Polygenic prediction of cardiorespiratory fitness in the Trøndelag health study (HUNT) Authors: Karsten Øvretveit, Marie Klevjer, Ben M. Brumpton, Ulrik Wisløff, Kristian Hveem, Anja Bye Published: 2025 ABSTRACT Cardiorespiratory fitness (CRF) has a strong genetic component and low CRF is a major risk factor for cardiovascular morbidity and mortality. The purpose of this study was to develop and validate a polygenic score (PGS) for CRF (CRFPGS) and assess its associations with cardiovascular disease (CVD) and all-cause mortality. We hypothesized that the CRFPGS would demonstrate similar cardioprotective benefits as the CRF phenotype. Effect estimates from a genome-wide association study on directly measured CRF in the Trøndelag Health Study (HUNT; n = 4525) were used in a Bayesian regression framework to develop multiple PGSs in an independent cohort from the UK Biobank (n = 65,165). The top performing score was applied in the HUNT target cohort, excluding the discovery sample (n = 82,109). The PGS-CRF association varied considerably as a function of model fit and phenotypic accuracy. There was a difference of 1.55 [95% confidence interval: 1.26, 1.84] mL·kg−1·min−1 between the bottom and top decile of the CRFPGS. Moreover, a high CRFPGS demonstrated cardioprotective effects, with reduced risk for CVD, myocardial infarction, hypertension, and all-cause mortality. Additionally, in women, we observed that the CRFPGS predisposed to lower risk of heart failure and hypertrophic cardiomyopathy. We developed the first PGS for CRF using gold standard phenotypes and multiple independent cohorts. Genetic susceptibility to high CRF may have a clinically meaningful impact on the phenotype. The CRFPGS was better to identify individuals with slightly higher lifelong levels of CRF, which appears to protect against cardiovascular morbidity and mortality.Supplementary InformationThe online version contains supplementary material available at 10.1038/s41598-025-28894-7. Introduction Cardiorespiratory fitness (CRF), typically expressed as maximal (V̇O2max) or peak (V̇O2peak) oxygen uptake, has a strong and independent association with cardiovascular morbidity and mortality, with no apparent upper level of benefit1–4. It is a complex phenotype that is influenced by multiple physiological parameters, including pulmonary, cardiovascular, and mitochondrial function5. The performance of these intermediate phenotypes converges to set the individual’s upper limit of oxygen uptake. The genetic component of V̇O2max is estimated to be between 44 and 68% from twin and family studies6. V̇O2max is considered the gold standard measurement of CRF, and it is typically obtained by direct gas exchange analysis during a graded cardiopulmonary exercise test (CPET)7. Although common practice in clinical research, CPET is less feasible in large cohort studies and often replaced by indirect measurements. The largest study with direct measurements of CRF to date was conducted as a part of the third wave of the Trøndelag Health Study (HUNT), where a total of 4631 individuals completed a V̇O2peak test8. Subsequently, a genome-wide association study (GWAS) in this population suggested 38 novel genetic variants associated with V̇O2peak9. Functional annotation of the results confirmed previous genetic associations between variants predisposing to endurance, cardiac function, and cardiovascular disease (CVD). A polygenic score (PGS) aggregates genetic effect estimates from GWAS into a single metric that quantifies the liability to a specific trait. To date, no study has used gold standard measurements of CRF in polygenic prediction of the phenotype. Here, we use genetic and CRF data from multiple large independent cohorts to develop and test a PGS for CRF (CRFPGS). In addition to exploring the association between CRFPGS and indices of CRF and related phenotypes, we assessed its impact on several CVD outcomes. We hypothesized that the CRFPGS would be associated with estimated CRF (eCRF), and that it would demonstrate similar cardioprotective benefits as the CRF phenotype. Methods Participants The total study sample comprised 151,852 individuals (Fig. 1). Beginning in 1984, HUNT has enrolled over 230,000 individuals in four separate surveys10. Nearly 100,000 participants have contributed with biological material, of whom approximately 88,000 have provided DNA and consent for genotyping11. Detailed information about genotyping is available in the supplemental material. We included up to 86,687 (45,970 women) genotyped and phenotyped participants from the HUNT study. The base data included individuals with directly measured V̇O2peak who had participated in the HUNT3 Fitness Study (n = 4578, 2343 women)8,9. To avoid overfitting, these were excluded from the HUNT target sample (n = 82,109, 43,627 women). The UK Biobank (UKB) is a prospective study of 500,000 participants from the United Kingdom that collects questionnaire data, clinical measurements, and biological samples. All participants in the UKB have been genotyped and imputed12, and participants with genotype data, CRF, and physical activity (PA) measurements were included, resulting in a sample size of 65,165 (34,793 women).Fig. 1Overview of the cohorts used as base, tuning, and target data Overview of the cohorts used as base, tuning, and target data Ethics The UKB research protocol and study design were approved by the NHS National Research Ethics Service, and all study participants provided written informed consent (project ID 40135). Ethical approval was obtained from the Northwest Centre for Research Ethics Committee (MREC, 11/NW/0382). In Scotland, the UKB has approval from the Community Health Index Advisory Group. The study was approved by the regional committee for medical research ethics (2019/29771), the Trøndelag Health Study, the Norwegian Data Inspectorate, and the National Directorate of Health. It was conducted in accordance with Norwegian laws and the Helsinki declaration. All participants provided written informed consent. Phenotyping Disease outcomes were derived from health registry data spanning January 1999 through March 2020. Outcomes were defined in accordance with the 10th revision of the International Statistical Classification of Diseases and Related Health Problems (ICD-10) and grouped in disease categories (Table S2). In the HUNT3 Fitness Study, V̇O2peak was measured by trained personnel with a CPET8. Self-reported outcomes were obtained from questionnaire data (Table S3). In HUNT4 participants, we also obtained body composition data from bioelectrical impedance analysis (InBody 770, Cerritos, CA, USA) and plasma hemoglobin (Hb). We calculated eCRF in HUNT3 and HUNT4 participants using a previously published nonexercise prediction model that incorporates age, waist circumference (WC), resting heart rate (RHR), and a PA-index (calculation available in the supplements) to determine sex-specific eCRF13:\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Men}}:{1}00.{27} - \left( {0.{296 } \times {\text{ age}}} \right) - \left( {0.{369 } \times {\text{ WC}}} \right) - \left( {0.{155} \times {\text{RHR}}} \right) + \left( {0.{226} \times {\text{PA}} - {\text{index}}} \right)$$\end{document}\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$${\text{Women}}:{74}.{74} - \left( {0.{247 } \times {\text{ age}}} \right) - \left( {0.{259 } \times {\text{ WC}}} \right) - \left( {0.{114 } \times {\text{ RHR}}} \right) \, + \, \left( {0.{198 } \times {\text{ PA}} - {\text{index}}} \right)$$\end{document} In the UKB, a submaximal bicycle test was used to assess CRF in a subsample of ~ 100,000 participants. In short, the participants were divided into five groups by health risk, and the participants in the three healthiest groups were tested at 50%, 30% and at constant level of the predicted maximum workload. CRF was then determined as the net oxygen consumption based on body weight and maximum workload14. Polygenic score To develop a PGS for CRF, we used PRS-CS, a Bayesian regression framework with continuous shrinkage priors15. The base data was the genotyped HUNT3 Fitness Study participants9, and an independent UKB cohort was used as the tuning cohort. After testing multiple PGSs, the score with the strongest association with CRF was identified and subsequently applied in an independent HUNT target cohort (Table S1). A total of 1,104,501 SNPs (99.1%) overlapped between the PGS developed for CRF in the tuning and testing cohort (Table 1).Table 1Development flow of a polygenic score (PGS) for cardiorespiratory fitness (CRF).TypeCohortn SNPsPurposeBase dataHUNT3 Fitness study (n = 4525)14,036,359Provide summary statistics of genetic effect estimates for CRFTuning dataUK Biobank (n = 65,165)1,115,064*Association testing of multiple PGSs with different tuning parametersTarget dataHUNT1-4 (n = 82,109)1,104,501*Validation and additional testing of the PGS on relevant outcomesHUNT, the Trøndelag Health Study; SNP, single-nucleotide polymorphism. * Restricted to HapMap3 variants. Table 1Development flow of a polygenic score (PGS) for cardiorespiratory fitness (CRF).TypeCohortn SNPsPurposeBase dataHUNT3 Fitness study (n = 4525)14,036,359Provide summary statistics of genetic effect estimates for CRFTuning dataUK Biobank (n = 65,165)1,115,064*Association testing of multiple PGSs with different tuning parametersTarget dataHUNT1-4 (n = 82,109)1,104,501*Validation and additional testing of the PGS on relevant outcomesHUNT, the Trøndelag Health Study; SNP, single-nucleotide polymorphism. * Restricted to HapMap3 variants. Development flow of a polygenic score (PGS) for cardiorespiratory fitness (CRF). HUNT, the Trøndelag Health Study; SNP, single-nucleotide polymorphism. * Restricted to HapMap3 variants. Statistical analysis Genotype–phenotype and phenotype-phenotype associations were assessed with the Pearson product-moment correlation coefficient. Disease risk estimates were derived from Cox proportional hazards models using the R package “survival”16. Separate models were fitted for each disease endpoint based on complete-case data from hospital records spanning January 1999 through March 2020, with age on the time scale. For each outcome, the first recorded diagnosis defined the event. Participants were not excluded on the basis of pre-existing disease recorded prior to January 1999. We modelled hazard ratios (HRs) adjusted for sex and the first ten principal components of population structure calculated with PLINK 2.0. The odds of self-reported hyperglycemia and medication use were assessed with logistic regression models adjusted for age, age2, sex, and the first ten principal components. In both the survival and logistic models, the CRFPGS was standardized to have a mean of “0” and standard deviation (SD) of “1” to estimate the effect of each SD change in the PGS. We also compared observed eCRF levels and mean change from HUNT3 to HUNT4 between the top and bottom deciles of the CRFPGS distribution. Due to nonnormal data distributions, group differences were assessed with the Mann–Whitney U test, with the Bonferroni correction being used to account for multiple comparisons. Results are presented with 95% confidence intervals (CI) and p-values where applicable. p-values were interpreted as continuous indicators of evidence strength and conclusions were drawn based on the magnitude and precision of effect sizes. Statistical analyses were performed with R (version 4.2.3). Results Participants characteristics and genotype–phenotype associations Genotype–phenotype correlations for CRFPGS varied considerably between cohorts, with the strongest relationship being observed in participants who were part of the base data (Table 2). The mean eCRF in HUNT3 and HUNT4 was 35.7 ± 8.2 mL·kg−1·min−1 and 35.0 ± 8.8 mL·kg−1·min−1, respectively. The r between eCRF at HUNT3 and measured V̇O2peak in the HUNT3 Fitness subset was 0.68 (n = 3474) and 0.58 at HUNT4 (n = 3026). The CRFPGS also correlated with all eCRF predictors individually (Table 3). It was also inversely correlated with BMI in both HUNT3 (r = − 0.062 [95% CI: − 0.071, − 0.052]) and HUNT4 (r = − 0.060 [95% CI: − 0.069, − 0.051]). In HUNT4, we observed an inverse relationship between CRFPGS and body fat mass (r = − 0.076 [95% CI: − 0.085, − 0.067]) and visceral fat (r = − 0.078 [95% CI: − 0.087, − 0.069]), but not fat-free mass.Table 2Genotype–phenotype associations for CRFPGS.VariableCohortSample sizePearson’s r95% CIMeasured V̇O2peakHUNT3 Fitness study45780.6930.677, 0.707eCRF (HUNT3)HUNT3 Fitness study34740.3450.315, 0.374eCRF (HUNT4)HUNT3 Fitness study30260.3520.320, 0.383eCRF (HUNT3)All HUNT32,3810.1130.102, 0.124eCRF (HUNT4)All HUNT40,2870.0940.085, 0.104eCRF (HUNT3)All HUNT w/o GWAS28,9070.0380.027, 0.050eCRF (HUNT4)All HUNT w/o GWAS37,2610.0520.042, 0.062CI, confidence interval; HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; V̇O2peak, peak oxygen uptake; eCRF, estimated cardiorespiratory fitness; GWAS, genome-wide association study. Sample size reflect complete cases for both variables in each cohort.Table 3Associations between CRFPGS and eCRF predictors.HUNT3HUNT4Pearson’s r95% CIPearson’s r95% CIAge0.0160.007, 0.0250.0130.004, 0.021Waist circumference− 0.054− 0.063, − 0.045− 0.075− 0.084, − 0.066Resting heart rate− 0.035− 0.045, − 0.025− 0.041− 0.050, − 0.032Exercise frequency0.0430.034, 0.0520.0430.034, 0.052Exercise intensity0.0210.010, 0.0310.0220.012, 0.032Exercise duration0.0130.002, 0.0230.0200.010, 0.030Physical activity index0.0300.019, 0.0410.0330.023, 0.043CRFPGS, polygenic score for cardiorespiratory fitness; eCRF, estimated cardiorespiratory fitness; HUNT, the Trøndelag Health Study; BMI, body mass index; CI, confidence interval. All coefficients derived from individuals who were not a part of the genetic base data. Table 2Genotype–phenotype associations for CRFPGS.VariableCohortSample sizePearson’s r95% CIMeasured V̇O2peakHUNT3 Fitness study45780.6930.677, 0.707eCRF (HUNT3)HUNT3 Fitness study34740.3450.315, 0.374eCRF (HUNT4)HUNT3 Fitness study30260.3520.320, 0.383eCRF (HUNT3)All HUNT32,3810.1130.102, 0.124eCRF (HUNT4)All HUNT40,2870.0940.085, 0.104eCRF (HUNT3)All HUNT w/o GWAS28,9070.0380.027, 0.050eCRF (HUNT4)All HUNT w/o GWAS37,2610.0520.042, 0.062CI, confidence interval; HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; V̇O2peak, peak oxygen uptake; eCRF, estimated cardiorespiratory fitness; GWAS, genome-wide association study. Sample size reflect complete cases for both variables in each cohort. Genotype–phenotype associations for CRFPGS. CI, confidence interval; HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; V̇O2peak, peak oxygen uptake; eCRF, estimated cardiorespiratory fitness; GWAS, genome-wide association study. Sample size reflect complete cases for both variables in each cohort. Table 3Associations between CRFPGS and eCRF predictors.HUNT3HUNT4Pearson’s r95% CIPearson’s r95% CIAge0.0160.007, 0.0250.0130.004, 0.021Waist circumference− 0.054− 0.063, − 0.045− 0.075− 0.084, − 0.066Resting heart rate− 0.035− 0.045, − 0.025− 0.041− 0.050, − 0.032Exercise frequency0.0430.034, 0.0520.0430.034, 0.052Exercise intensity0.0210.010, 0.0310.0220.012, 0.032Exercise duration0.0130.002, 0.0230.0200.010, 0.030Physical activity index0.0300.019, 0.0410.0330.023, 0.043CRFPGS, polygenic score for cardiorespiratory fitness; eCRF, estimated cardiorespiratory fitness; HUNT, the Trøndelag Health Study; BMI, body mass index; CI, confidence interval. All coefficients derived from individuals who were not a part of the genetic base data. Associations between CRFPGS and eCRF predictors. CRFPGS, polygenic score for cardiorespiratory fitness; eCRF, estimated cardiorespiratory fitness; HUNT, the Trøndelag Health Study; BMI, body mass index; CI, confidence interval. All coefficients derived from individuals who were not a part of the genetic base data. Correlation coefficients may be influenced by nonlinearities in the genotype–phenotype relationship, but the weighted average difference in eCRF between the bottom and top deciles of the CRFPGS in HUNT3 (1.43 [95% CI: 1.01, 1.86] mL·kg−1·min−1, n = 2891) and HUNT4 (1.66 [95% CI: 1.26, 2.06] mL·kg−1·min−1, n = 3727) was 1.55 [95% CI: 1.26, 1.84] mL·kg−1·min−1 (Fig. 2; Table 4).Fig. 2Cardiorespiratory fitness for each decile of the CRFPGS at HUNT3 and HUNT4. CRFPGS, polygenic risk score for cardiorespiratory fitness; HUNT, the Trøndelag Health Study. Data presented as mean with 95% confidence intervals for each decile.Table 4Descriptive characteristics of the bottom and top decile of the CRFPGS at HUNT3 and HUNT4.HUNT3HUNT4Bottom decileTop decileBottom decileTop decileWomen/men1687/12041618/12732185/15422063/1663Age (years)53.9 ± 15.254.4 ± 16.450.9 ± 12.851.6 ± 13.4BMI (kg/m2)27.6 ± 4.5*26.4 ± 4.0*27.4 ± 4.7*26.4 ± 4.2*eCRF (mL·kg−1·min−1)34.8 ± 8.0*36.2 ± 8.5*34.0 ± 8.5*35.7 ± 9.0*SBP (mmHg)131 ± 19131 ± 19129 ± 18128 ± 19DBP (mmHg)74 ± 1173 ± 1174 ± 1073 ± 10PA index10.2 ± 9.2*11.1 ± 9.2*10.6 ± 9.4*11.6 ± 9.6*HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; BMI, body mass index; SBP, systolic blood pressure; DBP, diastolic blood pressure; PA index, physical activity index. Individuals with data on both CRFPGS and estimated cardiorespiratory fitness (eCRF) were included in the comparison; * p < 0.004 (0.05/12) between deciles. Cardiorespiratory fitness for each decile of the CRFPGS at HUNT3 and HUNT4. CRFPGS, polygenic risk score for cardiorespiratory fitness; HUNT, the Trøndelag Health Study. Data presented as mean with 95% confidence intervals for each decile. Table 4Descriptive characteristics of the bottom and top decile of the CRFPGS at HUNT3 and HUNT4.HUNT3HUNT4Bottom decileTop decileBottom decileTop decileWomen/men1687/12041618/12732185/15422063/1663Age (years)53.9 ± 15.254.4 ± 16.450.9 ± 12.851.6 ± 13.4BMI (kg/m2)27.6 ± 4.5*26.4 ± 4.0*27.4 ± 4.7*26.4 ± 4.2*eCRF (mL·kg−1·min−1)34.8 ± 8.0*36.2 ± 8.5*34.0 ± 8.5*35.7 ± 9.0*SBP (mmHg)131 ± 19131 ± 19129 ± 18128 ± 19DBP (mmHg)74 ± 1173 ± 1174 ± 1073 ± 10PA index10.2 ± 9.2*11.1 ± 9.2*10.6 ± 9.4*11.6 ± 9.6*HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; BMI, body mass index; SBP, systolic blood pressure; DBP, diastolic blood pressure; PA index, physical activity index. Individuals with data on both CRFPGS and estimated cardiorespiratory fitness (eCRF) were included in the comparison; * p < 0.004 (0.05/12) between deciles. Descriptive characteristics of the bottom and top decile of the CRFPGS at HUNT3 and HUNT4. HUNT, the Trøndelag Health Study; CRFPGS, polygenic score for cardiorespiratory fitness; BMI, body mass index; SBP, systolic blood pressure; DBP, diastolic blood pressure; PA index, physical activity index. Individuals with data on both CRFPGS and estimated cardiorespiratory fitness (eCRF) were included in the comparison; * p < 0.004 (0.05/12) between deciles. Moreover, participants in the bottom decile of the CRFPGS had a 7% greater drop in eCRF from HUNT3 to HUNT4 compared to those in the top decile (p = 0.04), indicating that the CRFPGS may reflect preserved CRF throughout the lifespan. Although the CRFPGS reflects phenotypic differences at the tails of the distribution, it appears to be less suited to identify individuals with the highest values (Table 5).Table 5Frequency of the top 100 eCRF values in each PGS decile.Decile12345678910N512138812109518Mean eCRF (mL·kg−1·min−1)62.264.163.063.463.263.764.263.063.463.0eCRF, estimated cardiorespiratory fitness; PGS, polygenic score. Table 5Frequency of the top 100 eCRF values in each PGS decile.Decile12345678910N512138812109518Mean eCRF (mL·kg−1·min−1)62.264.163.063.463.263.764.263.063.463.0eCRF, estimated cardiorespiratory fitness; PGS, polygenic score. Frequency of the top 100 eCRF values in each PGS decile. eCRF, estimated cardiorespiratory fitness; PGS, polygenic score. PGS associations with disease outcomes Survival models for multiple disease outcomes suggested that a genetically high CRF may offer cardiovascular protection (Fig. 3). One SD increase in the CRFPGS was associated with reduced incidence of CVD, myocardial infarction (MI), hypertension, and all-cause mortality for both sexes, as well as heart failure (HF) and hypertrophic cardiomyopathy (HCM) in women.Fig. 3Cardiovascular disease risk CRFPGS, polygenic risk score for cardiorespiratory fitness; CI, confidence interval, SD; standard deviation. Disease outcomes were obtained from hospital records spanning January 1999 through March 2020 in participants from the Trøndelag Health Study that were not included in the base data (HUNT; n = 82,109). The models were adjusted for the first ten principal components of population structure with age as the time scale. Cardiovascular disease risk CRFPGS, polygenic risk score for cardiorespiratory fitness; CI, confidence interval, SD; standard deviation. Disease outcomes were obtained from hospital records spanning January 1999 through March 2020 in participants from the Trøndelag Health Study that were not included in the base data (HUNT; n = 82,109). The models were adjusted for the first ten principal components of population structure with age as the time scale. Higher CRFPGS was also associated with reduced frequency of antihypertensive and cholesterol-lowering medication, as well as self-reported hyperglycemia (Fig. 4). We observed an association between eCRF and Hb at HUNT4 (r = 0.181 [95% CI: 0.171, 0.191]), but no such relationship was apparent between CRFPGS and Hb (p = 0.11). Odds of hyperglycemia and use of lipid-lowering and blood pressure drugs. CRFPGS, polygenic risk score for cardiorespiratory fitness; CI, confidence interval; SD, standard deviation. The models were adjusted for age, age2, and the first ten principal components of population structure. The outcomes are derived from questionnaires distributed to participants in the latest Trøndelag Health Study (HUNT) survey (HUNT4) that were not included in the base data. The phrasing of each questionnaire item is available in the supplementary material (Table S3). Discussion By using large, independent cohorts, we developed and tested a genome-wide PGS for CRF and found potentially clinically meaningful differences between high- and low-PGS carriers, as well as reduced risk of CVD for each SD increase in the CRFPGS. Although the phenotypic correlation was low in the target sample, our findings suggest that a genetic predisposition to elevated CRF is associated with reduced risk of all-cause mortality, CVD, MI, HF, HCM, and hypertension, as well as self-reported use of antihypertensive drugs, cholesterol drugs and hyperglycemia, indicating shared genetics between CRF and these outcomes. The pathogenesis of CVD is influenced by both genetic and environmental factors, and the genetic predisposition is highly heterogenic and complex17,18. Low CRF has previously been identified as an important risk factor for both all-cause mortality and CVD, and different subtypes of CVD including MI, hypertension, atrial fibrillation (AF), stroke, and HF19. However, there are few studies investigating the genetic contribution to these relationships, mainly due to lack of large cohorts with directly measured CRF. The association between the CRFPGS and ACM, CVD, MI, HF, HCM, and hypertension observed in the present study suggests shared genetics between CRF and these diseases. In a Swedish nationwide twin study they found indications that genetic factors influence the association of CRF with CVD and all-cause mortality, emphasizing the importance of the genetic contribution to CRF, independent of the protective effect of exercise-induced CRF20. The CRFPGS was associated with lower risk of HF in women, but not in men, suggesting sex-specific differences in the overlapping genetic architecture between CRF and HF. The CRFPGS was also associated with lower risk for HCM in women, albeit with some uncertainty around the estimate due to few events (372 cases). A similar, but weaker association was observed in men. Left ventricular function is one of the most important components of a high CRF21. Furthermore, a reversal of left ventricular remodeling has been seen alongside exercise-induced improvements in V̇O2peak22. The association between CRFPGS and HCM risk may suggest that the CRFPGS captures genetic signals related to cardiac function. Indeed, the genes MYH7 and MYBPC3 are responsible for nearly 70% of all HCM mutations, with the MYH7 also being associated with CRF23, indicating a genetic overlap with the CRFPGS. Patients with HCM are recommended to perform a CPET to assess the complex pathophysiology and severity of disease, as V̇O2peak has a great prognostic value for HCM patients24. The CRFPGS was not associated with either increased or decreased risk of AF. It has been shown that exercise has a U-shaped association with AF, meaning that both people with low and (extremely) high exercise levels are at increased risk25,26. Interestingly, a genetically high RHR appears to protect against AF despite increasing the risk of other cardiovascular diseases27. Thus, although RHR is often used as a proxy for fitness status, the genetic contribution to RHR-associated AF risk is contradictory. The lack of association between the CRFPGS and AF might indicate that the previously reported inverse relationship between CRF and AF risk could be due to exercise-induced improvements in CRF, rather than a genetically high CRF. The CRFPGS had comparable and even slightly stronger associations with some of the individual predictors in the eCRF model than the eCRF itself, as well as anthropometric variables such as BMI and body fat mass. Thus, its cardioprotective benefits may in part be driven by factors such as a more favorable body composition, or higher PA, either at the genotype or phenotype level. Indeed, genetic liability to sedentary behavior is associated with CVD28, whereas genetically influenced leisure-time PA has been shown to protect against several metabolic and cardiovascular outcomes29. The lack of association between CRFPGS and Hb could also indicate that the PGS is not a direct predictor of oxygen transport capabilities, but rather a higher CRF as a function of lower body mass, in line with recent findings on the relationship between genetically predicted body composition and V̇O2max30. Although the evidence on the relationship between genetically predicted V̇O2max and a more favorable anthropometric and metabolic profile is equivocal30, our observation that a higher CRFPGS is associated with lower BMI, body fat mass, and odds of hyperglycemia is consistent with the findings of others23. A PGS is a single metric that incorporates multiple genetic effects to estimate the genetic liability for a specific trait. Since it can be calculated at birth, it can also proxy the time exposed to a beneficial or detrimental risk factor, e.g., high blood pressure or a high CRF. Although cross-sectional differences in phenotypic expression might be small, such as the ~ 1.6 mL·kg−1·min−1 observed between the bottom and top decile of the CRFPGS in the present study, this can have substantial influences on outcomes in disease cohorts31. Notably, an increase of 1 mL·kg−1·min−1 in V̇O2peak is associated with a ~ 15% decrease in mortality risk in men and women with coronary heart disease32. The bottom decile of the CRFPGS had a 7% greater drop in eCRF from HUNT3 to HUNT4 compared to the top decile, indicating that genetics may help preserve CRF throughout the lifespan and reduce disease risk. Indeed, small differences in cardiovascular traits between the tails of a polygenic risk distribution have been observed from the first years of life33. If we consider the area under the CRF curve to be the main driver of CRF-mediated disease protection, a slightly elevated CRF due to genetics may offer considerable protection towards the later stages of life. The top decile had higher PA indices and lower BMI, which indicates that the CRFPGS also captures signals related to PA patterns and body composition. However, completely disentangling genetic and environmental effects in these risk constructs for phenotypes that are a product of both is challenging, and a PGS will typically also contain environmental effects on modifiable phenotypes34. Using independent cohorts for each step of PGS development is important to avoid overfitting35. The lack of gold-standard CRF data in large cohorts makes the development of a CRFPGS challenging. In our three populations, available phenotypes were directly measured V̇O2peak with CPET, estimated V̇O2peak from submaximal exercise data, and predicted V̇O2peak from nonexercise data. To provide the most accurate base data for the PGS calculation, we used effect estimates from a GWAS on directly measured participants previously published by our group9. By performing association testing in all three cohorts, we illustrate how different the magnitude of genotype–phenotype correlation is. As expected, the CRFPGS showed the best fit in the population the genetic effect estimates were derived from. It had the lowest phenotypic correlation in the UKB tuning cohort, but although the strength of the coefficient increased about 3–5 times in the HUNT target sample, the relationship remained weak. This was likely influenced by a nonlinearity in the genotype–phenotype relationship, as evident by the eCRF in the 5th to 8th decile of the CRFPGS. In addition, lack of validation cohorts with directly measured CRF and small sample size in the base data might have contributed to the low phenotypic correlation. Strengths of the study include the use of independent cohorts to develop the CRFPGS, including the largest dataset on directly measured V̇O2peak in genotyped participants; the application of a novel Bayesian PGS framework to test multiple tuning parameters; and the use of registry-based disease outcomes with time to event. The study was limited by the intricate and highly polygenic genetic architecture of the CRF phenotype, which contributed to a low genotype–phenotype association in the target sample. Moreover, the reliance on estimated CRF, which only explains about ~ 60% of trait variance for the nonexercise model and ~ 50% (Pearson’s r range 0.68–0.74) for the submaximal exercise test36, makes it challenging to validate and quantify the true predictive accuracy of the CRFPGS. Expressing V̇O2max relative to body weight (mL·kg−1·min−1) may also disproportionately penalize heavy individuals37 and also explain the correlation between the CRFPGS, BMI, and indices of body fat. In addition to larger sample sizes with valid CRF phenotypes, elucidating the genetic basis of relevant intermediate phenotypes, e.g., left ventricular and mitochondrial function, may be required to more accurately predict CRF with genetic data. Conclusion We developed the first genome-wide PRS for CRF using a Bayesian regression framework and large, genotyped cohorts with either measured or estimated CRF. Despite a weak correlation with the CRF phenotype, the CRFPGS was associated with cardiovascular and metabolic benefits. A genetic susceptibility to a high CRF appears to lead to slightly higher lifelong levels of CRF within the normal range, rather than those with supraphysiological CRF levels. The CRFPGS appears to have cardioprotective effects by reducing the risk of CVD, MI, HF, HCM, hypertension, as well as all-cause mortality. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5