Main

Genome-wide association studies (GWAS) of HF have provided valuable translational insights, yet existing studies are often limited by phenotypic heterogeneity and a lack of stratification by etiology1,2,3. HF often develops following significant risk factor exposures, such as myocardial infarction and valvular heart disease4. We hypothesized that a study of HF occurring without major antecedent risk factors (specifically, myocardial infarction, coronary revascularization or marked structural or valvular heart diseases) could better capture genetic factors directly influencing HF susceptibility5 (Extended Data Fig. 1 and Supplementary Fig. 1). We refer to this form of HF as ni-HF. Where data permitted, ni-HF cases were further stratified by left ventricular systolic dysfunction, defining those with left ventricular ejection fraction (LVEF) above or below 50% as ni-HFpEF or ni-HFrEF, respectively. We performed GWAS meta-analyses to identify genomic loci, prioritize putative effector genes and characterize the key pathways, tissues and cell types involved in etiology. We further investigated etiologic clusters through a pleiotropy scan and used Mendelian randomization (MR) to assess the genetic evidence supporting risk factor associations. Our study advances the understanding of the genetic architecture of HF and its subtypes.

Results

GWAS meta-analysis

We conducted a GWAS meta-analysis of HF and its subtypes across 42 studies participating in the Heart Failure Molecular Epidemiology for Therapeutic Targets (HERMES) consortium. The study population comprised 1,946,349 individuals, including 153,174 diagnosed cases of HF (hereafter termed HFall), of which 44,012 had a nonischemic etiology (ni-HF). Among these, 5,406 were classified as ni-HFrEF (LVEF < 50%) and 3,841 were classified as ni-HFpEF (LVEF ≥ 50%; Supplementary Figs. 1 and 2). Antecedent myocardial infarction was reported for 29% of the HFall cases and <1% of the ni-HF phenotypes (Supplementary Table 1). The identified HF cases encompassed five major ancestry groups: 139,533 (91%) European; 9,413 (6.2%) East Asian; 3,292 (2.2%) African; 779 (0.5%) South Asian; and 157 (0.1%) admixed American (Supplementary Table 2).

We performed a cross-ancestry, fixed-effect inverse-variance-weighted (IVW) meta-analysis to investigate the associations of 10,199,961 genetic variants (minor allele frequency (MAF) >1%) with the risk of HF and its subtypes. We identified 59 conditionally independent (sentinel) genetic variants associated with HFall at P < 5 × 10−8, at 56 nonoverlapping genomic loci (distance > 500 kb; Fig. 1 and Supplementary Table 3). We did not observe heterogeneity of effects between ancestry groups, except for one sentinel variant at the LPA locus (Phet-ancestry < 0.05/59; Supplementary Fig. 3 and Supplementary Table 4). Ten additional variants were associated with ni-HF subtypes but not with HFall. In total, 66 independent genomic loci were identified across the HF phenotypes at P < 5 × 10−8, of which 46 (70%) loci passed a multiplicity-adjusted threshold of P < 5 × 10−8/4 phenotypes. Of the 66 identified loci, 46 (70%) were associated with at least one nonischemic subtype at adjusted P < 0.05/66 (hereafter termed ni-HF loci). The remaining 20 loci were associated with HFall but not with the nonischemic subtypes (hereafter termed other HF loci; Fig. 2c). We identified 37 variants not previously reported and successfully replicated 76 of 87 (87%) previously reported HF variants available in our study (P < 0.05/87; Supplementary Fig. 4 and Supplementary Table 5)1,2,3,6,7.

Fig. 1: GWAS meta-analysis results across heart failure phenotypes.
figure 1

a, Phenotyping schema. b, Manhattan plots of four HF subtypes showing −log10(P) for genetic associations from GWAS meta-analysis (y axis) across genetic variants ordered by chromosome and base pair positions (x axis). c, Summary of conditionally independent lead variants effect across HF phenotypes. Lead variants are denoted using chromosome, base pair position according to the GRCh37 assembly, risk-increasing allele in HF phenotype with the lowest P value for association and the other allele. Loci are categorized by the strength of genetic associations across HF phenotypes, with labels on the right edge of the plot. The presented P values are derived from two-sided association tests as described in Methods. Chr1, chromosome 1.

Full size image
Fig. 2: Prioritization of effector genes across heart failure genetic susceptibility loci.
figure 2

a, Circular heatmap showing predictor score and Boolean classifier scores for 142 candidate genes across HF susceptibility loci. Blue tiles indicate a 'true' value and prioritized genes within each locus are highlighted in red. Coloc refers to colocalization evidence due to shared causal variants with gene expression level in the most relevant tissue; Mendel indicates overlap with Mendelian disease genes; ABC refers to the ABC measure for enhancer-gene activity of overlapping lead variants and their proxies or fine-mapped variants; overall refers to the weighted average of PoPS, TWAS statistics and V2G scaled predictor score (with 2:1:2 weight ratio). b, Schematic diagram of locus categorization by genetic associations across HF subtypes. c, Upset plot showing number of loci per phenotype category. d, Schematic diagram of effector gene prioritization.

Full size image

To investigate sex as a potential modifier of variant effects, we conducted a meta-regression analysis of sex ratio (proportion of male sex) and additive effects of risk-increasing alleles for 66 identified HF sentinel variants across the participating studies. Although limited by statistical power, we observed nominally significant evidence of effect modification by sex at 3 of the 66 loci (P < 0.05; Supplementary Table 6).

Genetic architecture and heritability

The genetic architecture of HF was found to be highly polygenic, evidenced by an elevated genomic inflation factor (λGC = 1.22) in the absence of population stratification (linkage disequilibrium score (LDSC) regression intercept = 1.01; Supplementary Figs. 5 and 6 and Supplementary Table 7). We observed an exponential relationship between allele frequency and effect size for the associated variants (Extended Data Fig. 2). The estimated proportion of variance in disease liability explained by common genetic variants under an additive model, or SNP-based heritability (h2g), was 5.4 ± 0.2% for HFall, 6.1 ± 0.5% for ni-HF, 11.8 ± 2.6% for ni-HFrEF and 1.8 ± 1.3% for ni-HFpEF (Extended Data Fig. 3)8. Positive genetic correlations were observed across between all HF subtype pairs, with estimates ranging from 0.42 (s.e. = 0.18) between ni-HFrEF and ni-HFpEF to 0.93 (s.e. = 0.15) between ni-HF and ni-HFpEF (Supplementary Fig. 7). We derived a polygenic score (PGSHF) from the HFall GWAS, excluding UK Biobank (UKB) participants, and evaluated its association with HF in UKB. Among 347,235 UKB European participants (13,793 HFall cases), the PGSHF was associated with HFall (odds ratio (OR) per PGS s.d. = 1.37 (95% confidence interval (CI) = 1.35–1.39), P < 2 × 10−16), after adjusting for sex, age and first ten genetic principal components (PCs). At a 5% false positive rate, 19% of the cases were identified. Individuals in the top decile of PGSHF had 1.70-fold higher odds of developing heart failure compared to those in the fifth decile (OR = 1.70, 95% CI = 1.59–1.82, P = 1.7 × 10−142) and 2.89-fold compared to those in the bottom decile (OR = 2.89, 95% CI = 2.66–3.14, P = 2 × 10−51; Supplementary Table 8 and Extended Data Fig. 4).

Prioritization of effector variants, genes and pathways

We then examined the functional properties of variants and genes within each GWAS locus. Through functionally informed fine mapping, we identified 70 credible sets containing 547 putative causal variants at 47/66 HF loci (cumulative posterior inclusion probability (PIP) >0.95)9. Eleven fine-mapped variants, including exonic variants in established dilated cardiomyopathy (DCM) genes FLNC, BAG3 and HSPB7, had high predicted deleteriousness (Combined Annotation-Dependent Depletion (CADD) Phred score > 20)10,11,12,13 (Extended Data Fig. 5 and Supplementary Table 9). Next, we used a two-stage approach to identify potential effector genes by evaluating 758 protein-coding genes located within heart failure loci. In the first stage, we identified candidate genes ranked highest by one or more of the following gene prioritization methods: (1) variant-to-gene (V2G) scoring of fine-mapped variants14, (2) gene-level polygenic priority scores (PoPS)15 and (3) association of predicted gene expression with HFall in multitissue transcriptome-wide association study (TWAS)16. In the second stage, we nominated a single best-prioritized gene for each of the 66 loci based on the following: (1) predicted enhancer-gene activity using the activity-by-contact (ABC) model17,18, (2) colocalization of gene expression in a relevant tissue19 and (3) association with a phenotypically relevant Mendelian disorder20 (Fig. 2 and Supplementary Table 10). Results for regional genetic associations, gene prioritization scores, cross-trait association and study-level estimates are provided in Supplementary Data 1.

The prioritized genes in ni-HF loci included those implicated in cardiac development and cardiomyopathy (ACTN2, BAG3, CAMKD2 (ref. 21), CAND2 (ref. 22), FLNC, ACTN2 and HSPB7 (refs. 11,12,13,23), CAMKD2 (ref. 21), NKX2-5 and STRN24), cardiac hypertrophy (CAMKD2 (ref. 21) and CAND2 (ref. 22)) and cardiac arrhythmia (PITX2, KLF12 and ATP1B1 (refs. 25,26,27)). Genes with an established role in atherosclerosis (LDLR28, LPL29, ABCG5 (refs. 30,31) and LPA32) were notable among loci associated with HFall but not with ni-HF.

We also performed an exploratory enrichment analysis of the candidate gene sets to identify additional pathways and cellular components associated with HFall and ni-HF (insufficient genes were identified for ni-HFrEF and ni-HFrEF; Extended Data Fig. 6 and Supplementary Table 11). Cellular growth and senescence pathways were highly enriched for both HFall and ni-HF. Notably, a key senescence-associated secretory phenotype (SASP) gene, insulin-like growth factor binding protein 7 (IGFBP7), was prioritized at one of the ni-HFpEF loci. IGFBP7 encodes a circulating anti-angiogenic factor produced by the vasculature, which has recently been implicated in cardiomyocyte senescence and cardiac remodeling33. The glucocorticoid pathway was enriched across all HF phenotypes, and genes associated with the sarcomeric Z-disk were overrepresented in both HFall and ni-HF.

Identifying organs, tissues and cells involved in etiology

To identify the biological components contributing to HF etiology, we performed a heritability enrichment analysis across 206 tissues and cell types, representing 12 organs and systems (Fig. 3 and Supplementary Table 12). We observed enrichment of 46 unique tissues and cell types (one-sided P < 0.05), highlighting the diverse range of organ systems involved in HF etiology. Overall, cardiac tissues were the most frequently enriched (15 enrichments), followed by musculoskeletal/connective tissues (12 enrichments) and nervous system tissues (8 enrichments). Cardiac tissues showed the highest enrichment for HFall, ni-HF and ni-HFrEF, whereas for niHFpEF, the kidney and pancreas showed the highest enrichment. We then investigated the contribution of 14 cardiac cell populations using single-nucleus transcriptomics derived from 185,185 nuclei from 16 nonfailing human hearts. Heritability enrichment and gene-based association enrichment using MAGMA34 revealed cardiomyocytes as the major effector cell type for ni-HF. Additionally, vascular smooth muscle cells were also enriched for HFall but not for ni-HF (Supplementary Fig. 8 and Supplementary Table 13).

Fig. 3: Heritability enrichment of tissues and cell types across heart failure phenotypes.
figure 3

Estimates for heritability enrichment of 206 tissues and cell types in 12 system/organ categories across four HF phenotypes are presented. Top five tissues/cell types per phenotype are highlighted with labels. The presented −log10(P) is derived from a one-sided statistical test of heritability enrichment using LDSC-SEG as described in Methods.

Full size image

To investigate whether genes prioritized at the identified HF loci were differentially regulated in failing hearts, we compared the cell-type transcriptomic profile of the nonfailing heart samples with those from 28 DCM and hypertrophic cardiomyopathy (HCM) patients (n nuclei = 344,797). We found that 30/95 (30%) of ni-HF genes were differentially regulated, predominantly within fibroblasts and cardiomyocytes (Extended Data Fig. 7 and Supplementary Table 14). Notably, nuclear receptor subfamily 3 group C member 1 (NR3C1), which encodes glucocorticoid receptors, was upregulated in 12 of 14 cardiac cell types from failing hearts. This upregulation aligns with the glucocorticoid pathway enrichment among the prioritized genes identified at HF GWAS loci. These findings suggest that glucocorticoid signaling may have a cardioprotective role, potentially through the inhibition of pathological mineralocorticoid receptor signaling35, which could help explain, in part, the beneficial effects of mineralocorticoid receptor antagonist drugs in HF36. Another pro-survival HF gene, KLF12, was upregulated in both cardiomyocytes and cardiac fibroblasts. KLF12 is a transcription factor that upregulates p53, which in turn leads to upregulation of p21 (CDKN1A)37. Activation of p53 has been implicated as an important mechanism in LMNA cardiomyopathy38. In contrast to prior reports, we did not observe differential expression of IGFBP7 in cardiomyocytes from failing hearts33.

Mapping genetic loci to etiologic clusters

To explore potential mechanisms by which HF genetic loci influence disease risk, we conducted a pleiotropy scan to examine associations between the identified loci and 294 clinical traits and diseases systematically ascertained from UKB39. We identified 207 pleiotropic associations at false discovery rate (FDR) < 1% (corresponding to P < 0.001) between 46/66 (70%) HF loci and 79 (27%) of the phenotypes tested (Extended Data Fig. 8, Supplementary Fig. 9 and Supplementary Table 15). To visualize these associations, we constructed a pleiotropy graph network by representing loci and diseases as nodes and locus–phenotype associations as edges (Extended Data Fig. 9a). Using a community detection algorithm, we identified 18 etiological clusters of locus–disease associations (Fig. 4, Extended Data Fig. 9b and Supplementary Table 16). The largest cluster, cluster 1, included traits and diseases related to atherosclerosis, with loci that were not associated with ni-HF subtypes. Cluster 2 encompassed cardiac arrhythmias and DCM and loci associated with ni-HF. Cluster 3 centered on the pleiotropic effects of a locus prioritizing ALDH2, which encodes an enzyme involved in removing toxic acetaldehydes, with deficiencies in ALDH2 being linked to increased risks of cancer and cardiovascular diseases40. The sentinel variant at the corresponding locus has been associated with reduced ALDH2 expression in blood41. Cluster 4 was centered around hypertension, a major risk factor for all HF subtypes. Finally, cluster 5 comprised ni-HF and ni-HFpEF loci associated with adiposity, diabetes and carpal tunnel syndrome, which are notable clinical risk factors for HFpEF.

Fig. 4: Etiologic clusters of heart failure identified through pleiotropy network analysis.
figure 4

The network is constructed from 207 genotype–phenotype associations across 79 unique diseases and 46 HF susceptibility loci identified from phenome-wide association (PheWAS) analysis in UKB at FDR <1%. Nodes represent genetic loci labeled by the prioritized gene (solid background with bold–italic label, colored by categorical association across HF phenotypes) and phenotypes (translucent background, colored by phenotype category), sized proportionally to centrality measure. Edges (arrows connecting locus nodes to phenotype nodes) represent the association, with thickness representing the strength of association measured by absolute z score. The full pleiotropy network and phenotype category color codes are presented in Extended Data Fig. 9. Phenotype abbreviations, phenotype categories and phenome-wide association results are presented in Supplementary Table 15 and Extended Data Fig. 8. Cluster membership and centrality measures of nodes are presented in Supplementary Table 26 and Supplementary Fig. 13. CHD, coronary heart disease; NOS, not-otherwise-specified (nonspecific cause), AF, atrial fibrillation; AV block, atrioventricular block; LRTI, lower respiratory tract infection; COPD, chronic obstructive pulmonary disease; LBBB, left bundle branch block; BPH, benign prostatic hyperplasia.

Full size image

We then investigated the genetic associations of sentinel variants across 66 HF loci with 24 related diseases, lifestyle factors and quantitative traits related to cardiovascular physiology and cardiac structure and function. This analysis revealed 207 pleiotropic associations at FDR < 1% (Extended Data Fig. 10 and Supplementary Table 17). Notably, 24 of the 66 HF loci (36%) were not associated with any of the traits investigated (P ≥ 5 × 10−8). Cross-trait colocalization analysis revealed evidence for shared causal variants for 105 locus–phenotype pairs (posterior probability of shared causal variants > 0.8; Fig. 5 and Supplementary Table 18). For example, the BTB domain and CNC homolog 1 (BACH1) ni-HFrEF locus colocalized with both coronary artery disease (CAD) and DCM. BACH1, a transcription factor induced by mechanical stress in cardiac fibroblasts and endothelial cells, regulates Yes1-associated transcriptional regulator (YAP) expression by binding to its promoter42. The BACH1–YAP transcriptional network plays a crucial role in cardiac regeneration by modulating proliferative programs in cardiomyocytes43. Additionally, six HF loci colocalized with either estimated glomerular filtration rate (eGFR) or chronic kidney disease (CKD), highlighting the key role of the cardio–renal axis in HF. The prioritized gene at one of these loci, PFDN1, encodes prefoldin subunit 1, a molecular chaperone involved in protein folding. This gene has been associated with mortality and cardiovascular phenotypes in mouse knockout models44,45.

Fig. 5: Genetic colocalization between the overall heart failure phenotype and related phenotypes.
figure 5

a, Chord diagram showing connections between 22 (of 24 tested) related phenotypes across 42 (of 66) HF susceptibility loci with posterior probability of shared causal variants (PPcoloc H4) > 0.8. Each band connects a locus to a phenotype, representing the sharing of causal genetic variants (colocalization) between the tested phenotype and HF at the locus. b, Total number of colocalized phenotypes across HF susceptibility loci. c, Total number of HF susceptibility loci with genetic colocalization across tested phenotypes. DBP, diastolic blood pressure; FVC, forced vital capacity; FEV1, forced expiratory volume in 1 s; LVEDV, left ventricular end diastolic volume; PDSR-r, peak diastolic strain rate radial; PDSR-l, peak diastolic strain rate longitudinal; RVEF, right ventricular ejection fraction; RVEDV, right ventricular end diastolic volume; LAVi, left atrial volume index.

Full size image

Estimation of risk factor effects on heart failure

Previous genetic studies have identified certain exposures as potential causal factors for HF; however, these studies have been limited in the range of traits examined and characterization of their association with HF subtypes1,2,3. To address this gap, we investigated the potential role of 24 related phenotypes (used in the etiologic cluster mapping above) in the etiology of HF and its subtypes. First, we estimated the genetic correlation (rg) between the HF phenotypes and the exposures of interest using bivariate LDSC. We found evidence of shared additive genetic effects for 21/24 phenotypes, with rg values ranging from −0.25 to 0.67 (Fig. 6 and Supplementary Table 19). Next, we estimated the associations using MR with an IVW fixed effects model and performed sensitivity analyses using MR–Egger and the weighted median estimator (WME). Overall, 18 of the 24 of the exposures were associated with at least one HF subtype (FDR < 1% of IVW estimate, with directional concordance in sensitivity analyses; Fig. 6, Supplementary Figs. 10–12 and Supplementary Table 20).

Fig. 6: Genetic correlation (rg) and MR estimates across 24 traits and four heart failure phenotypes.
figure 6

Asterisks (*) indicate binary traits. MR effect estimates are reported as OR (ORMR) per doubling prevalence for binary traits or per s.d. increase for quantitative traits. Estimates that were robust to multiple-testing adjustment at FDR <1% and sensitivity analyses were indicated by light blue shade (for rg < 0 and ORMR < 1) or light red shade (for rg > 0 and ORMR > 1). The heatmaps represent two-sided P values for associations from different MR models, color-coded with the direction of MR estimates and strength of associations. Diamonds and whiskers on the forest plots represent point estimates and 95% CIs for rg (left) and ORMR (right). Missing diamonds/point whiskers with arrows represent point estimates/95% CIs outside the scale. HR, heart rate; RA-area, right atrial area.

Full size image

The estimated effects of risk factors on HFall and ni-HF outcomes were similar, with the exception of CAD. Consistent with prior studies, liability to CAD was associated with HFall (MR OR per doubling prevalence with IVW estimator (ORMR–IVW) = 1.20, 95% CIs = 1.18–1.22), but not with ni-HF (ORMR–IVW = 1.02, 1.00–1.04)1. We found weak evidence supporting a causal effect of liability to type 2 diabetes (T2D) on HFall (ORMR−IVW = 1.04, 1.03–1.05 per doubling prevalence) as previously reported1,46; however, evidence of horizontal pleiotropy was observed (PMR–Egger intercept = 1 × 10−6), and the association was attenuated when estimated using pleiotropy-robust methods (ORMR–Egger = 0.98, 0.96–1.00, ORMR–WME = 1.01, 1.00–1.02 per doubling prevalence). We did not find evidence supporting a causal T2D effect on ni-HF risk (ORMR–IVW = 1.01, 0.99–1.02 per doubling prevalence).

Despite the kidney being prioritized in heritability enrichment analyses and colocalization of HF phenotypes with renal traits, neither lower eGFR or CKD were associated with ni-HFpEF in MR analysis (Figs. 3 and 5) and there was limited genetic correlation between CKD and HFall (rg = 0.17, P = 0.0002)47,48. In MR analysis, genetically predicted high systolic blood pressure (SBP) and higher body mass index (BMI) were associated with risk of all HF phenotypes, with the effect estimates differing between ni-HF subtypes stratified by LVEF; SBP had the largest effect on ni-HFrEF (ORMR–IVW = 1.92, 1.70–2.17 per s.d., Pinteraction versus ni-HFpEF estimate = 5.7 × 10−7), while BMI had the greatest magnitude of effect on ni-HFpEF (ORMR–IVW = 1.87, 1.70–2.06 per s.d., Pinteraction versus ni-HFrEF estimate = 0.002).

We identified potential causal effects of liability to chronic obstructive pulmonary disease (ORMR–IVW = 1.10, 95% CI = 1.06–1.14 per doubling prevalence) and lower forced expiratory volume on risk of HF (ORMR–IVW = 0.91, 95% CI = 0.88–0.95 per s.d.) and risk of ni-HF (ORMR–IVW = 0.91, 95% CI = 0.88–0.95 per s.d.). We also found evidence for effects of smoking behavior on HF (ORMR–IVW = 1.26, 95% CI = 1.19–1.33 per s.d. pack-years of smoking). Given the prioritization of ALDH2 at an identified GWAS locus for HF and ni-HF, we estimated the effects of higher alcohol consumption on HF risk. We found evidence of a positive association between higher alcohol consumption and risk of both HF (ORMR–IVW = 1.17, 95% CI = 1.06–1.29 per s.d. drinks per week) and ni-HFrEF (ORMR–IVW = 1.87, 95% CI = 1.37–2.55 per s.d. drinks per week), consistent with clinical studies that identify alcohol as a risk factor for dilated cariomyopathy49.

In MR analyses, lower LVEF, a trait linked to impaired myocardial contractility, was associated with a higher risk of all HF phenotypes except for ni-HFpEF (Fig. 6 and Supplementary Fig. 12).The strongest association was observed with ni-HFrEF (ORMR–IVW = 0.18, 0.12–0.27 per s.d. LVEF; ORMR–IVW = 1.83, 1.64–2.03 per doubling prevalence of DCM). Conversely, a higher liability to hypertrophic cardiomyopathy, a condition linked to increased myocardial contractility, was protective against ni-HFrEF (ORMR–IVW = 0.88, 0.85–0.91 per doubling prevalence of HCM).These findings are consistent with the reported opposing genetic associations between dilated and HCM and with observational studies reporting divergent association between lower and higher LVEF and the risk of HFrEF and HFpEF, respectively50,51. Despite the widely held hypothesis that diastolic dysfunction is the key driver for HFpEF, we did not find evidence supporting a causal association (ORMR–IVW for ni-HFpEF = 1.16, 0.91–1.48 per s.d. peak diastolic strain rate radial), although the limited set of diastolic parameters investigated are not established as clinical indices of diastolic dysfunction52.

Discussion

We report a GWAS of heart failure in 1.9 million individuals, including over 150,000 cases, with a focus on HF subtypes defined by etiology and LVEF. Our analysis identifies 66 distinct genetic loci associated with HF, including 37 new loci, with specific associations identified for ni-HFpEF. Through a comprehensive framework integrating functional genomic data at the variant, locus and genome-wide levels, we uncover new genes and pathways underlying HF and its subtypes. We leverage genetic association data to map the organs, tissues and cells contributing to HF pathogenesis and estimate subtype-specific effects of key risk factors.

Our study focusses on HF occuring in the absence of major secondary causes, specifically, previous myocardial infarction, revascularization, congenital heart disease and significant valvular heart disease. By focusing on cases without these common upstream drivers, our findings highlight mechanisms intrinsic to HF, including cardio–renal mechanisms and primary cardiomyopathies53. The ni-HFrEF phenotype captures both HF with mildly reduced (LVEF = 41–49%) and reduced (LVEF < 40%)54 ejection fraction and is enriched for DCM. Among the subtypes analyzed, this phenotype showed the highest heritability. In contrast, ni-HFpEF had low heritability but significant enrichment for renal tissues, reflecting a more heterogeneous phenotype consistent with a multisystem disorder.

Our comprehensive appraisal of the genetic and phenotypic etiology of HF and its subtypes offers key insights with tangible translational implications. First, by integrating functional genomic data to identify likely causal variants, genes and pathways, we identify new loci and genes, as well as prioritizing alternative candidate genes at previously reported HF loci. Molecular mechanisms regulating growth emerged as potentially important, including modulation of BACH1–YAP42 and KLF12–CDKN1A37 activity. Notably, we prioritized ALDH2 at a previously known HF locus, a gene that has been associated with increased cardiovascular risk via its effect on alcohol metabolism. We also found enrichment of the SASP pathway and identified IGFBP7 as a potential HFpEF gene. IGFBP7 is a well-established clinical biomarker of acute renal failure and a driver of cell cycle arrest, fibrosis and vascular rarefaction, presenting a potentially promising target for new therapeutics33,55,56. Second, our findings underscore the significance of noncardiac tissues in HF etiology. In particular, the kidney, vascular system and metabolic tissues were identified as important for HFpEF, although we did not find evidence of associations between CKD and eGFR in our MR analysis57. Third, our appraisal of cardiovascular risk factor effects across HF subtypes yielded several findings with potential implications for preventive strategies. For example, the weak evidence of the causal effect of T2D on HF suggests that observational associations between these conditions might partially stem from shared upstream mechanisms. Notably, certain diabetes treatments, such as SGLT2 inhibitors, have demonstrated substantial benefits in patients with HF without diabetes. This observation may reflect the drugs’ effects on common upstream pathways influencing both HF and diabetes58. We also found no evidence supporting the causal effects of CAD on liability to HF in the absence of previous myocardial infarction or revascularization (ni-HF).Finally, we provide new evidence implicating modifiable lifestyle factors, including smoking and alcohol consumption, in HF risk, highlighting potentially actionable targets for prevention59,60.

Our study has important limitations that suggest directions for future research. Despite efforts to include data from diverse populations, participants of non-European ancestry were underrepresented in our analysis. Additionally, only a minority of HF cases could be stratified by LVEF due to limited data availability, and key measures of ventricular size and diastolic function were largely unavailable. Large prospective studies with detailed phenotypic measurements are particularly needed for HFpEF to enable a more precise characterization of the genetic basis underlying HF subtypes. The variants and genes identified in the discovery analysis require functional validation, and studies with larger sample sizes are needed to improve the precision of effect size estimates. Our meta-regression analysis of sex ratio likely underestimates genetic effect modification by sex due to the limited statistical power and potential confounding by sex-related factors. To address this, sex-stratified genetic association studies are needed to uncover sex-specific effects. In our cross-trait genetic analyses, we note a minor sample overlap between samples in the current GWAS and those used in GWAS of CAD and atrial fibrillation; however, this overlap is unlikely to have significantly biased the MR estimates due to the limited scope of the overlap and the strength of the exposure instruments. Nevertheless, further studies are needed to model relationships between individual risk factors more effectively, incorporating multivariable and mediation analysis.

In summary, we used human genetic analysis to map the etiologies underlying HF and its subtypes, classified by etiology and LVEF. Our findings provide new evidence supporting the concept of HF as a multisystem clinical syndrome and provide insights into the underlying biological mechanisms. These insights may guide the development of more effective strategies for prevention and treatment.

Methods

Ethics statement

This study complies with the ethical regulations provided by the University College London Research Ethics Committee. All participating studies were ethically approved by local committees, and all study participants provided written informed consent (Supplementary Table 21).

Phenotype definition

The present meta-analysis included 1,946,349 individuals from 42 studies (Supplementary Table 21). Each study defined up to four HF phenotypes, which are as follows: (1) overall HF (HFall), (2) ni-HF, (3) ni-HFrEF and (4) ni-HFpEF. The HFall phenotype includes any diagnosis of HF based on a physician’s adjudication, hospital record review or diagnosis codes. ni-HF was defined by excluding antecedent ischemic, valvular and congenital heart diseases. ni-HFrEF was defined as ni-HF with LVEF < 50% based on cardiac imaging or diagnosis of LVSD at any point. ni-HFpEF was defined as ni-HF with LVEF ≥ 50% based on cardiac imaging without a record of LVSD at any point. Phenotyping was performed separately in each participating study (Supplementary Table 22), using a harmonized multimodal phenotyping framework (Supplementary Methods).

Genome-wide association analysis

Study-level GWAS was performed using logistic regression (for prevalent cases) or Cox proportional hazard (for incident cases) assuming additive genetic effect with adjustment for sex, age at DNA draw, genetic PCs and study-specific covariates (Supplementary Table 23). Study-level GWAS results were meta-analyzed using a fixed-effect IVW model implemented in METAL61. As study-level estimates include both results from logistic regression and Cox proportional hazard regression, we report the meta-analysis effect estimates as relative risk per additional risk allele. Sample genotyping and quality control of study-level and meta-analysis GWAS summary statistics are detailed in Supplementary Methods and Supplementary Data 2.

Identification of genetic susceptibility loci

To identify genetic susceptibility loci for HF, we performed a chromosome-wide stepwise conditional-joint analysis in each heart failure phenotype using Genome-wide Complex Trait Analysis (GCTA) software62. Conditionally independent variants across HF phenotype with joint P < 5 × 10−8 that are physically located within 500 kb of each other were aggregated into one locus set. A genomic locus was then defined as the genomic region within 500 kb upstream and downstream of the farthest variants in each aggregated set.

The identified genomic loci were labeled with an incremental one-based integer sequence based on phenotype order (HFall, ni-HF, ni-HFrEF and ni-HFpEF), chromosome and base pair positions. A genomic locus was declared as new if all conditionally independent variants within the locus and any of the sentinel variants reported at P < 5 × 10−8 in previous GWAS of HF1,2,3,6,7 were physically located more than 250 kb away and not in linkage disequilibrium (LD; r2 < 0.2). Analysis was performed using reference genotype from randomly sampled 10,000 UKB participants with proportionally matched ancestry composition (admixed American ancestry was not included due to unavailability in UKB and the small proportion of this ancestry in the overall sample).

Furthermore, genetic associations were categorized based on P values at cutoff values of P < 5 × 10−8 (genome-wide significant), P < 0.05/number of identified loci (replicated at a Bonferroni-adjusted threshold), P < 0.05 (nominally significant) and P ≥ 0.05 (no evidence of association). Loci associated with any ni-HF subtype at Bonferroni-adjusted or genome-wide significance thresholds were labeled as ni-HF loci, and the remaining were labeled as other HF loci. ni-HF loci that are associated with ni-HFrEF at Bonferroni-adjusted or genome-wide significance thresholds were labeled as ni-HFrEF loci, and the remaining were labeled as ni-HFpEF loci. Sentinel variants for HF were defined as conditionally independent variants with the lowest P value for association across HF phenotypes within a locus. For loci associated with both HFall and ni-HF phenotypes, sentinel variants identified in the HFall analysis were prioritized.

Cross-ancestry allelic effect heterogeneity

To assess the heterogeneity of allelic effects across ancestries, we performed a meta-analysis accounting for ancestry-specific allelic effects with a meta-regression framework implemented in MR-MEGA63. Using this technique, we estimated the P value for heterogeneity correlated with ancestry, the P value for residual heterogeneity and the concordance of genetic association P values from the fixed-effect meta-analysis across sentinel variants for the HFall phenotype (Supplementary Fig. 3).

Genetic effect modification by sex

We estimated potential genetic effect modification by sex across sentinel genetic variants for HF through a meta-regression analysis using the R meta v7.0.0 package64,65. For each sentinel variant, we extracted study-specific effect estimates across studies with available sex ratio information (Supplementary Table 1) and aligned the effect alleles to risk-increasing alleles in the GWAS meta-analysis. We then performed a meta-regression of study-level effect estimates on the proportion of males, weighted by the inverse variance of the effect size. We report the regression coefficient and the corresponding P value, whereby a positive value represents a positive correlation between the estimated additive genetic effect of risk-increasing alleles and a larger proportion of males.

Genetic architecture assessment

To assess genetic architecture and polygenicity across HF phenotypes, we compared quantiles of the expected and observed genome-wide genetic association P values using quantile–quantile plot and calculated the genomic inflation coefficient (λGC). To distinguish between polygenicity and confounding by population stratification, we calculated λGC assuming 1,000 participants (λGC-1,000), and estimated the LD score regression slope using LDSC66. To minimize bias, LDSC regression was performed using the European ancestry meta-analysis subset with reference genotypes from 10,000 random UKB European participants. We investigated the relationship between allelic effect on HF and allele frequency by plotting the relative risk per additional minor allele as a function of MAF and fitted two separate local polynomial regressions with locally estimated scatter-plot smoothing for groups of conditionally independent variants associated with increased and decreased risk of HF. To increase precision, we fitted the regression using conditionally independent variants associated with HFall at an FDR < 1%, estimated using the q value package in R67.

SNP-based heritability estimation

The proportion of variance in HF risk explained by common SNPs, that is, SNP-based heritability (h2SNP), was estimated from GWAS meta-analysis summary statistics using Linkage-Disequilibrium-Adjusted Kinships (LDAK) SumHer software68 with LDAK-Thin and BLD-LDAK heritability models69 on a liability scale. To minimize bias, we used the European ancestry meta-analysis subset with precomputed tagging files derived from 2,000 UKB European participants68. The conversion to liability scale was calculated using population prevalence derived from the meta-analysis sample and sample prevalence derived using correction for effective sample size assuming an equal number of cases and controls70.

Polygenic risk score analysis

We explored the association between PGS of HFall (PGSHF) and HFall risk in 347,235 UKB European participants, including 13,793 HFall cases. The PGSHF was constructed as a weighted sum of the allelic count of 1,012,059 genetic variants selected using the LDpred2-auto model71, with weights derived from the present GWAS meta-analysis of HF excluding UKB and reference genotype from 362,320 UKB European participants provided by LDpred2 authors. OR of HFall per s.d. of PGSHF was estimated using logistic regression with binary HFall status as response variable, standardized PGSHF as predictor and sex, age and first ten genetic PCs as covariates. To estimate the risks of HF in individuals with high PGSHF, we grouped participants into deciles of PGSHF and calculated ORs of the top decile as compared against the fifth and the first (bottom) decile using logistic regression.

Fine mapping and functional consequences of causal variants

We performed functionally informed fine mapping using PolyFun9 and sum of single effects (SuSiE)72 to identify causal variants within HF genetic susceptibility loci. We used precomputed, functionally informed prior causal probabilities of 19 million genetic variants based on a meta-analysis of 15 UKB traits9 and genome-wide association estimates from the current analysis to calculate prior causal probability proportional to per-SNP heritability of phenotype under analysis. The resulting estimates were used to calculate per-SNP PIP and to construct 95% credible sets of likely causal variants using the sum-of-single-effects fine-mapping model implemented in SuSiE, assuming at most five causal variants per locus. To minimize bias, we used effect estimates from the European ancestry meta-analysis subset and reference genotypes from 10,000 random UKB European participants. To assess the functional consequences of fine-mapped variants, we extracted variant-level information on the nearest gene(s), genic functions and CADD10 Phred score from ANNOVAR73 and OpenTargets Genetics14.

Prioritization of effector genes

To identify effector genes for HF, we implemented a two-step prioritization approach. In step 1, we identified candidate genes using a combination of the following three predictors: (1) PoPS15, (2) V2G—OpenTargets V2G score14,74 and (3) TWAS—multitissue TWAS using S-MulTiXcan16. Genes with the highest PoPS, highest V2G score or lowest TWAS P value within a locus were considered as candidate effector genes for HF.

In step 2, we further prioritized these genes using the following three Boolean (true/false) classifiers: (1) ABC—overlap with enhancer as predicted by ABC17,18 score > 0.02, (2) Mendel—association with enriched Mendelian disease term estimated using MendelVar20 and (3) Coloc—sharing of causal variants (colocalization) between gene expression in relevant tissue with at least one HF phenotype under study at posterior probability >0.8 estimated using R coloc package19 using gene expression data from Genotype-Tissue Expression project (GTEx, v.8)75.

In addition, we derived an overall predictor score based on weighted average of PoPS, highest V2G and −log10(P) MulTiXcan values with 2:2:1 weight ratio, scaled to 0–100 value using a quantile transformation with uniform output distribution as implemented in the Python scikit-learn76 library. Finally, for each locus, we ranked genes based on total classifier score (sum of true values) and the overall predictor score. Genes that are top-ranked or have a total classifier score ≥2 were prioritized as effector genes for HF. More details are provided in Supplementary Methods.

Tissue- and cell-type enrichment

To identify tissues and cell types involved in HF pathology, we estimated heritability enrichment of specifically expressed gene sets across 206 tissues and cell types77,78 within 12 organ/system categories (Supplementary Table 24). For each set, we performed a one-sided enrichment test for per-SNP heritability attributed to the given set, conditional on the set that includes all genes and a baseline model consisting of 52 genomic annotations using LDSC applied to specifically expressed genes (LDSC-SEG)79. For this analysis, we used the European ancestry meta-analysis subset with precomputed LD score weights derived from the 1000 Genomes European reference panel80.

Based on the observation that heart tissues are the major contributors to SNP heritability across HF phenotypes, we extended this analysis to 15 cell types derived from single-nucleus RNA-sequencing of 185,185 nuclei from 16 nonfailing human heart donors43 (Supplementary Table 25). Identification of specifically expressed genes across cardiac cell types is described in Supplementary Methods.

As sensitivity analysis, we further performed a gene-property enrichment analysis using MAGMA81,82. We first assigned SNPs to 20,260 coding genes from Ensembl v.92 with a 1-kb window and computed gene-level association statistics using the SNP-wise mean model. We then computed average gene expression per cell type and provided this as gene covariates to MAGMA to calculate a one-sided P value for enrichment of a given cell type conditioned on the average expression of other cell types.

Cellular expression of heart failure effector genes

To assess changes in the expression of HF effector genes in failing heart, we performed single-nucleus differential gene expression analysis of candidate effector genes within heart failure loci by comparing transcription levels in cell nuclei from nonfailing heart described above with cell nuclei from 28 failing heart donors diagnosed with end-stage DCM (ndonors = 12 and nnuclei = 142,490) or HCM (ndonors = 16 and nnuclei = 202,307) from Myocardial Applied Genetic Network34,83. Samples are processed using CellBender84 and Cell Ranger85. Differential gene expression was calculated using the limma–voom86,87 model adjusting for age and sex (see more details in Supplementary Methods).

Differential expression was defined using the following criteria: two-sided adjusted P < 0.01 (following Benjamini–Hochberg adjustment for multiple testing), sign concordance between estimates derived from CellBender and Cell Ranger quantifications and no background contamination in CellBender. To test whether differentially expressed genes in each cell type are overrepresented among prioritized HF effector genes, we performed a one-sided Fisher’s exact test implemented in R, using all genes tested in the given cell type as the background set.

Pathway enrichment

To identify biological pathways relevant to HF, we tested for overrepresentation of biological terms in HF effector gene set using g:Profiler88. To account for etiological differences and uncertainty in gene prioritization, we tested six gene sets constructed from combinations of three phenotypic classifications of loci (HF, ni-HF and other HF) and two gene prioritization categories (candidate and prioritized effector genes). For each gene set, we performed an unordered enrichment analysis with a one-sided Fisher’s test for biological terms in the Kyoto Encyclopedia of Genes and Genomes89, Reactome90, Wiki Pathways91 and Gene Ontology (GO)92,93, excluding terms with more than 2,000 genes. Multiple-testing adjustment was performed by multiplying the enrichment P values by the ratio of the approximate threshold given the number of genes in the set and the initial experiment-wide type-I error rate of 0.05, using the g:SCS algorithm in g:Profiler. Driver GO terms were identified using the g:Profiler clustering algorithm to represent groups of more specific terms.

Phenome-wide association of heart failure sentinel variants

To characterize pleiotropic effects of HF genetic loci across human diseases, we performed a phenome-wide association study (PheWAS) in 408,480 UKB European participants. We tested the associations between sentinel variants at 66 HF loci with 294 disease phenotypes derived using a curated phenotype definition and category39. For each phenotype, participants were classified as cases if at least one event was recorded in linked hospital admission, death certificate, primary care visit, self-reported cancer diagnosis, noncancer diagnosis or procedure history. For each phenotype–sentinel variant pair with at least 100 cases and 100 noncases (controls), we ran a case–control additive genetic association test in PLINK2 with a logistic regression model adjusted for sex, genotyping array and first ten genetic PCs. For presentation, we report effects per additional HF risk-increasing allele.

Pleiotropy network analysis

Using the PheWAS results, we performed a network analysis to evaluate the connections between HF genetic loci based on their pleiotropic effects across human diseases. First, we constructed a network dataset with nodes representing loci or phenotypes and edges representing unidirectional associations of the locus sentinel variant with a phenotype. Eigenvector centrality measure was calculated to estimate node influence within the network (Supplementary Fig. 13 and Supplementary Table 26). Furthermore, we performed a community detection analysis using a walktrap algorithm94 weighted by the absolute z score of the edges to identify distinct etiologic clusters underlying HF. The network analysis and visualization were performed using igraph, tidygraph and ggraph package in R.

Locus-specific pleiotropy assessment

We investigated the local pleiotropic effects of sentinel variants in 66 HF genetic loci on 24 related GWAS traits (Supplementary Table 27). Estimates for sentinel variants that were not reported in a target GWAS were substituted with those of a proxy variant with the highest LD r2 (requiring at least LD r2 > 0.8), identified using PLINK95 with LD reference panel derived from 10,000 random samples of UKB European participants. For presentation (Extended Data Fig. 10), effect estimates were converted to z scores derived from the regression coefficients (β) divided by their s.e., with direction of effect aligned to reflect the additive effect of HF risk-increasing allele. Loci and GWAS traits were ordered using hierarchical agglomerative clustering (implemented with the hclust function in R 4.2.0.) with average linkage method based on the Euclidean distance measure derived from absolute z scores.

Cross-trait colocalization

To test whether HFe and related traits share causal genetic variants, we performed pairwise cross-trait colocalization analysis using the R coloc package19. Analyses were performed across pairwise combinations of credible sets corresponding to the tested traits identified using SuSiE regression72. We used the default marginal prior probabilities for association with trait 1 (p1) and trait 2 (p2) of 10−4, as well as a prior probability of joint causal association (p12) of 10−5. A posterior probability of common causal variants (PPcoloc H4) > 0.8 in at least one pair of credible sets was considered as evidence of colocalization. The colocalization analysis was performed for each identified HF locus using an overlapping set of variants available in both GWAS of heart failure and trait under analysis situated within the locus. Genetic association estimates for HF were extracted from GWAS of the HFall phenotype for loci with PHF < 5 × 10−8 (locus 1–56) or otherwise from GWAS of ni-HF phenotypes for subtype-specific loci—ni-HF for locus 57–61, ni-HFrEF for locus 62–64 and ni-HFpEF for locus 65–66. Genetic association estimates for tested traits are extracted from the source GWAS (Supplementary Table 27).

Genetic correlation and MR

We tested genetic correlation across pairs of four HF phenotypes under analysis and 24 related GWAS traits using bivariate LDSC regression96. We further estimated the causal effects of each tested trait on four HF phenotypes under analysis using two-sample MR as implemented in the MendelianRandomization package97. MR instruments for the exposure traits were selected from overlapping genetic variants available in both GWAS of exposure and outcome traits using an LD-based clumping algorithm with a P value threshold of 5 × 10−8 and LD r2 threshold of 0.05 implemented in PLINK95. For each exposure trait, we estimated the causal association with each HF phenotype using the IVW MR estimator and performed sensitivity analyses with MR–Egger and WME98,99. An adjusted P < 0.01 following the Benjamini–Hochberg correction in IVW analysis with the consistent direction of effect in sensitivity analyses was considered supportive evidence of causality. Reference genotype from 10,000 random UKB European participants was used to perform these analyses, given the majority of European ancestry in the tested GWAS traits. Details of the tested traits, source GWAS and model parameters are provided in Supplementary Table 27.

Reporting summary

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