Estimating genetic correlation jointly using individual-level and summary-level GWAS data Authors: Yiliang Zhang, Wei Jiang, Jiangnan Shen, Youshu Cheng, Yixuan Ye, Qiongshi Lu, Hongyu Zhao Published: 2026 ABSTRACT AbstractWith increasing access to individual-level data from genome-wide association studies, it is now common for researchers to have individual-level data of some traits, whereas for other traits only publicly released summary statistics are available due to privacy and safety concerns. Current genetic-correlation methods require both traits to be in the same format. When one trait has individual-level data and the other has only summary statistics, researchers often convert individual-level data to summary statistics and then apply summary-based methods, which is inefficient and can lose information. Here we introduce GENJI, a method for estimating within-population or transethnic genetic correlation based on individual-level data for one trait and summary-level data for the other. Across extensive simulations and real data analyses, GENJI produces more reliable and efficient estimation than summary data-based methods. We further show that more accurate genetic correlation estimation can improve cross-population polygenic risk prediction.Subject terms: Genome-wide association studies, Statistics Introduction Genetic correlation analysis, which quantifies the correlation of additive genetic effects on different traits across a set of genetic markers, has gained popularity owing to the remarkable success of genome-wide association studies (GWAS) in the past 20 years1. Compared with traditional family-based approaches2, estimating genetic correlation from GWAS data does not require samples from large pedigrees. The phenotypes of interest also do not have to be measured on the same individuals. These advances have effectively increased the sample size and improved statistical power in genetic correlation studies. Consequently, genetic correlation estimation has become a routine for post-GWAS analysis3 and has shed light on the shared genetic basis of numerous human complex traits and diseases4,5. A plethora of GWAS-based methods for genetic correlation estimation have been developed6, such as genome-based restricted maximum likelihood7 (GREML) for individual-level data and linkage disequilibrium (LD) score regression8 (LDSC) for summary-level data. The GREML approach is based on variance components estimation in a linear mixed model framework, which needs individual-level genotype and phenotype data as input. Many computational tools have been developed to implement GREML9–11, which mainly differ by the algorithms used for likelihood optimization. LDSC, in comparison, only requires GWAS summary data and is thus more broadly used in GWAS analyses. Web servers have also been built to facilitate the computation and visualization of genetic correlations12. Despite the popularity, estimates of LDSC are substantially less precise with larger standard errors compared to GREML6,13 due to information loss from individual-level data to summary statistics. This suggests that when individual-level data are available for both traits, GREML (or similar methods based on individual-level data) should be the method of choice. In addition, transethnic genetic correlation is recently introduced to quantify how the genetic architecture of complex traits varies across populations14. Summary statistics-based approaches are particularly less reliable for transethnic genetic correlation due to typically smaller sample sizes of non-European GWAS and a lack of robustness to mismatched LD between GWAS samples and reference panels. Even for analysis of GWAS from the same population, summary statistics-based methods struggle when the input GWAS have limited power. In many biomedical settings, researchers collect individual-level data for complex or expensive phenotypes in small cohorts. Examples include neuroimaging measures, molecular biomarkers, or pharmacogenomic response variables. These traits are often not measured at biobank scale but are highly informative. Currently, if individual-level data are available for a trait of interest, in order to test its genetic correlations with published GWAS, researchers would have to first generate summary data for the trait and then estimate genetic correlations based on the summary statistics. Such a procedure is conspicuously inefficient. Our main goal in this study is to find a way to jointly model individual-level data (e.g., for a disease with limited samples or a GWAS in minority populations) and summary statistics (e.g., from published meta-analyses) to efficiently estimate genetic correlation, which could be used to test whether these traits share genetic determinants with large-scale phenotypes such as BMI, blood pressure, or lipid levels, potentially identifying cost-effective proxies for early screening or risk prediction. In this work, we introduce GENJI (Genetic correlation EstimatioN Jointly using Individual-level and summary data), a method that estimates genetic correlation with individual-level data for one trait and summary-level data for the other trait. This enables researchers to explore whether a common and inexpensive trait is genetically correlated with a complex or costly phenotype, potentially supporting proxy-based inference or indirect selection strategies grounded in classical quantitative genetics. Through extensive simulations, we demonstrate that GENJI provides statistically rigorous and computationally efficient inference for both within-population and transethnic genetic correlations and substantially outperforms summary data-based methods. We demonstrate the effectiveness of GENJI through applications to many datasets, including UK Biobank (UKBB)15, the Wellcome Trust Case Control Consortium (WTCCC)16, the Northern Finland Birth Cohorts program (NFBC), and numerous GWAS of complex human traits spanning European, African, and East Asian populations from published studies and BioBank Japan (BBJ)17,18. While existing methods such as GREML9 and MTG210 provide accurate estimates of genetic covariance when individual-level data are available for both traits, such access is often infeasible in real-world settings due to privacy constraints, cohort access restrictions, or population mismatches. In contrast, GENJI is designed for a hybrid-access setting—where one trait has individual-level data and the other only summary statistics—a common scenario in cross-biobank or trans-ethnic studies. This work addresses this underexplored but practically important use case by providing a robust and efficient framework for estimating genetic correlation when joint individual-level modeling is not possible. Results Overview of GENJI Genetic correlation (covariance) is the correlation (covariance) of additive genetic components of complex traits across a set of genetic markers. It is commonly used as an informative metric to quantify the shared genetic basis between two traits and has also been extended to study the shared and unique genetic effects on the same trait in multiple populations. In this paper, we refer to the former as within-population genetic correlation and the latter as transethnic genetic correlation. In what follows, without stated explicitly, within-population genetic correlation and transethnic genetic correlation are collectively referred to as genetic correlation. The proposed method can estimate both types of genetic correlations. The primary contribution of GENJI is to enable genetic correlation estimation when researchers have access to individual-level data for one trait and only GWAS summary statistics for another. This hybrid-access setting is increasingly common and cannot be addressed by existing methods. In what follows, we demonstrate that GENJI produces valid and efficient estimates under this data configuration, and we benchmark its performance relative to both individual-level and summary-level methods. Genetic correlation rg is genetic covariance ρg normalized by SNP heritability, i.e., rg=ρg/h12h22, where h12 and h22 are the heritability of the two traits, respectively. Assume that we have individual-level GWAS data of the first trait (referred to as study 1 in the following) but only summary-level data of the second trait (referred to as study 2 in the following). For the first trait, we model phenotype ϕ as the sum of genetic component Xβ and random noise, where X and β represent genotype matrix and genetic effects, respectively. Denote ϕij as the phenotype of individual j from study i and z2 as the vector of z-scores in study 2. The LD matrix for study 2 can be estimated by an external reference panel or even genotype data in study 1 if they are conducted on the same population. We compute expectations and variances of XiTz2ϕ1i and apply weighted least squares to estimate genetic covariance: where Xi is the genotype of individual i in study 1, ns is the number of individuals shared by the two GWAS studies, and Xs is the genotype of the ns shared individuals. Introduction of more notations and detailed derivations are provided in the “Methods” section. Instead of the block jackknife used in LDSC8, the estimates of GENJI can use the theoretical standard error from weighted regression because individuals in study 1 are presumably unrelated to each other. Conceptually, GENJI estimates genetic correlation by aligning individual-level genotype–phenotype associations from one dataset with genome-wide summary statistics from another. It computes the expected cross-product between each individual’s genotype and the summary-based z-scores of the second trait, and then models the extent to which these values reflect shared genetic signal rather than noise. This approach allows GENJI to extract information about genetic covariance without requiring individual-level data for both traits. In practice, there can be some individuals shared by the two studies. Although we have individual-level data for the overlapped samples in study 1, we assume the phenotypes of them for study 2 are unknown. When we know which samples are shared by the two studies, we assume that the first ns samples in both studies are the overlapped samples. Similar to LDSC8, sample overlap inflates the covariance between ϕ1i and XiTz2 for each overlapped individual. Genetic covariance ρg and noise covariance ρe appear in different terms in the expectation of XiTz2ϕ1i in (1) and hence can be distinguished by regression (“Methods”). When we don’t have the information of which samples are shared by the studies (i.e., we cannot assume the first ns samples are common in both studies), we extend GENJI by first estimating nsρe and then adjusting the bias in the estimation of ρg caused by ρe (“Methods”). We refer to this version of GENJI as GENJI_adjusted in the simulation results. We also compared the computational complexity of different methods to estimate genetic correlation as a function of the number of individuals. We changed the sample sizes from 1748 to 270,000. We included GREML (GCTA9), LDSC8, Popcorn14, and GENJI in this comparison. For the dataset of 270,000 individuals, GENJI can fit its model in 6 h, while GREML took more than 24 h to fit. Our results showed that GENJI significantly improved computational complexity compared to GREML across all the settings. Summary-statistic-based methods, which didn’t take individual genotype and phenotype data as input, consistently fit the model in only a few minutes. The details of our simulation settings can be found in “Methods”. (Supplementary Fig. 19 and Supplementary Data 6). These runtime benchmarks confirm that GENJI is computationally tractable even at biobank scale, making it a practical choice for hybrid-access analyses. Simulations for within-population genetic covariance We performed simulations to assess the performance of GENJI on within-population genetic covariance estimation. We compared GENJI with two state-of-the-art methods, LDSC and GREML, on both quantitative and binary traits. We provided GREML with individual-level data, LDSC with summary-level data, and GENJI with individual-level data for the first study, while summary-level data for the second one. To assess the robustness of different methods to sample overlap, we changed the proportion of overlapping samples from zero to half of the sample sizes of the studies. We used real genotype data from WTCCC16 to simulate traits. We equally divided 15,918 samples in the dataset into two subsets, which we denote as set 1 and set 2, respectively. There are no shared individuals between set 1 and set 2. Randomly combining the samples from set 1 and set 2, we created set 3, set 4, and set 5, which had 10, 25, and 50% overlapping samples with set 1, respectively. The simulated GWASs were conducted on these subsets, and the sample sizes of them were all equal to 7959. The SNP effects were generated by a bivariate normal distribution with genetic covariance ranging from 0 to 0.25, and heritability was fixed at 0.5. Each setting was repeated 100 times. Detailed simulation settings and quality control procedures are described in the “Methods” section. The three competing methods in this section showed well-calibrated type I error rates when the true covariance was zero (Supplementary Figs. 1–4). LDSC and GENJI provided unbiased estimates across all the settings. GREML overestimated genetic covariance for the 10 and 25% sample overlap settings (Fig. 1 and Supplementary Data 1). Simulation results for within-population genetic covariance estimation. We compare the performance of LDSC, GREML, and GENJI by point estimation, type-I error, and statistical power. Boxplots show the quantiles of the estimates of genetic covariance in different settings across n = 100 independent computational simulation replicates for each method and each true genetic covariance setting. The centre line denotes the median, box limits denote the first and third quartiles, whiskers extend to the most extreme values within 1.5 times the interquartile range, and points beyond the whiskers denote outliers. Each simulated GWAS included 7959 individuals from WTCCC. The red dashed lines represent the true value of genetic covariance. We use the proportion of p-values that are less than 0.05 to estimate type-I error or statistical power when true parameters are zero or nonzero, respectively. A, B Two GWASs are simulated on two non-overlapping datasets (set 1 and set 2). Because there is no shared individual between the studies, we also include LDSC that constrain the intercept to zero in the comparison and denote it as “LDSC_no_intercept”. C–H compare LDSC, GREML, and GENJI using GWASs (C, D) with a 10% sample overlap (set 1 and set 3), E, F with a 25% sample overlap (set 1 and set 4), and G, H with a 50% sample overlap (set 1 and set 5). GENJI produces more accurate estimates than LDSC and estimates that are consistent with GREML, while enabling use of hybrid data settings that LDSC and GREML cannot support. Source data are provided as a Source Data file. In the case of a zero-overlapping sample (set 1 and set 2), we also included LDSC with an intercept fixed at zero (denoted as “LDSC_no_intercept”) in the comparison. When there was no overlapping sample, LDSC_no_intercept, GREML, and GENJI had comparable estimates. At the same time, the statistical power of GREML or GENJI was larger than that of LDSC_no_intercept when the true value of genetic covariance was relatively low. Estimates of LDSC without constraint had substantially larger variance and lower power than the other methods (Fig. 1A, B). This indicates that although no extra information for point estimation of genetic covariance can be acquired from individual-level data compared with summary-level data for studies with disjoint cohorts, the standard errors from parametric methods are more stable than those from resampling-based methods (e.g., block jackknife). When we did not fix the intercept for LDSC, the variability of intercept estimates increased the standard error of genetic covariance by 30% compared to LDSC with a constraint, a phenomenon that was also reported in the paper of LDSC8. For simulations in set 1 and set 3 (10% overlap), set 4 (25% overlap), and set 5 (50% overlap), GENJI outperformed LDSC when the overlapping samples were less than half of the sample sizes of the studies (Fig. 1C–F) and achieved comparable results with LDSC when the proportion of overlapping samples was 50% (Fig. 1G, H). GENJI was as powerful as GREML and had more accurate estimates under a low sample overlap setting (< 10% sample overlap). Under a moderate sample overlap setting (25%), the power of GENJI was between LDSC and GREML. We observe that the advantage of GENJI over LDSC reduces as the overlapped sample size increases, and reaches zero when half of the study 1 samples are included in study 2 due to collinearity in the terms of (1) under a large proportion of sample overlap (Supplementary Fig. 6). We also performed additional simulations to evaluate the performance of the methods in more general settings. Firstly, for phenotypes based on the LDAK models, the point estimates obtained from GENJI under the Uniform (formerly referred to as the GCTA model) and LDAK models are very close (Supplementary Figs. 8–12). The accommodation of the LDAK model improves the precision of genetic correlation estimates. Second, for binary traits, we obtained similar results except that genetic covariance is on the observed scale (Supplementary Fig. 13). Next, we investigated the influence of the sample size ratio of trait 1 versus trait 2 on the performance of these methods (Supplementary Fig. 5 and Supplementary Data 2). We observed that GENJI consistently outperformed LDSC and performed comparably to GREML across all the sample size ratios. Then, we compared HDL19, which is a new summary data-based method to estimate genetic correlation with GENJI and GENJI_adjusted under different settings. We observed similar results as in a previous benchmarking work6 that HDL cannot provide reliable estimation or powerful statistical inference using an external reference panel. Additionally, HDL showed inflated type I error when the proportion of sample overlap is large6 (Supplementary Figs. 14–16 and Supplementary Data 4 and 5). As a comparison, GENJI_adjusted can still provide unbiased estimates with well-calibrated type I error, although not provided with sample overlap information. Detailed descriptions for the additional simulations are in the “Methods” section. Simulations for transethnic genetic correlation We performed simulations to assess the performance of GENJI on transethnic genetic covariance estimation. We compared GENJI with Popcorn14 and GREML. In this simulation, the first GWAS was of African ancestry, and the second was of European ancestry. Genotypes of 6992 related samples of African ancestry from UKBB were used to simulate the phenotypes of study 1. We still used samples from WTCCC (set 1; n = 7959) to simulate phenotypes of European ancestry for study 2. The SNP effects were generated by a bivariate normal distribution with genetic covariance ranging from 0 to 0.5, and heritability was fixed at 0.5. We note that when genetic covariance is 0.5, the genetic correlation is equal to 1, which means the genetic effects are perfectly correlated between the populations. Each simulation setting was repeated 100 times. Detailed simulation settings and quality control procedures are described in the “Methods” section. Popcorn needs two external reference panels to estimate LD matrices in two populations. To make fair comparisons, we also added an additional setting for Popcorn using in-sample African references, which are also the African genotype data used for simulations (denoted as “Popcorn_in_sample”). We included two kinds of hypothesis testing: the null hypothesis is either (1) genetic correlation = 0; or (2) genetic correlation = 1. GENJI outperformed both Popcorn using an external reference panel and Popcorn using an in-sample reference panel, showing more accurate estimates and improved statistical power (Fig. 2). Due to the constraint of genetic correlation being less than 1 for GREML, GREML underestimates genetic covariance for the setting of 0.5. GENJI showed unbiased estimates across all the settings (Fig. 2A). Estimates of GENJI were as powerful as those of GREML in most settings (Fig. 2B, C and Supplementary Data 7 and 8). All methods showed well-controlled type-I error for both kinds of testing (Supplementary Figs. 17 and 18). Simulation results for transethnic genetic covariance estimation. We compare the performance of Popcorn, GREML, and GENJI by point estimation, type-I error, and statistical power. We include Popcorn using both external and in-sample reference panel and denote the latter as “Popcorn_in_sample”. A Boxplots show genetic covariance estimates across n = 100 independent computational simulation replicates for each method and each true genetic covariance setting. The centre line denotes the median, box limits denote the first and third quartiles, whiskers extend to the most extreme values within 1.5 times the interquartile range, and points beyond the whiskers denote outliers. The simulated transethnic pair of GWAS datasets consisted of an African ancestry GWAS with 6992 UKBB individuals and a European ancestry GWAS with 7959 WTCCC individuals. The red dashed lines represent the true value of genetic covariance. B, C Type-I error and statistical power were calculated as the proportion of simulation replicates with P < 0.05 across n = 100 independent computational simulation replicates for each setting. P values were obtained from z-tests based on the estimated transethnic genetic correlation and its standard error. B used two-sided z-tests for the null hypothesis that the transethnic genetic correlation equals 0. C used one-sided z-tests for the null hypothesis that the transethnic genetic correlation equals 1. Source data are provided as a Source Data file. Real data applications: within-population genetic correlation We applied GREML, LDSC, and GENJI to estimate genetic correlations across 12 traits in WTCCC (n ~ 16,000) and NFBC (n ~ 5300) (“Methods”). Both studies were conducted on samples of European ancestry and shared zero samples, hence LDSC_no_intercept can also be applied. Detailed information about the GWAS data is summarized in Supplementary Data 9 and 10. Since individual-level data from both studies are available, GREML can be treated as the gold standard for the other two methods. Consistent with our simulation results, genetic covariance and genetic correlation estimates of GENJI and LDSC_no_intercept were more consistent with GREML estimates, measured by R2, than those of LDSC without a constraint on intercept (Supplementary Fig. 20 and Supplementary Data 3). Although the point estimates of GENJI and LDSC_no_intercept on the 36 trait pairs were similar, GENJI achieved higher statistical power. After Bonferroni correction (p < 0.05/36 = 1.39e-3), genetic correlations of two trait pairs were identified by both GREML and GENJI, while LDSC or LDSC with intercept fixed at zero failed to identify any trait pairs (Fig. 3 and Supplementary Data 11). The significant trait pairs are type-2 diabetes (T2D)20 versus body mass index (BMI)21 and hypertension (HT) versus high-density lipoprotein (HDL)22. The positive association between obesity (BMI) and T2D has long been observed23 and is replicated in our following applications on GWAS summary data with larger sample sizes (Fig. 4). Although low levels of HDL cholesterol are associated with increased risk of coronary artery disease24, positive correlations between HDL and HT were observed across different populations25,26. In a recent study, the authors showed that the biological mechanism of the positive association between HDL and HT is related to circulating CD34-positive cell levels27. In Supplementary Fig. 21, we also demonstrate the genetic covariance and correlation estimates of GENJI under the LDAK model. The results are very close to the results of the default setting of GENJI under the Uniform model. Genetic correlations estimated by GREML, LDSC, and GENJI across traits in WTCCC and NFBC. The heatmaps reflect the genetic correlation estimates of A GREML, B LDSC, C LDSC with intercept fixed at zero, and D GENJI. Asterisks indicate statistically significant genetic correlations based on two-sided z-tests for the null hypothesis that the genetic correlation equals zero. Significance was determined after Bonferroni correction for 36 tested trait pairs, with P < 0.05/36 = 1.39 × 10⁻³. Due to low heritability estimation, estimates of LDSC for the genetic correlation of two trait pairs are unavailable and colored as grey in (B). We summarize detailed information about each trait, including abbreviations, in Supplementary Data 9 and 10. Source data are provided as a Source Data file. Genetic correlation estimates of GENJI and LDSC for 150 trait pairs. The upper and lower heatmaps reflect the estimates of GENJI and LDSC, respectively. Asterisks indicate statistically significant genetic correlations based on two-sided z-tests for the null hypothesis that the genetic correlation equals zero. Significance was determined after Bonferroni correction for 150 tested trait pairs, with P < 0.05/150 = 3.3 × 10−4. We summarize detailed information about each trait, including abbreviations, in Supplementary Data 12 and 13. Source data are provided as a Source Data file. To illustrate the superior power of GENJI over LDSC on more GWASs with larger sample sizes, we applied GENJI and LDSC to estimate the genetic correlations between 25 common traits with publicly available GWAS summary data and 6 traits from UKBB (n ~ 270,000), which are the same set of 6 NFBC traits we used before. We fixed the intercept of LDSC at zero because there are no overlapped samples between the trait pairs. We provided GENJI with the individual-level data from UKBB. Detailed information of the 25 GWAS summary data is summarized in Supplementary Data 12. Detailed information of the 6 traits from UKBB is summarized in Supplementary Data 13. The genetic correlation estimates of GENJI and LDSC are highly consistent (R2 = 0.96). However, the standard errors from GENJI are mostly smaller than those from LDSC. Of the 150 trait pairs, significant genetic correlations of 82 and 33 trait pairs are identified by GENJI and LDSC, respectively, after Bonferroni correction (p < 0.05/150 = 3.3e-4) (Fig. 4 and Supplementary Data 14). All the trait pairs identified by LDSC are also identified by GENJI. Real data applications: transethnic genetic correlation We applied GENJI and compared its performance with Popcorn on transethnic genetic correlation estimation. We estimated the transethnic genetic correlation of 12 complex traits in the African, East Asian, and European populations. The choice of these traits was based on data availability and trait heritability (“Methods”). The GWAS data of African and European populations were from UKBB. The GWAS data of the East Asian population were from BBJ17,18 due to the small sample size of East Asian ancestry in UKBB (n ~ 1500). We only have access to summary-level GWAS data of BBJ. The details about the sample sizes and the sources of the data that we use in this section can be found in Supplementary Data 13 and 15. We first compared GENJI, Popcorn, and GREML on the transethnic genetic correlation between European and African populations in UKBB. To reduce the computational burden for GREML, we used a subset of European ancestry samples from UKBB (n = 10,000). For comparison fairness, GENJI and Popcorn also used the reduced dataset as input in this analysis. The individual-level data of African GWASs and summary-level data of European ancestry were provided to GENJI. We applied Popcorn twice in the comparison: (1) we used genotype data from the 1000 Genomes Project28 as the reference panels for African and European GWASs; (2) we used genotype data of African samples from UKBB as the reference panel for African GWASs (in-sample reference panel) and European samples from the 1000 Genomes Project as the reference panel for European GWASs. It was also for comparison fairness because GENJI took advantage of the information from the individual-level data of African ancestry. The results of the GREML, GENJI, and Popcorn are presented in Supplementary Data 16. The estimates of GENJI were substantially more consistent with GREML estimates (R2 = 0.94) compared to Popcorn (R2 = 0.41) or Popcorn_in_sample (R2 = 0.45) (Supplementary Fig. 22). The variability of the Popcorn estimates yielded relatively unreliable and improper outcomes. To be specific, the transethnic genetic correlation estimates of Popcorn or Popcorn_in_sample for four traits were outside of [−1,1], which is the range of genetic correlation. Three and two traits were unavailable for Popcorn and Popcorn_in_sample, respectively, because of negative estimates for heritability. As a comparison, two traits’ GREML estimates were outside the proper range, while no genetic correlation estimates of GENJI were out of bounds. Estimates for all traits were available for GREML and GENJI. Consistent with our simulation results, standard errors of GENJI were substantially smaller than those of Popcorn and similar to those of GREML (Supplementary Fig. 23). No traits were identified by Popcorn_in_sample with transethnic genetic correlation significantly larger than zero after Bonferroni correction (p < 0.05/24 = 2.1e-3), while both GREML and GENJI identified three significant traits: Height, BMI, and HDL. We then used the whole dataset of European ancestry samples in UKBB and GWAS summary data from BBJ to estimate the transethnic genetic correlation across African, East Asian, and European populations for the 12 traits. The genotype data provided to GENJI were also used as the reference panel for Popcorn (in-sample reference panel) (“Methods”). The point estimates from GENJI and Popcorn were similar (Fig. 5). Since the sample sizes of the GWAS summary data from BBJ were larger than those of African population GWASs (Supplementary Data 13 and 15), the standard errors for the transethnic genetic correlation between European and East Asian populations were smaller than those of the other two population pairs. The standard errors of GENJI estimates were consistently smaller than those of Popcorn. Due to the unstable estimates for heritability, the transethnic genetic correlation estimates of Popcorn for prostate cancer (PC)29 were unavailable. The genetic correlations between European and East Asian populations were significantly larger than those between European and African populations (Wilcoxon test p = 4.9e-4) and those between East Asian and African populations (Wilcoxon test p = 9.8e-4) across the 12 traits. The differences of genetic correlation between African and European populations and African and East Asian populations were not statistically significant (Wilcoxon test p = 0.42). The transethnic genetic correlation estimates of GENJI for most traits were significantly larger than zero and smaller than one across the populations (Supplementary Data 17–19), which indicates that the genetic effects for most traits are neither independent nor perfectly correlated between the populations. All traits show a genetic correlation significantly larger than zero between European and East Asian populations according to the results of GENJI. Only two traits, age at natural menopause30 and PC, did not show genetic correlations significantly larger than zero between European and African populations. A recent trans-ancestry GWAS of PC showed that the gap of R2 of the polygenic risk score (PRS) for African ancestry was substantially lower than that of East Asian and European ancestry if the PRS is trained by the GWAS, which mostly consists of European ancestry31. Transethnic genetic correlation estimates of GENJI and Popcorn for 12 traits cross different populations. To make the comparisons fair, we use an in-sample reference panel for Popcorn for the population for which GENJI takes the individual-level data as input. African, East Asian, and European are denoted as AFR, EAS, and EUR, respectively. Data are presented as point estimates +/− one standard error. Panels show estimates between A East Asian and European populations, B European and African populations, and C African and East Asian populations. Some of the point ranges are truncated due to their extraordinary values. Some of the estimates of Popcorn are not available due to low heritability estimation. We summarized the abbreviations and detailed information about the data sources in Supplementary Data 13 and 15. Source data are provided as a Source Data file. GENJI improves cross-population risk prediction To illustrate the benefits of improved transethnic genetic correlation estimation, we show that cross-population PRS is more predictive with the help of more accurate genetic correlation estimates. It has been shown that leveraging transethnic genetic correlation can substantially improve PRS performance, especially for minority groups32. Here, based on a commonly used method for computing PRS, clumping and thresholding (C + T)33–35, we highlight the advances of GENJI, which shows that the cross-population PRS improved by genetic correlation estimates of GENJI is more predictive than PRSs constructed in other ways. We constructed the PRSs of Height and BMI for the samples of African ancestry in Population Architecture Genomics and Epidemiology (PAGE): Multiethnic Cohort (MEC) (n = 3520). The genetic weights were trained from UKBB data. We used the estimates of transethnic genetic correlation between African and European populations in the previous section to improve the prediction (“Methods”). We also included PRS constructed by meta-analysis of African and European populations using METAL36 in the comparison. The predictive R2 of PRSs improved by GENJI was higher than all the other approaches (Fig. 6 and Supplementary Fig. 7) as a result of the more accurate estimates from GENJI. The R2 of the PRSs that were constructed only by GWAS of the African population were significantly the lowest. Since the transethnic genetic correlation of BMI is relatively low (GENJI corr = 0.61) and the transethnic genetic correlation of BMI estimated by Popcorn (Popcorn corr = 0.68) is relatively close to that estimated by GENJI (Fig. 5), PRSs improved by GENJI and Popcorn outperformed the PRS constructed by METAL for BMI. On the other hand, the transethnic genetic correlation of Height is relatively high (GENJI corr = 0.74). In addition, the estimate of Popcorn (Popcorn corr = 0.60) is far from that of GENJI and might underestimate the transethnic genetic correlation for Height (Fig. 5). Consequently, PRS improved by GENJI and the PRS constructed by METAL outperformed PRS improved by Popcorn for Height. The R2 of the cross-population PRS. Bar plots visualize the R2 of the PRSs for BMI and Height. The exact values of R2 are annotated in the figure. The PRSs are constructed by four different ways: (1) Original, only using the GWAS of the African population to construct the PRS; (2) METAL, using meta-analysis of GWASs of African and European populations to construct the PRS; (3) Popcorn, using estimates of Popcorn to adjust the genetic weights in PRS; (4) GENJI, using estimates of GENJI to adjust the genetic weights in PRS. Discussion Genetic correlation is a powerful metric to effectively measure the etiologic sharing of numerous phenotypes. Recently, increasing attention has been paid to the genetic architecture of non-European populations37–39. Transethnic genetic correlation can provide insights into how the genetic architecture of complex phenotypes varies between population14. Methods to estimate genetic correlation based on individual-level data or summary-level data have achieved some success. The individual-level data-based methods (e.g., GREML7) have proved to be statistically more efficient than the summary-level data-based methods6. When individual-level data are available for one trait and summary-level data are available for the other trait, to estimate genetic correlation, researchers have to transform the individual-level data to summary statistics and apply summary-statistics-based methods, leading to information loss and suboptimal estimates. Now that individual-level data are increasingly accessible, we propose a statistical framework that can estimate within-population or transethnic genetic correlation with individual-level data for one trait and summary-level data for the other trait, which can provide more accurate estimates than summary-statistics-based methods. GENJI provides statistically rigorous and computationally efficient inference for both within-population and transethnic genetic correlations and substantially outperforms summary data-based methods in our simulations. The computational complexity of GENJI increases linearly with sample size, which is more efficient than individual-level data-based methods. We also observe that the information loss from individual-level data to summary-level data is likely to be induced by the estimation of the covariance of non-genetic effects for overlapped samples, which is parameterized as the intercept of LDSC. The point estimation of LDSC with intercept fixed at zero is comparable with GREML and GENJI and is more efficient than LDSC without constraint on intercept by 30% under the circumstance of no overlapped samples. Since LDSC uses block jackknife to estimate the standard errors, the statistical power of LDSC is suboptimal, no matter whether its intercept is fixed or not. The “samples” in the regression model of GENJI are the individuals in the studies. So, if the GWAS samples are unrelated, we can directly compute the standard errors of the estimates in closed form instead of applying resampling-based methods. Applied to different datasets to estimate within-population genetic correlation, GENJI consistently identified more significant trait pairs than LDSC. The findings of GENJI have been validated by GREML on both point estimation and false discovery control. Notably, GENJI identified 148% more trait pairs than LDSC in an application to 150 trait pairs from UKBB and 25 publicly available GWAS summary datasets. While GENJI shows comparable or better accuracy and efficiency to established methods, its primary advantage lies in enabling analyses that would otherwise be infeasible due to format constraints of the input data. GENJI is particularly useful in settings where researchers aim to study the genetic architecture of phenotypes that are difficult or costly to measure. Examples include molecular or imaging biomarkers, treatment response indicators, and phenotypes that require invasive or time-intensive data collection. These traits are often available only in smaller, specialized cohorts. GENJI enables researchers to estimate the genetic correlation between such rare phenotypes and more widely available traits, such as blood pressure or BMI, by combining individual-level data with large-scale summary statistics. Discovering strong genetic correlations in this context can suggest that a simple, low-cost trait may serve as a proxy or early screening indicator for a more complex trait. This insight has the potential to inform study design, improve targeting for downstream testing, and enhance translational relevance, especially in resource-constrained settings. The ability to estimate genetic correlation between difficult-to-measure traits and more widely accessible traits has important implications for applied study design. This idea has long been central in quantitative and animal genetics, where the concept of indirect selection is used to predict gain in a focal trait based on selection on a genetically correlated, easier-to-measure proxy40. In human studies, analogous challenges arise when individual-level data are collected for phenotypes that are expensive, invasive, or technically demanding, such as imaging biomarkers, cognitive tests, or drug response profiles. If a common trait, such as BMI or blood pressure, is strongly genetically correlated with one of these complex traits, it may serve as a low-cost proxy to identify individuals for follow-up studies or clinical targeting. GENJI facilitates this type of analysis by enabling genetic correlation estimation between small-scale individual-level datasets and large-scale public GWAS summary statistics. However, as Gianola et al. (2015) discuss41, genome-wide genetic correlation should be interpreted carefully, since molecular markers may reflect pleiotropy only partially, and correlations can be shaped by LD structure, trait architecture, and population differences. Users of GENJI should consider these factors when drawing mechanistic conclusions or designing proxy-based strategies. GENJI can also yield reliable outcomes for transethnic genetic correlation estimation. Estimates of GENJI are more consistent with those of GREML than the estimates from Popcorn. In addition, the standard errors from GENJI are substantially smaller than those from Popcorn. The results show that the genetic effects of East Asian and European are more correlated than African and European or African and East Asian, which the closer genetic distance between Europeans and Asians might interpret. Furthermore, for most traits, we observe that the transethnic genetic correlations across populations are significantly different from either zero or one. These results suggest that more sophisticated methods are needed to study the genetic architectures for non-European populations. On one hand, it is inappropriate to generalize the findings in the European population to the non-European population37,38. On the other hand, a large amount of information can be borrowed by cross-population analysis. For example, elaborated modeling of GWAS from multiple populations can improve PRS prediction. Researchers have proposed several methods for cross-population PRS that leverage GWAS data from other populations to boost the performance of PRS prediction on the target population32,42–44. Genetic correlation has proved to be a powerful way to improve PRS performance32,44. PRS prediction is bound to benefit from a more accurate estimation of genetic correlation. As a showcase, we used the estimates for transethnic genetic correlation between African and European populations from GENJI and Popcorn to construct cross-population PRS for the African population. We compared the PRS improved by genetic correlation with two alternative approaches: PRS based on the target population alone and PRS built on a meta-analysis combining multiple populations, which are equivalent to the cases of genetic correlation equal to zero and one, respectively. The performance of PRS can be evaluated by predictive R2 in external testing datasets. We find that PRS improved by the GENJI outperformed other approaches, which suggests that estimation from GENJI is closer to the underlying true value. A variety of existing methods, such as GREML and MTG2, provide highly accurate estimates of genetic covariance when individual-level data are available for both traits. These methods represent the current gold standard for within-population genetic correlation estimation under full data access. However, such scenarios are increasingly uncommon in large-scale, multi-cohort studies, particularly in the trans-ethnic setting. In many practical cases, individual-level data are available for only one trait. At the same time, the other is only accessible via summary statistics from external sources such as consortia, biobanks, or meta-analyses. Moreover, privacy concerns, restricted access policies, and cohort-specific data harmonization issues often make it infeasible to analyze full individual-level data across studies or populations jointly. GENJI is specifically designed to address this hybrid-access setting, where one trait is represented by individual-level data and the other by summary-level data. This framework complements—but does not aim to replace—multi-trait methods designed for joint individual-level analysis. Instead, it extends the utility of genetic correlation estimation to a broader range of real-world applications where traditional joint modeling is not feasible. An alternative strategy, when individual-level data are available for one trait, is to first perform a GWAS and then apply summary-based methods such as LDSC or Popcorn to estimate genetic correlation. While computationally straightforward, this approach can introduce estimation error due to loss of information during the GWAS summarization step. In particular, it discards residual variance, within-sample LD, and the joint genotype–phenotype relationship that may be informative for cross-trait analysis. GENJI retains this information by modeling the individual-level data directly, which leads to greater estimation precision and robustness, especially when sample sizes are moderate or when studying trans-ethnic correlations. In such settings, where LD and allele frequency differences across populations may confound summary-based methods, GENJI leverages observed genotypes to infer shared genetic effects more accurately41. This is especially important when genetic correlation is driven by subtle or heterogeneous signals, as molecular markers may only weakly inform pleiotropic relationships. We view GENJI as a complementary tool that extends the reach of genetic correlation analysis to hybrid-access settings while retaining statistical rigor and biological interpretability. Our method has some limitations. First, to implement the weighted regression in GENJI, if the two studies share non-zero samples, we need to know who is included in the set of the shared samples. Second, the advantage of GENJI diminishes with the increase of the proportion of overlapping samples. Too many overlapping samples between the studies make the regression problem in GENJI suffer from multicollinearity and result in unstable estimates (“Methods”). We note that these two limitations are both induced by sample overlapping, which does not exist in transethnic genetic correlation applications. Third, genome-wide genetic correlations only reflect the average concordance of genetic effects across the genome and often fail to reveal the local, heterogeneous pleiotropic effects, especially when the underlying genetic basis involves multiple etiologic pathways43,45,46. Future directions include extending our method to estimate local genetic correlation jointly using individual-level and summary-level data. While genome-wide genetic correlation is a valuable tool for assessing shared genetic architecture, its interpretation requires care. A significant genetic correlation may reflect true pleiotropy, where the same genetic variants influence both traits, or it may arise from correlation due to LD between nearby causal variants. Additionally, unaccounted environmental correlations may contribute to observed associations. As noted by Gianola et al. (2015)41, marker-based estimates reflect statistical rather than purely biological relationships. GENJI, like other correlation methods, captures genome-wide average sharing and should be interpreted accordingly, especially in trans-ethnic or polygenic settings. Taken together, GENJI provides a biologically motivated and statistically principled analytical strategy to tackle etiologic sharing of complex traits within and across populations. A combination of individual-level and summary-level data is more desirable to fully utilize the information provided by the data. Moreover, it is doable in prospect owing to the increasing availability of individual-level data. We believe GENJI will have wide applications in complex traits and cross-population genetics research. Methods This research complies with all relevant ethical regulations. All UKBB participants provided written informed consent, and additional information is available at https://www.ukbiobank.ac.uk/learn-more-about-uk-biobank/governance. Access to WTCCC data was obtained through the WTCCC. Access to NFBC1966 data can be obtained through the Northern Finland Birth Cohorts research program at the University of Oulu and its NFBC Scientific Committee; access information is available at https://www.oulu.fi/en/university/faculties-and-units/faculty-medicine/northern-finland-birth-cohorts-and-arctic-biobank/nfbc-aineistopyynto. Access to all restricted datasets was subject to the original ethical approvals, participant consent and data-use agreements. Statistical model The statistical frameworks of GENJI for the estimation of within-population genetic correlation and transethnic genetic correlation are nearly the same. Assume that the two studies have sample sizes n1 and n2, respectively. Standardized trait values ϕ1 and ϕ2 follow the linear models below: where X and Y are n1×m and n2×m standardized genotype matrices; m is the number of shared SNPs between the two studies; ε and δ are the noise terms; and β and γ denote the genetic effects for ϕ1 and ϕ2. The genetic covariance ρg is defined as the covariance of the (population-specific) allele-variance-normalized random SNP effect sizes. So, the combined random vector of β and γ follows a multivariate normal distribution given by: where Im is the identity matrix of size m. In the estimation of transethnic genetic correlation, this assumption for the distribution of SNP effect sizes corresponds to genetic-impact correlation in the Popcorn paper14. Since trait values ϕ1 and ϕ2 are standardized, we have Varεi1=1−h12 and Varδi2=1−h22 for 1≤i1≤n1 and 1≤i2≤n2. The covariance of the genetic factors of the two traits is equal to ρg. In fact, for an individual with normalized genotype vector X0, CovX0Tβ,X0Tγ=EβTX0X0Tγ=ρg. Here, we assume that we have individual-level data for the first study and only summary-level data for the second study. With this assumption, we treat X it as a fixed genotype matrix and Y as an unknown random design matrix. We further assume that each row of Y is independently drawn from a distribution with covariance matrix R, the m×m LD matrix for study 2, i.e., CovYi2⋅=R. The LD matrix can be estimated from an external reference panel (e.g., 1000 Genomes Project28) or the genotype of study 1 if the two studies are conducted on the same population. For the phenotype vectors ϕ1 and ϕ2, we assume ϕ1 is known while ϕ2 is unknown. For the GWAS summary data of study 2, we only observe z-scores, and we can approximate the z-score of the SNP j by z2j≈Y⋅jTϕ2/n2. We use z2 to denote the vector of length m which contains all the z-scores in study 2. In practice, for within-population genetic correlation estimation, two different studies may share a subset of samples. Without loss of generality, we assume the first ns samples in each study are shared (ns≤n1 and ns≤n2). The non-genetic effects of the shared samples for the two studies are correlated: We denote the genotype of the overlapped samples as Xs, which consists of the first ns rows of X. We note that the genotype matrix Xs is known to us because Xs is a submatrix of X. We also note that the phenotypes for study 2 of the overlapped samples are unknown, although we know their genotypes. Parameter estimation The primary parameter to be estimated is genetic covariance ρg. To fully utilize the individual-level information, we relate the expectation of XiTz2ϕ1i to the parameter of interest, where Xi and ϕ1i are the genotype vector and phenotype value of the i th sample in study 1, respectively. It can be shown that where 1i≤ns is the indicator function for the overlapped samples. We provide detailed derivations of (2) in the Supplementary Note. We estimate genetic covariance by regressing XiTz2ϕ1i against XiTXsTXs+n2−nsRXi and then multiplying the resulting slope by mn2. We note that for transethnic genetic covariance, since there is no overlapped sample between the studies, (2) reduces to and we constrain the intercept of the regression to be zero. To obtain the optimal estimator of the regression, we apply weighted regression. The weights are given by the reciprocal of the variance of XiTz2ϕ1i: The weights depend on h12, h22, ρg and ρe that are unknown. The parameters for heritability h12 and h22 can be estimated by GREML and LDSC, respectively. For ρg and ρe, we estimate them in two steps. In the first step, we set ρg=ρe=0 to calculate the weights. Then, we use the estimates of ρg and ρe in the first step to calculate the weights for the second step. The result of genetic covariance estimation is the estimate from the second regression. The performance of (2) is deteriorated by the multicollinearity between XiTXsTXs+n2−nsRXi and XiTXi1i≤ns under large sample overlap. In that case, GENJI_adjusted could be applied as an alternative, which is introduced in detail in the next section. The detailed derivations to calculate the variance of XiTz2ϕ1i is provided in the Supplementary Note. The estimate for genetic covariance is on the observed scale if one or both studies are case-control study. The observed scale genetic covariance is ρobs=ρgϕτ1ϕτ2P11−P1P21−P2/K11−K1K21−K2 when both studies are case-control studies, where ρg is the liability scale genetic covariance, ϕ is the standard normal density, τ1 and τ2, P1 and P2, and K1 and K2 are the liability threshold, sample prevalence, and population prevalence of study 1 and study 2, respectively. If one study is case-control study, the observed scale genetic covariance is given by ρobs=ρgϕτiPi1−Pi/Ki1−Ki, i=1, 2. The detailed derivations are similar to those presented in the Supplementary Note of Zhang et al.43. Since the samples are independent of each other, we can directly use the standard errors of the coefficients estimated by the weighted regression. Z-test is applied to determine the statistical significance. For transethnic genetic correlation, we also use the standard errors from weighted regression to test whether the genetic correlation is significantly below 1 with a one-tailed test. We can extend the above method to include the covariates of study 1: Here A is an n1×p matrix of covariates and b is a vector of length p for fixed effects. Let p0 be the rank of A and U be an n1×n1−p0 matrix that consists of the orthonormal bases of the orthogonal complement space of A. U satisfies UTU=In1−p0 and UUT=In1−A0A0TA0−1A0T, where A0 is an n1×p0 full rank submatrix of A. We multiply by UT on both sides of (3) which yields: We can simply use standardized UTϕ1 as the corrected phenotypes and column standardized UTX as the corrected genotypes and then input the corrected phenotypes and genotypes to the regression introduced above in (2). Correction for unknown overlapped samples When the samples shared between the studies are unknown, regression in (2) cannot be applied. To quantify the bias in the estimation of ρg caused by ρe, we treat the model as if there were no overlapped samples and obtain the misspecified estimator ρ~g. The bias is given by where wi is the regression weight for individual i which is usually the reciprocal of variance. We apply a regression model similar to LDSC to estimate nsρe. We denote Xj as the j th column of the genotype matrix X and Rj as the LD of SNP j. We have and We use a similar two-step weighted regression procedure introduced previously to calculate the regression weights, which involves unknown parameters. Although we can obtain the estimation ρg from this regression, the estimator ρg derived from (2) shows more statistical efficiency and convenience in inference. The detailed derivations are given in the Supplementary Note. Accommodation for a flexible heritability model Estimating genetic correlation between traits requires an underlying model of how genetic variants contribute to heritability and co-heritability. Like many early methods, such as GREML and LDSC, GENJI implicitly assumes that each SNP contributes equally to heritability, known as the uniform or GCTA model. While convenient, this assumption ignores variation in SNP effect sizes due to factors such as local LD, minor allele frequency (MAF), and functional annotations36. The SumHer framework directly addresses this limitation by introducing a flexible class of non-uniform heritability models, particularly the LDAK-Thin Model, which downweighs SNPs in high-LD regions and accounts for MAF-dependent architecture47. SumHer demonstrated that incorporating such biologically informed priors can substantially improve the accuracy and precision of SNP-heritability and genetic correlation estimates, particularly in the presence of polygenicity and LD heterogeneity. These findings have motivated a broader shift toward modeling heterogeneous SNP contributions in genetic association studies. To support this advancement, our method GENJI has been extended to accommodate arbitrary SNP-specific heritability weights, allowing users to incorporate models such as LDAK-Thin directly into the estimation of genetic correlation. This generalization is implemented by adjusting the regression framework to weight SNP contributions according to user-specified values. As a result, GENJI can now be applied under both traditional uniform and more realistic non-uniform heritability models. The details of the accommodation are provided in the Supplementary Note. Simulation settings for within-population genetic correlation In this section, we introduce the procedure that we use to simulate data to compare the performance of GENJI, LDSC, and GREML. Phenotypes in our simulations were generated based on unimputed genotype data from the WTCCC for within-population genetic correlation estimation. Samples were randomly divided into two equal subgroups, and each subgroup had 7959 individuals. We denote them as set 1 and set 2, respectively. We randomly sampled individuals from set 1 and set 2 and created set 3, set 4, and set 5, whose sample sizes were all equal to 7959, such that they had 10, 25, and 50% overlapped samples with set 1, respectively. Genotype data of 503 individuals with European ancestry from the 1000 Genomes Project phase III28 were used as the LD reference in our simulations. In other words, we were using an external reference panel to estimate LD scores. The comparison between LDSC and GENJI is fair because we assume that we only have access to the individual-level data of only one GWAS. Since the genotype data of the second GWAS were unknown, we used external reference panel for the second GWAS, for both LDSC and GENJI. GREML is supposed to outperform LDSC and GENJI because GREML has more information from individual-level data. SNPs with ambiguous alleles or minor allele frequencies (MAF) less than 5% were removed. 253,196 SNPs presented in both WTCCC and the 1000 Genomes Project remained. The effect sizes of SNPs were generated by a multivariate normal distribution, and we applied Genome-wide Complex Trait Analysis (GCTA)9 to simulate the phenotypes. We used PLINK35 to run GWAS and obtain summary statistics of the two simulated phenotypes. We repeated each simulation setting 100 times. Detailed simulation settings are summarized below. We fixed the heritability of the phenotypes as 0.5, and the values of genetic covariance were set from 0 to 0.25. The first GWAS was simulated on the individuals in set 1, and the second GWAS was simulated for set 2, set 3, set 4, and set 5 in four scenarios with different sample overlapping. The covariance of non-genetic effects on overlapped samples was set to be 0.2, i.e., ρe=0.2. When there was no overlapped sample between the two studies, we also included LDSC with the intercept fixed at zero in our comparison. GCTA was applied for GREML implementation. To check the performance of GENJI under both Uniform and LDAK models, we also conduct another set of experiments in which the phenotypes were simulated under the LDAK model. In the experiments, we compared four methods: (i) GENJI with uniform (GCTA) weights, (ii) GENJI with LDAK-Thin weights, (iii) SumHer under the LDAK-Thin Model, and (iv) REML (GCTA). The results are in Supplementary Figs. 8–12. We also simulated binary traits using the liability model for the no-sample-overlap scenario to investigate the performance of the methods for a case-control study. The liability scale genetic covariance was still set from 0 to 0.25. We set the threshold for the liability to be the 80% quantile of the standard normal distribution, such that the prevalence of the binary traits was 0.2. The related results are in Supplementary Fig. 13. To study the influence of sample size ratio of trait 1 versus trait 2 on the performance of the methods, we randomly assigned the 15,918 samples from the WTCCC to study 1 and study 2 in three difference ways: (1) study 1/study 2/overlap: 11370/5685/1137; (2) study 1/study 2/overlap: 7959/7959/796; (3) study 1/study 2/overlap: 5489/10978/549. Namely, the sample ratio of study 1 versus study 2 is 1:2, 1:1, or 2:1. The number of overlapped samples was fixed to be 10% of study 1. We used real genotype data and simulated phenotypes. The heritability was set to be 0.5, and the genetic covariance was 0.1. The covariance of the random error of the overlapped samples was 0.2. The simulations were repeated 100 times. We compared the point estimates and statistical power of LDSC, GREML, and GENJI on the dataset. The related results are in Supplementary Fig. 5 and Supplementary Data 2. In another set of simulations, we compared GENJI with HDL19, a new method to estimate genetic correlation. We obtained the real genotype data from WTCCC and the simulated phenotype data through the same procedure as the simulations in Fig. 1. We changed the value of true genetic covariance and the proportion of overlapped samples in the comparison. We provided HDL with two different reference panels, which were the genotypes from the 1000 Genomes Project and UKBB, because the authors of HDL suggested that HDL performs better with a reference panel of a large sample size19. The corresponding results were referred as HDL for the 1000 Genomes Project and HDL_ref_ukb for UKBB in our simulations, respectively. We also included GENJI_adjusted in the comparison to evaluate the performance of GENJI when the information of overlapped samples was unknown. We note that GENJI_adjusted does not need any information from overlapped samples between the studies as input. The results of point estimates are in Supplementary Figs. 14–16, and the results of statistical power are in Supplementary Data 4 and 5. Simulation settings for transethnic genetic correlation In this section, we used a similar procedure introduced above to compare the performance of GENJI, Popcorn, and GREML on transethnic genetic correlation estimation. The unimputed European genotype data from the WTCCC (set 1), and genotype data of genetically unrelated UKBB African ancestry samples were used to generate phenotypes, with sample sizes of 7959 and 6992, respectively. We provided GENJI with individual-level data of African ancestry from UKBB and summary-level data generated from WTCCC European samples. Individuals of European and African ancestry from the 1000 Genomes Project phase III were used as the reference for Popcorn. For fair comparison, we also added an additional setting for Popcorn, where the reference panel was the genotype data used to generate the phenotype of the African traits (in-sample reference) because GENJI uses the individual-level data of the GWASs of African ancestry to estimate genetic correlation. SNPs with ambiguous alleles or MAFs less than 5% were removed. 159,093 shared SNPs remained in the genotype data of WTCCC, UKBB, and the 1000 Genomes Project. Genetic effect sizes were generated by a multivariate normal distribution. Similarly, phenotypes and summary statistics were generated by GCTA and PLINK, respectively. The heritability of the phenotypes was still set to be 0.5, and the values of genetic covariance were set to range from 0 to 0.5. Each setting was repeated 100 times. We used the genetic impact correlation for Popcorn. Since it was also of great interest to investigate whether the genetic effects of a specific trait between two populations were perfectly correlated (transethnic genetic correlation = 1), besides the traditional null that genetic correlation is equal to zero, we also tested for H0:corr=1 with alternative H1=corr<1. Computational efficiency We compared the computational efficiency of different genetic correlation estimation methods, including GREML, LDSC, Popcorn, and GENJI. We investigated the relation between sample size and the computational efficiency of these methods. We followed a procedure similar to our simulation introduced above, which used real genotype data and simulated phenotype data. For trait 1 (which is the trait of individual-level data for GENJI), we randomly sampled 1748, 3496, 6992, 50,000, 75,000, and 270,000 individuals from UKBB and generated the phenotypes. For trait 2 (which is the trait of summary-statistic data for GENJI), we fixed the sample size and generated the phenotype based on WTCCC European samples (n = 7959). Though Popcorn was proposed to estimate transethnic genetic correlation, in this comparison, for the sake of fairness, we used it on the same datasets as other methods for within-population genetic correlation estimation. The number of SNPs is 253,196 across the datasets with different sample sizes in this comparison. Genotype data from the 1000 Genomes Project were used as the reference panel for GENJI and summary-level data-based methods. We set a 24-h time limit for all methods. Only GREML in the setting of 270,000 individuals cannot fit the model within the time constraint. We run all these methods on Cascade Lake CPU. The detailed results can be found in Supplementary Fig. 19 and Supplementary Data 6. Within-population genetic correlation estimation for European ancestry GWASs We used GREML, LDSC, and GENJI to estimate genetic correlations across the 12 phenotypes in the WTCCC and NFBC datasets. We downloaded genotype and phenotype data of NFBC from dbGaP48 (accession: phs000276.v2.p1). Since only a small proportion of SNPs were shared by these two datasets, we imputed WTCCC using the Michigan Imputation server49, which resulted in 227,383 shared SNPs with NFBC with MAF larger than 0.05. Sex, age, and the top 4 principal components were included as covariates to implement GREML, GENJI, and perform GWAS. Sex information was obtained from the original studies and data providers. To run GENJI and LDSC, we generated summary data for both cohorts using PLINK. We provided GENJI with individual-level data from NFBC and summary data from WTCCC. Genotypes of European ancestry in the 1000 Genomes Project were still used as a reference panel for GENJI and LDSC. The details of the phenotypes and sample sizes are summarized in Supplementary Data 9 and 10. We then applied GENJI on GWASs of European ancestry with much larger sample sizes. We estimated the genetic correlation between 25 common traits with publicly available GWAS summary data and 6 common traits from UKBB. We included the genotypes of 276,731 genetically unrelated samples of European ancestry and 306,579 Axiom Array (unimputed) SNPs with MAF larger than 0.05 from UKBB in the analysis. All the trait pairs of WTCCC and NFBC were also included between UKBB traits and the 25 GWASs summary data. Due to the unavailability of the individual-level data of 25 GWASs, GREML was not implemented. Sex, age, and the top 4 principal components of UKBB samples were included as covariates to implement GENJI and perform GWAS. We transformed GWASs of UKBB traits to summary data by PLINK as the input to LDSC. Since there was no overlapped sample between UKBB and the 25 GWASs, we fixed the intercept of LDSC at zero. The details about the phenotypes and sample sizes of the UKBB traits are given in Supplementary Data 13. The details about the sample sizes and the sources of the 25 GWASs are given in Supplementary Data 12. Transethnic genetic correlation estimation for European, African, and East Asian ancestry GWASs We used individual-level data of samples with European ancestry and African ancestry in UKBB and summary-level data from BBJ17,18 to estimate the transethnic genetic correlation of 12 traits across the three populations. Our choice of these traits was based on the availability of the data and the heritability of the traits. All the chosen traits have nominally significant heritability (p < 0.05) across all the populations (Supplementary Data 20). We used GWAS summary data from BBJ because they are publicly available and have larger sample sizes than samples of East Asian ancestry in UKBB. Similarly, we only used unimputed SNPs from UKBB. We used the genetic impact correlation for Popcorn. Detailed information about the datasets used in this section is summarized in Supplementary Data 13 and 15. We first compared the performance of GENJI with Popcorn by the consistency with GREML on transethnic genetic correlation estimation between the European ancestry and African ancestry populations. Since it is an expensive computational burden to implement GREML on Biobank-scale data (n = 276,731 for samples of European ancestry), we randomly extracted 10,000 individuals of European ancestry and performed the three methods on the GWASs based on this subset. Sex, age, and the top 4 principal components were included as covariates. We provided GENJI with individual-level data of African ancestry and summary-level data generated from European ancestry. For the African population, Genotype data of African ancestry from the 1000 Genome Project (external reference panel) and UKBB (in-sample reference panel) were used as a reference panel to estimate LD in two implementations of Popcorn, respectively, for the sake of comparison fairness between GENJI and Popcorn. For the European population, genotype data of European ancestry in the 1000 Genomes Project (external reference panel) was used as a reference panel for both GENJI and Popcorn. Then, GENJI and Popcorn were further compared on the estimation of transethnic genetic correlation among European, African, and East Asian ancestry GWASs with larger sample sizes. For the results between African and European populations, genetic correlations were estimated using the whole dataset of UKBB. Similarly, we provided GENJI with individual-level data of African ancestry and summary-level data generated from European ancestry. Individuals of African ancestry from UKBB (in-sample reference panel) and of European ancestry from the 1000 Genomes Project were used as the reference for Popcorn. GENJI also used the European ancestry genotype from the 1000 Genomes Project as the reference panel. The GWAS summary data for the East Asian population were downloaded from the BBJ website (URLs). For the genetic correlation between African and East Asian populations, and European and East Asian populations, we provided GENJI with individual-level data from UKBB (African or European population) and summary data for the East Asian population. Individuals of African or European ancestry in UKBB (in-sample reference panel), and of East Asian ancestry from the 1000 Genomes Project were used as the reference for Popcorn, while GENJI used the genotype of East Asian ancestry from the 1000 Genomes Project as its reference panel. Using transethnic genetic correlation to improve cross-population PRS We used the transethnic genetic correlation estimated by GENJI and Popcorn to improve PRS prediction for Height and BMI in the African population. We compared the PRS adjusted by genetic correlation with the unadjusted PRS and the PRS based on the meta-analysis combining the GWASs of European and African populations using METAL50. We used the GWAS data of African and European populations for Height and BMI from UKBB as the training data. The transethnic genetic correlations for Height and BMI were directly applied from the previous section. Samples of African ancestry from PAGE: MEC were included in the testing set. We downloaded the individual-level phenotype and genotype data from dbGaP48 (accession: phs000220.v2.p2). We imputed genotype from PAGE using the Michigan Imputation server49. We restricted the analysis to autosomal variants with genotype missing rate per marker <0.05, missing rate per individual <0.1, Hardy–Weinberg Equilibrium p-value > 1e-9, and MAF > 0.05. After quality control, we ended with 3520 samples of African ancestry with both genotype and phenotype data available. 1,334,601 SNPs remained in the imputed genotype data of PAGE. For the unadjusted PRS, we first clumped the SNPs by PLINK35. We set the significance threshold for index SNPs as 1, LD threshold for clumping as 0.1, and physical distance threshold for clumping as 250 kb. We then used the –score function in PLINK to calculate the unadjusted PRS. The PRS from meta-analysis was constructed similarly. We used the inverse variance-based analytical strategy given by Table 1 of METAL36 and generated genetic weights to compute PRS. For PRS adjusted by genetic correlation, we applied similar methods in Multi-Trait Analysis of GWAS51 and Multi-Ancestry Meta-Analysis52 to adjust the effect sizes of SNPs. The effect size vector for the first study (population) is β, which is used to construct the PRS for the target population. The effect size vector for the second study (population) is γ, which is the auxiliary data. We only used the SNPs after clumping, and hence, the SNPs were considered to be independent (no LD). So, without LD, conditioned on the true effect size βi, the marginal effect of normalized genotype for SNP i from the GWAS of the target population is subject to normal distribution: β^i∣βi~Nβi,1/n1. The conditional distribution of γ^i given βi is: Finally, the conditional expectation of βi given β^i and γ^i is: We plugged the estimates from GENJI and Popcorn into (4) and used the results given by (4) as the genetic weight improved by transethnic genetic correlation to compute the new PRS. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Data 1-20 Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements This study uses data generated by the Wellcome Trust Case-Control Consortium. We conducted the research using the UKBB resource under approved data requests (refs: 29900). This study makes use of summary statistics from many GWAS consortia. We thank the investigators in these GWAS consortia for generously sharing their data. The NFBC1966 Study was conducted and supported by the National Heart, Lung, and Blood Institute (NHLBI) in collaboration with the Broad Institute, UCLA, University of Oulu, and the National Institute for Health and Welfare in Finland. This manuscript was not prepared in collaboration with investigators of the NFBC1966 Study and does not necessarily reflect the opinions or views of the NFBC1966 Study Investigators, Broad Institute, UCLA, University of Oulu, National Institute for Health and Welfare in Finland and the NHLBI. Funding support for the PAGE Multiethnic Cohort study was provided through the National Cancer Institute (R37CA54281, R01CA6364, P01CA33619, U01CA136792, and U01CA98758) and the National Human Genome Research Institute (U01HG004802). Assistance with phenotype harmonization, SNP selection, data cleaning, meta-analyses, data management and dissemination, and general study coordination, was provided by the PAGE Coordinating Center (U01HG004801-01). GWAS data of east Asian ancestry were downloaded from BBJ website, which were all publicly available. Author contributions Y.Z. conceived and designed the study. Y.Z. developed the statistical framework. Y.Z., W.J., and J.S. implemented the software. Y.Z., W.J., J.S., and Y.C. performed simulations. Y.Z., W.J., J.S., Y.C., and Y.Y. performed real data analysis. Y.Y. processed UKBB data. W.J. downloaded and curated NFBC and PAGE data. Y.Y., Q.L., and H.Z. advised on the biological interpretation of the real data findings. W.J., Q.L., and H.Z. advised on statistical and genetics issues. Y.Z., W.J., and Y.C. wrote the manuscript. All authors contributed to manuscript editing and approved the manuscript. Peer review Peer review information Nature Communications thanks Doug Speed and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding H.Z. discloses support for the research of this work from the National Institutes of Health (NIH) [grant number 1R01HG012735]. Data availability The raw individual-level genotype and phenotype data from UK Biobank, the Wellcome Trust Case-Control Consortium, NFBC1966 and PAGE/MEC are available under restricted access and cannot be shared by the authors because they contain participant-level genetic and phenotypic information protected by participant consent, ethical approvals and data-use agreements. Access to these datasets can be requested from the corresponding data repositories or data access committees. Access to UK Biobank data can be obtained through UK Biobank; this study used UK Biobank data under approved application number 29900 [https://www.ukbiobank.ac.uk/]. Access to WTCCC data can be obtained through the Wellcome Trust Case-Control Consortium [https://www.wtccc.org.uk/]. Access to NFBC1966 data can be obtained through dbGaP under accession code phs000276.v2.p1 [https://dbgap.ncbi.nlm.nih.gov/beta/study/phs000276.v2.p1/#study]. Access to PAGE/MEC data can be obtained through dbGaP under accession code phs000220.v2.p2 [https://dbgap.ncbi.nlm.nih.gov/beta/study/phs000220.v2.p2/#study]. The expected timeframe for responses to access requests and the duration of access after approval are determined by the respective repositories. The BBJ GWAS summary-level statistics data used in this study are available from BioBank Japan and the NBDC Human Database under accession codes hum0014-v36 [https://humandbs.dbcls.jp/en/hum0014-v36] and hum0197-v23 [https://humandbs.dbcls.jp/en/hum0197-v23]. The publicly available GWAS summary-level statistics data used for the 25-trait within-population analysis are listed in Supplementary Data 12. The 1000 Genomes Project genotype data used as reference panels in this study are available from the 1000 Genomes Project [https://www.internationalgenome.org/data/]. Source data are provided with this paper. Code availability The GENJI software is publicly available via GitHub at https://github.com/YiliangTracyZhang/GENJI and via Zenodo https://doi.org/10.5281/zenodo.2096912653. External software used in this study is publicly available from the following sources: GCTA/GREML (https://cnsgenomics.com/software/gcta/#GREML), LDSC (https://github.com/bulik/ldsc), Popcorn (https://github.com/brielin/popcorn), PLINK (https://zzz.bwh.harvard.edu/plink/profile.shtml), the Michigan Imputation Server (https://imputationserver.sph.umich.edu/index.html#!) and METAL (https://genome.sph.umich.edu/wiki/METAL_Documentation). Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yiliang Zhang, Wei Jiang, Jiangnan Shen. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-76693-z. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Data 1-20 Reporting Summary Transparent Peer Review file Source Data Data Availability Statement The raw individual-level genotype and phenotype data from UK Biobank, the Wellcome Trust Case-Control Consortium, NFBC1966 and PAGE/MEC are available under restricted access and cannot be shared by the authors because they contain participant-level genetic and phenotypic information protected by participant consent, ethical approvals and data-use agreements. Access to these datasets can be requested from the corresponding data repositories or data access committees. Access to UK Biobank data can be obtained through UK Biobank; this study used UK Biobank data under approved application number 29900 [https://www.ukbiobank.ac.uk/]. Access to WTCCC data can be obtained through the Wellcome Trust Case-Control Consortium [https://www.wtccc.org.uk/]. Access to NFBC1966 data can be obtained through dbGaP under accession code phs000276.v2.p1 [https://dbgap.ncbi.nlm.nih.gov/beta/study/phs000276.v2.p1/#study]. Access to PAGE/MEC data can be obtained through dbGaP under accession code phs000220.v2.p2 [https://dbgap.ncbi.nlm.nih.gov/beta/study/phs000220.v2.p2/#study]. The expected timeframe for responses to access requests and the duration of access after approval are determined by the respective repositories. The BBJ GWAS summary-level statistics data used in this study are available from BioBank Japan and the NBDC Human Database under accession codes hum0014-v36 [https://humandbs.dbcls.jp/en/hum0014-v36] and hum0197-v23 [https://humandbs.dbcls.jp/en/hum0197-v23]. The publicly available GWAS summary-level statistics data used for the 25-trait within-population analysis are listed in Supplementary Data 12. The 1000 Genomes Project genotype data used as reference panels in this study are available from the 1000 Genomes Project [https://www.internationalgenome.org/data/]. Source data are provided with this paper. The GENJI software is publicly available via GitHub at https://github.com/YiliangTracyZhang/GENJI and via Zenodo https://doi.org/10.5281/zenodo.2096912653. External software used in this study is publicly available from the following sources: GCTA/GREML (https://cnsgenomics.com/software/gcta/#GREML), LDSC (https://github.com/bulik/ldsc), Popcorn (https://github.com/brielin/popcorn), PLINK (https://zzz.bwh.harvard.edu/plink/profile.shtml), the Michigan Imputation Server (https://imputationserver.sph.umich.edu/index.html#!) and METAL (https://genome.sph.umich.edu/wiki/METAL_Documentation).