Denver, Colorado, 80206, US; 120 - Washington University in St Louis, St Louis, Missouri, 63130, US; 121 - Fred Hutchinson Cancer Research Center, Seattle, Washington, 98109, US; 122 - New York Genome Center, New York City, New York, 10013, US; 123 - Icahn School of Medicine at Mount Sinai, New York, New York, 10029, US; 124 - University of Pittsburgh, Pittsburgh, Pennsylvania, US; 125 - Beth Israel Deaconess Medical Center, Boston, Massachusetts, 2215, US; 126 - University of Pittsburgh, Pittsburgh, Pennsylvania, 15260, US; 127 - Boston Children's Hospital, Harvard Medical School, Department of Psychiatry, Boston, Massachusetts, 2115, US; 128 - University of Texas Rio Grande Valley School of Medicine, San Antonio, Texas, 78229, US; 129 - University of Colorado Anschutz Medical Campus, Aurora, Colorado, 80045, US; 130 - Mass General Brigham, Obstetrics and Gynecology, Boston, Massachusetts, 2115, US; 131 - Lundquist Institute, Torrance, California, 90502, US; 132 - Broad Institute, Broad Institute, Cambridge, Massachusetts, 2142, US; 133 - University of Mississippi, Cardiology, Jackson, Mississippi, 39216, US; 134 - University of Calgary, Medicine, Calgary, CA; 135 - University of Maryland, Genetics, Philadelphia, Pennsylvania, 19104, US; 136 - Yale University, Department of Chronic Disease Epidemiology, New Haven, Connecticut, 6520, US; 137 - University of Washington, Epidemiology, Seattle, Washington, 98195-9458, US; 138 - Wake Forest Baptist Health, Winston-Salem, North Carolina, 27157, US; 139 - Brigham & Women's Hospital, Channing Division of Network Medicine, Boston, Massachusetts, 2115, US; 140 - University of Texas Health at Houston, Houston, Texas, 77225, US; 141 - Regeneron Genetics Center, Boston, Massachusetts, 2115, US; 142 - University of Iowa, Iowa City, Iowa, 52242, US; 143 - National Health Research Institute Taiwan, Institute of Population Health Sciences, NHRI, Miaoli County, 350, TW; 144 - Blood Works Northwest, Seattle, Washington, 98104, US; 145 - Taichung Veterans General Hospital Taiwan, Taichung City, 407, TW; 146 - Oklahoma State University Medical Center, Internal Medicine, DIvision of Endocrinology, Diabetes and Metabolism, Columbus, Ohio, 43210, US; 147 - National Heart, Lung, and Blood Institute, National Institutes of Health, NHLBI, Bethesda, Maryland, 20892, US; 148 - University of Washington, Medicine, Seattle, Washington, 98109, US; 149 - University of Michigan, Biostatistics, Ann Arbor, Michigan, 48109, US; 150 - University of California, San Francisco, San Francisco, California, 94118, US; 151 - Harvard University, Cambridge, Massachusetts, 2138, US; 152 - McGill University, Montréal, QC H3A 0G4, CA; 153 - University of Colorado at Denver, Epidemiology, Aurora, Colorado, 80045, US; 154 - Blood Works Northwest, Medicine, Seattle, Washington, 98104, US; 155 - Loyola University, Public Health Sciences, Maywood, Illinois, 60153, US; 156 - Harvard School of Public Health, Biostats, Boston, Massachusetts, 2115, US; 157 - University of Colorado at Denver, Medicine, Aurora, Colorado, 80048, US; 158 - Boston University, University of Massachusetts Chan Medical School, Worcester, Massachusetts, 1655, US; 159 - Brown University, Epidemiology and Medicine, Providence, Rhode Island, 2912, US; 160 - Duke University, Cardiology, Durham, North Carolina, 27708, US; 161 - Stanford University, Cardiovascular Institute, Stanford, California, 94305, US; 162 - Boston University, Boston, Massachusetts, 2215, US; 163 - The Ohio State University, Division of Pulmonary, Critical Care and Sleep Medicine, Columbus, Ohio, 43210, US; 164 - Broad Institute, Harvard University, Massachusetts General Hospital; 165 - University of Alabama, University of Alabama at Birmingham, Birmingham, Alabama, 35487, US; 166 - Brown University, Epidemiology, Providence, Rhode Island, 2912, US; 167 - University of Washington, Genome Sciences, Seattle, Washington, 98195, US; 168 - RTI International, US; 169 - Massachusetts General Hospital, Medicine, Boston, Massachusetts, 2114, US; 170 - University of Arizona, Tucson, Arizona, 85721, US; 171 - Stanford University, Center For Sleep Sciences and Medicine, Palo Alto, California, 94304, US; 172 - National Institute of Child Health and Human Development, National Institutes of Health, Bethesda, Maryland, 20892, US; 173 - Oklahoma Medical Research Foundation, Genes and Human Disease, Oklahoma City, Oklahoma, 73104, US; 174 - Ministry of Health, Government of Samoa, Apia, WS; 175 - Howard University, Washington, District of Columbia, 20059, US; 176 - University of Washington, Department of Genome Sciences, Seattle, Washington, 98195, US; 177 - University of Maryland, Balitmore, Maryland, 21201, US; 178 - University at Buffalo, Buffalo, New York, 14260, US; 179 - University of Pennsylvania, Division of Sleep Medicine/Department of Medicine, Philadelphia, Pennsylvania, 19104-3403, US; 180 - Stanford University, Stanford Cardiovascular Institute, Stanford, California, 94305, US; 181 - University of Minnesota, Minneapolis, Minnesota, 55455, US; 182 - RTI International, Biostatistics and Epidemiology Division, Research Triangle Park, North Carolina, 27709-2194, US; 183 - Fred Hutchinson Cancer Research Center, Fred Hutch and UW, Seattle, Washington, 98109, US; 184 - Johns Hopkins University, Cardiology/Medicine, Baltimore, Maryland, 21218, US; 185 - University of Colorado at Denver, Medicine, Denver, Colorado, 80204, US; 186 - University of Colorado at Denver, CCPM, Denver, Colorado, 80045, US; 187 - Northwestern University, Chicago, Illinois, 60208, US; 188 - New York Genome Center, New York Genome Center, New York City, New York, 10013, US; 189 - National Jewish Health, Medicine, Denver, Colorado, 80206, US; 190 - Lutia I Puava Ae Mapu I Fagalele, Apia, WS; 191 - University of Ottawa, Sleep Research Unit, University of Ottawa Institute for Mental Health Research, Ottawa, ON K1Z 7K4, CA; 192 - Vanderbilt University, Medicine, Pharmacology, Biomedicla Informatics, Nashville, Tennessee, 37235, US; 193 - University of Washington, Seattle, Washington, 98104, US; 194 - Universidade de Sao Paulo, Faculdade de Medicina, Sao Paulo, 1310000, BR; 195 - Columbia University, New York, New York, 10027, US; 196 - University of Maryland, Pathology, Seattle, Washington, 98195, US; 197 - Lundquist Institute, TGPS, Torrance, California, 90502, US; 198 - Harvard University, Division of Hematology/Oncology, Boston, Massachusetts, 2115, US; 199 - Harvard Medical School, Genetics, Boston, Massachusetts, 2115, US; 200 - Harvard Medical School, Boston, Massachusetts, 2115, US; 201 - Emory University, Pediatrics, Atlanta, Georgia, 30307, US; 202 - Emory University, Human Genetics, Atlanta, Georgia, 30322, US; 203 - Vanderbilt University, Medicine/Cardiology, Nashville, Tennessee, 37235, US; 204 - UMass Memorial Medical Center, Worcester, Massachusetts, 1655, US; 205 - University of Saskatchewan, Saskatoon, SK S7N 5C9, CA; 206 - University of Michigan; 207 - University of Washington, Epidemiology, Seattle, Washington, 98195, US; 208 - Albert Einstein College of Medicine, New York, New York, 10461, US; 209 - Wake Forest Baptist Health, Biostatistical Sciences, Winston-Salem, North Carolina, 27157, US; 210 - Stanford University, Genetics, Stanford, California, 94305, US; 211 - University of Colorado at Denver, Genomic Cardiology, Aurora, Colorado, 80045, US; 212 - Brigham & Women's Hospital, Channing Department of Medicine, Boston, Massachusetts, 2115, US; 213 - Université Laval, Quebec City, G1V 0A6, CA; 214 - University of Washington, University of Washington, Department of Genome Sciences, Seattle, Washington, 98195, US; 215 - Fred Hutchinson Cancer Research Center, Cancer Prevention Division of Public Health Sciences, Seattle, Washington, 98109, US; 216 - University of Pennsylvania, Genetics, Philadelphia, Pennsylvania, 19104, US; 217 - University of Washington, Department of Biostatistics, Seattle, Washington, 98195, US; 218 - University of Vermont, Pathology & Laboratory Medicine, Burlington, Vermont, 5405, US; 219 - University of Southern California, USC Methylation Characterization Center, University of Southern California, California, 90033, US; 220 - Brigham & Women's Hospital, Mass General Brigham, Boston, Massachusetts, 2115, US; 221 - University of Michigan, US; 222 - University of Pittsburgh, Department of Human Genetics, Pittsburgh, Pennsylvania, 15260, US; 223 - Brigham & Women's Hospital, Channing Division of Network Medicine, Department of Medicine, Boston, Massachusetts, 2115, US; 224 - Indiana University, Epidemiology, Indianapolis, Indiana, 46202, US; 225 - Henry Ford Health System, Detroit, Michigan, 48202, US; 226 - Case Western Reserve University; 227 - Beth Israel Deaconess Medical Center, Cardiology, Cambridge, Massachusetts, 2139, US; 228 - Henry Ford Health System, Department of Medicine, Detroit, Michigan, 48202, US; 229 - University of Pittsburgh, Medicine, Pittsburgh, Pennsylvania, 15260, US; 230 - University of Michigan, Department of Epidemiology, Ann Arbor, Michigan, 48109, US; 231 - Case Western Reserve University, Department of Population and Quantitative Health Sciences, Cleveland, Ohio, 44106, US; 232 - University of California, San Francisco, Medicine, San Francisco, California, 94143, US; 233 - Mayo Clinic, Health Quantitative Sciences Research, Rochester, Minnesota, 55905, US; 234 - Washington University in St Louis, Department of Medicine, Cardiovascular Division, St. Louis, Missouri, 63110, US

Data availability

This paper used the TOPMed Freeze 8 WGS data and lipids phenotype data. Genotype and phenotype data are both available in database of Genotypes and Phenotypes. The TOPMed WGS data were from the following twenty study cohorts (accession numbers provided in parentheses): Old Order Amish (phs000956.v1.p1), Atherosclerosis Risk in Communities Study (phs001211), Mt Sinai BioMe Biobank (phs001644), Coronary Artery Risk Development in Young Adults (phs001612), Cleveland Family Study (phs000954), Cardiovascular Health Study (phs001368), Diabetes Heart Study (phs001412), Framingham Heart Study (phs000974), Genetic Study of Atherosclerosis Risk (phs001218), Genetic Epidemiology Network of Arteriopathy (phs001345), Genetic Epidemiology Network of Salt Sensitivity (phs001217), Genetics of Lipid Lowering Drugs and Diet Network (phs001359), Hispanic Community Health Study - Study of Latinos (phs001395), Hypertension Genetic Epidemiology Network and Genetic Epidemiology Network of Arteriopathy (phs001293), Jackson Heart Study (phs000964), Multi-Ethnic Study of Atherosclerosis (phs001416), San Antonio Family Heart Study (phs001215), Genome-wide Association Study of Adiposity in Samoans (phs000972), Taiwan Study of Hypertension using Rare Variants (phs001387), and Women’s Health Initiative (phs001237). The sample sizes, ancestry and phenotype summary statistics of these cohorts are given in Supplementary Table 2.

The functional annotation data are publicly available and were downloaded from the following links: GRCh38 CADD v1.4 (https://cadd.gs.washington.edu/download); ANNOVAR dbNSFP v3.3a (https://annovar.openbioinformatics.org/en/latest/user-guide/download); LINSIGHT (https://github.com/CshlSiepelLab/LINSIGHT); FATHMM-XF (http://fathmm.biocompute.org.uk/fathmm-xf); FANTOM5 CAGE (https://fantom.gsc.riken.jp/5/data); GeneCards (https://www.genecards.org; v4.7 for hg38); and Umap/Bismap (https://bismap.hoffmanlab.org; ‘before March 2020’ version). In addition, recombination rate and nucleotide diversity were obtained from Gazal et al⁵⁰. The whole-genome individual functional annotation data was assembled from a variety of sources and the computed annotation principal components are available at the Functional Annotation of Variant-Online Resource (FAVOR) site (https://favor.genohub.org)⁵¹ and the FAVOR database (https://doi.org/10.7910/DVN/1VGTJI)⁵².

Code availability

MultiSTAAR is implemented as an open source R package available at https://github.com/xihaoli/MultiSTAAR and https://content.sph.harvard.edu/xlin/software.html. Data analysis was performed in R (4.1.0). STAAR v0.9.7 and MultiSTAAR v0.9.7 were used in simulation and real data analysis and implemented as open-source R packages available at https://github.com/xihaoli/STAAR and https://github.com/xihaoli/MultiSTAAR. The assembled functional annotation data were downloaded from FAVOR using Wget (https://www.gnu.org/software/wget/wget.html).