Abstract Biological mechanisms underlying multimorbidity remain elusive. To dissect the polygenic heterogeneity of multimorbidity in twelve complex traits across populations, we leveraged biobank resources of genome-wide association studies (GWAS) for 232,987 East Asian individuals (the 1st and 2nd cohorts of BioBank Japan) and 751,051 European individuals (UK Biobank and FinnGen). Cross-trait analyses of respiratory and cardiometabolic diseases, rheumatoid arthritis, and smoking identified negative genetic correlations between respiratory and cardiometabolic diseases in East Asian individuals, opposite from the positive associations in European individuals. Associating genome-wide polygenic risk scores (PRS) with 325 blood metabolome and 2917 proteome biomarkers supported the negative cross-trait genetic correlations in East Asian individuals. Bayesian pathway PRS analysis revealed a negative association between asthma and dyslipidemia in a gene set of peroxisome proliferator-activated receptors. The pathway suggested heterogeneity of cell type specificity in the enrichment analysis of the lung single-cell RNA-sequencing dataset. Our study highlights the heterogeneous pleiotropy of immunometabolic dysfunction in multimorbidity. Subject terms: Genome-wide association studies, Type 2 diabetes, Dyslipidaemias, Chronic obstructive pulmonary disease, Asthma __________________________________________________________________ Here, the authors perform cross-trait analyses of respiratory and cardiometabolic diseases, rheumatoid arthritis, and smoking and identify negative genetic correlations between respiratory and cardiometabolic diseases in individuals of East Asian ancestry, in opposition from the positive association in European ancestry individuals. Introduction Multimorbidity, two or more coexisting diseases in an individual, burdens individuals and societies globally^[64]1,[65]2. People with multimorbidity show impaired physical functions, frequent hospitalization, and high mortality^[66]3–[67]5. Multimorbidity consequently increases healthcare costs in individuals, for instance, those with asthma and chronic obstructive pulmonary disease (COPD)^[68]6,[69]7. Unveiling the biological mechanism of multimorbidity can provide a stepping stone for personalized medicine and eventually contribute to decreasing the disease-associated burden in terms of healthcare and socio-economy. However, the complex structure of diseases and their interactions prevented the comprehensive understanding of multimorbidity. As a population-based study shows that the combination of traits comprising multimorbidity differs between populations^[70]8, we hypothesized that genetic analysis could disentangle the biological and epidemiological complexity of multimorbidity. Genome-wide association studies (GWAS) for decades have identified genetic risks of complex diseases^[71]9,[72]10. Recent studies generally identify the positive genetic correlations among diseases in the same category^[73]11–[74]13, the classification criteria for which are based on increasing knowledge about organ functions and diagnostic tests^[75]14. On the other hand, diseases in different categories present complex genetic correlations. A preceding study dissected the genome into regions with high linkage disequilibrium (LD) and described the negative local genetic correlations in the major histocompatibility complex (MHC) regions among multicategorical traits, such as that between asthma and blood triglyceride levels^[76]15. The genetic heterogeneity of multimorbidity may exist among different populations but remains unclear because of the limited GWAS datasets in non-European (non-EUR) populations^[77]16. Therefore, assessing phenotypic correlations among diverse populations is effective in seeking clues for research investigating heterogeneous genetic correlations. In respiratory and cardiometabolic diseases, some studies reported the phenotypic correlations of lipid metabolism-related traits with asthma and COPD^[78]17–[79]20. Obesity occurs less frequently with severe asthma in East Asian (EAS) population than in EUR^[80]21. Furthermore, individuals with COPD are more underweight than healthy individuals in EAS^[81]22,[82]23 but not in EUR^[83]24. In contrast, a prospective observational study for EAS population showed that the prevalence of dyslipidemia is associated positively with the severity of asthma and the frequency of asthma exacerbation^[84]25. Comparing the genetic correlations between populations can offer novel insights into the biological mechanism underlying the heterogeneous phenotypic correlations. As respiratory, autoimmune, and cardiometabolic diseases are related to the immune system and inflammation^[85]26–[86]28, the dissimilar phenotypic correlations may be derived from the immune response network associated with lipid metabolism. The development of GWAS downstream analysis will contribute to the precise understanding of the genetic heterogeneity of multimorbidity. A recent pathway polygenic risk scores (PRS) analysis enables researchers to analyze the direction of functional associations between traits^[87]29. Pathway PRS accounts for genomic substructure and reflects disease heterogeneity. Instead of aggregating the estimated effects of risk alleles across the entire genome (genome-wide PRS), pathway PRS aggregates risk alleles per pathway. Although other post-GWAS analyses do not account for cross-trait associations per function, pathway PRS provides a detailed insight into the genetics underlying the common and heterogeneous dysfunctions in multimorbidity. A recent development of Bayesian PRS in genome-wide PRS^[88]30–[89]32 may improve the predictive performance of genetic liability in pathway PRS. Here, we performed GWAS for twelve complex traits relevant to the immune system and inflammation, namely respiratory and cardiometabolic diseases, rheumatoid arthritis (RA), and smoking, in EAS (n = 232,987) and EUR (n = 408,552) populations. For the multi-population GWAS, we leveraged data obtained from BioBank Japan (BBJ) and UK Biobank (UKB)^[90]33,[91]34. We merged summary statistics from FinnGen^[92]35 (n = 342,499) in EUR population using a standard fixed-effect meta-analysis and compared the global and local genetic correlations in EAS and EUR populations. We constructed genome-wide PRS to assess the genetic correlation in the individuals with and without multimorbidity. Utilizing the blood metabolome and proteome datasets of the EAS (BBJ1) and EUR individuals (UKB), we assessed the associations of respiratory diseases with circulating lipid biomarkers. We then constructed Bayesian pathway PRS using PRS-CSx^[93]32 to identify pathways with cross-trait associations in the immune system and lipid metabolism. Finally, we applied scDRS^[94]36 to investigate phenotype-relevant cells based on the human lung single-cell RNA-sequencing (scRNA-seq). Results Characteristics of GWAS meta-analysis obtained from three biobank resources An overview of this study is shown in Fig. [95]1. To investigate the heterogeneity of genetic correlations in individuals with and without multimorbidity related to the immune system and inflammation, we leveraged the biobank resources of BBJ, UKB, and FinnGen. BBJ collected about 200,000 participants for its first cohort (BBJ1) and 67,000 participants for its second cohort (BBJ2)^[96]33. This study enrolled individuals with asthma, COPD, interstitial lung disease (ILD), RA, smoking, obesity, dyslipidemia, type 2 diabetes (T2D), hypertension, coronary artery disease (CAD), heart failure (HF), or stroke. Cases were individuals with the target phenotypes, and controls were those without target or related phenotypes (Supplementary Tables [97]1 and [98]2). Briefly, we analyzed the samples of EAS (BBJ1: 801–49,217 cases; and BBJ2: 962–17,342 cases) and EUR (UKB: 2199–126,436 cases; and FinnGen: 2922–98,683 cases). We conducted GWAS for the BBJ1, BBJ2, and UKB individuals and obtained FinnGen GWAS summary statistics. We then meta-analyzed the GWAS summary statistics for each population and phenotype using an inverse-variance-weighted fixed-effect method implemented in RE2C^[99]37. Fig. 1. The study overview. [100]Fig. 1 [101]Open in a new tab We performed a GWAS meta-analysis on twelve complex traits examining 232,987 East Asian individuals (EAS) from BioBank Japan (BBJ) and 751,051 European individuals (EUR) from UK Biobank and FinnGen. We estimated the heritability and genetic correlations among the complex traits and found significant negative genetic correlations between respiratory and cardiometabolic diseases in BBJ (bottom left corner). Association analyses for genome-wide polygenic risk scores (PRS) and nuclear magnetic resonance (NMR) metabolite and Olink protein biomarkers showed the negative associations between regression coefficients for dyslipidemia and respiratory diseases (bottom left). Cross-trait pathway association analysis using Bayesian pathway PRS detected five pathways with negative risk associations, the functions of which regulate lipid metabolism (bottom right). Further pathway enrichment analysis of cell types demonstrated the enrichment of the lipid pathway in T cells of asthma (bottom right corner). Heterogeneity of cross-trait genetic correlations across populations We evaluated the genetic multimorbidity among respiratory and cardiometabolic diseases with multifaced approaches. We applied LDSC^[102]38 to estimate the heritability of the twelve phenotypes and analyze the global genetic correlations for each population (Fig. [103]2 and Supplementary Data [104]1). The direction of phenotypic and genetic correlations (r[g]) was concordant in most trait pairs. The EAS analysis showed negative values of genetic correlations in most of the 28 pairs of cardiometabolic diseases × respiratory and autoimmune diseases, while there were positive values in the EUR analysis. We then identified the negative genetic correlations satisfying the significance level (P < 0.05/66) in the four disease pairs of EAS analysis (asthma−dyslipidemia: r[g] = −0.29, P = 7.5 × 10^−6; COPD−dyslipidemia: r[g] = −0.26, P = 6.0 × 10^−4; asthma−T2D: r[g] = −0.15, P = 6.0 × 10^−4; and RA−hypertension: r[g] = −0.25, P = 7.0 × 10^−4). The genetic correlations between respiratory and cardiometabolic diseases in EAS population remained negative among the independent datasets (Supplementary Fig. [105]1). To further investigate the negative genetic correlation between asthma and dyslipidemia, we analyzed additional EAS GWAS datasets (asthma from Tohoku Medical Megabank [TMM] and dyslipidemia from Korean Genome and Epidemiology Study [KoGES])^[106]39,[107]40. We observed that 9/12 of the pairs showed negative values of r[g] (Supplementary Fig. [108]2). Since these disease pairs displayed positive genetic correlations in the EUR analyses, our results highlighted the heterogeneity of genetic correlations between the two populations. Fig. 2. Analysis of heritability, genetic correlations, and phenotypic correlations in EAS and EUR populations. [109]Fig. 2 [110]Open in a new tab a A bar plot of the heritability for twelve complex traits in EAS and EUR populations. Trait labels are colored based on the disease categories. b A heatmap of genetic correlations in the twelve complex traits colored by LDSC genetic correlation estimates. P-values of the two-sided tests are adjusted using Bonferroni corrections. The upper and lower triangular matrices show EAS and EUR analyses, respectively. c A heatmap illustrating the phenotypic correlations in the twelve complex traits colored by the natural logarithm of odds ratio (OR). P-values are calculated from two-sided Fisher’s exact tests and adjusted using Bonferroni corrections. EUR populations included UKB samples. OR was calculated via Fisher’s exact tests. h^2: heritability. *: P[uncorrected] < 0.05/66. CAD: coronary artery disease; COPD: chronic obstructive pulmonary disease; HF: heart failure; ILD: interstitial lung disease; RA: rheumatoid arthritis; and T2D: type 2 diabetes. To assess the influence of polygenicity on the heterogeneity of cross-trait genetic correlations, we conducted a local genetic correlation analysis. We partitioned the genome into blocks with high levels of LD using LAVA supplementary program^[111]15 (EAS: 2,197 blocks; and EUR: 2,495 blocks). We then applied SUPERGNOVA^[112]41 for a local genetic correlation analysis for each population (Supplementary Fig. [113]3). After calculating per-block local genetic correlations of all LD blocks for all 66 disease pairs, we compared the proportion of significant blocks with negative correlations between EAS and EUR populations. We observed the consistency of positive directions in phenotypic and genetic correlations between asthma and obesity for both populations^[114]19,[115]21 (38 and 3 positive blocks in EUR and EAS, respectively). In twelve EAS and one EUR phenotype pairs of respiratory and cardiometabolic diseases, LD blocks with negative correlations were the majority. While we confirmed that the genetic correlation between COPD and dyslipidemia supported the population difference in phenotypic correlations (3/5 blocks in EAS and 1/53 blocks in EUR with significant negative correlations), we saw several genetic correlations that disagreed with the known phenotypic correlations^[116]25,[117]42 (asthma–dyslipidemia, asthma–T2D, and COPD–T2D in EAS analysis). The pair of asthma and dyslipidemia showed a remarkable difference in the proportion of blocks with negative correlations (11/12 blocks in EAS and 7/31 blocks in EUR). To further validate the negative genetic correlation between asthma and dyslipidemia, we analyzed asthma from TMM and dyslipidemia from KoGES^[118]39,[119]40. Among the pairs of asthma-dyslipidemia EAS GWAS datasets, we observed significant LD blocks in 10/12 pairs (Supplementary Fig. [120]4). LD blocks with significant negative correlations were the majority among the 9/10 pairs, supporting the negative local genetic correlation between asthma and dyslipidemia in EAS population. The analyses of EAS population showed the negative genetic correlation between asthma and dyslipidemia at both global and local levels. We used Popcorn^[121]43 to analyze the heterogeneity of genetic correlations across the populations and to identify whether respiratory or cardiometabolic diseases have heterogeneity toward genetic correlations (Supplementary Fig. [122]5). As described in the LDSC within-population analysis, many pairs of respiratory and cardiometabolic diseases presented the opposite direction of genetic correlations in EAS population. We next focused on cross-population analysis between EAS and EUR populations. Although cardiometabolic diseases in EAS and respiratory diseases in EUR had positive correlations (for instance, EAS T2D and EUR COPD: r[g] = 0.19, P = 7.8 × 10^−5), respiratory diseases in EAS suggested the negative genetic correlations with cardiometabolic diseases in EUR (EAS COPD and EUR T2D: r[g] = −0.22, P = 0.014). Our results implied that EAS respiratory diseases had the opposite genetic risk components from cardiometabolic diseases. Directions of cross-trait associations in genome-wide polygenic risk scores analysis To validate the per-individual negative genetic correlations between respiratory and cardiometabolic diseases in the independent datasets, we constructed Bayesian PRS using PRS-CSx^[123]32 for each phenotype and population (Fig. [124]3a). For the PRS analyses, we assigned the BBJ1 (EAS) and FinnGen (EUR) to the training datasets and the BBJ2 (EAS) and UKB (EUR) to the testing datasets. We showed the predictive performances of genome-wide PRS in Supplementary Fig. [125]6. The average proportion of heritability genome-wide PRS explained was 53.5% in EAS and 61.1% in EUR analyses, comparable to the preceding PRS study ([126]Supplementary Methods)^[127]44. The standardized regression coefficients (β) of genome-wide PRS showed consistency with the estimates of genetic correlations (Supplementary Fig. [128]7). Because the previous study of genetic correlations in UKB EUR individuals with multimorbidity showed the positive genetic correlations of asthma with dyslipidemia and cardiometabolic diseases^[129]45, we hypothesized that cross-trait genetic correlations between respiratory and cardiometabolic diseases might be in the same direction among the individuals with and without multimorbidity. For validation, we analyzed the individuals with and without multimorbidity separately and compared the directions of genetic correlations. Fig. 3. Predictive performance and cross-trait associations of genome-wide PRS. [130]Fig. 3 [131]Open in a new tab a Results from logistic regression analyses testing the associations between target phenotypes in testing datasets and genome-wide PRS calculated from the training datasets using PRS-CSx. The analyses used identical phenotypes in the training and testing datasets. In the forest plots, dots indicate standardized regression coefficients, and whiskers represent 95% confidence intervals. Disease labels are colored based on the disease categories. b Results from logistic regression analyses testing the cross-trait associations of base and target phenotypes. After excluding all individuals with overlapping base and target phenotypes from the testing datasets, we analyzed the associations between the target phenotype and the PRS generated from the training datasets. The heatmaps are colored based on standardized regression coefficients. P-values of the two-sided tests are adjusted using Bonferroni corrections. *: P[uncorrected] < 0.05/132. CAD: coronary artery disease; COPD: chronic obstructive pulmonary disease; HF: heart failure; ILD: interstitial lung disease; RA: rheumatoid arthritis; and T2D: type 2 diabetes. To analyze the individuals without multimorbidity, we excluded multimorbid individuals for the base and target phenotypes from the testing datasets in the analysis of each phenotype pair. We then tested the associations of binary target phenotypes with genome-wide PRS of base phenotypes (significance level: P < 0.05/132). We adjusted the results from the analysis for age, sex, and top ten genetic principal components (PCs). Concordant with the earlier genetic correlation analysis using the GWAS meta-analysis summary statistics, the genome-wide PRS analysis in EAS presented the negative genetic risk associations between respiratory and cardiometabolic diseases among the independent datasets (Fig. [132]3b). In the EAS analysis of asthma (base phenotype) and dyslipidemia (target phenotype), asthma showed a negative association with dyslipidemia (β = −0.039, P = 1.4 × 10^−4). Furthermore, we tested the cross-trait associations across independent EAS biobanks. We constructed genome-wide PRS using EAS GWAS meta-analysis (BBJ1 + BBJ2). We then tested the cross-trait associations on 99,561 individuals registered in TMM, one of the largest EAS biobanks in Japan. As shown in Supplementary Fig. [133]8 and Supplementary Table [134]3, the negative associations between respiratory and cardiometabolic diseases were validated in the phenotype pairs (i.e., asthma–dyslipidemia and COPD–dyslipidemia). We then investigated genetic risk associations in the individuals with multimorbidity by linear regressions between genome-wide PRS for base and target phenotypes (Supplementary Fig. [135]9). The genetic associations between respiratory and cardiometabolic diseases presented similar results in the individuals with and without multimorbidity. For instance, the genome-wide PRS for respiratory and cardiometabolic diseases presented suggestive negative correlations in the individuals with multimorbidity (e.g., β = −0.077 and P = 0.050 for asthma and dyslipidemia). Our PRS analysis revealed the heterogeneity of genetic associations between respiratory and cardiometabolic diseases in the EAS individuals with and without multimorbidity. Motivated by the different directions between phenotypic and genetic correlations in asthma and smoking and sex-specific smoking behavior in EAS population^[136]46, we performed Fisher’s exact tests between smoking status and asthma stratified by sex (Supplementary Fig. [137]10). Consistent with the previous study from Japan^[138]46, there was a male-specific negative correlation between smoking status and asthma in EAS population (odds ratio=0.84, P = 2.8 × 10^-4). Therefore, we conducted an interaction analysis of cross-trait PRS associations including age, sex, smoking amount (pack-years), sex * smoking amount as the interaction term, and top ten genetic PCs in the models (Supplementary Data [139]2). The interaction term presented significant associations in the analysis of asthma PRS and dyslipidemia in EAS population (EAS: P = 1.7 × 10^−5; and EUR: P = 0.0016). Even after accounting for the interaction, the analysis yielded similar results for the association of asthma with dyslipidemia in EAS population (β = −0.047, P = 1.7 × 10^−5; Supplementary Fig. [140]11). Because sex differences affect asthma risk and serum lipid profiles^[141]47,[142]48, we hypothesized that there might be a sex difference in the negative genetic association between asthma and dyslipidemia. Accordingly, we performed sex-stratified cross-trait association analyses of genome-wide PRS for individuals without multimorbidity (Supplementary Fig. [143]12). The analysis identified the negative association of asthma with dyslipidemia in the EAS males (males: β = −0.051, P = 2.1 × 10^−4; and females: β = −0.051, P = 0.20). The multifaceted analyses for genetic and phenotypic associations supported the negative association between asthma and dyslipidemia, especially in the EAS males. Associations with genome-wide PRS and circulating metabolites and proteins We investigated associations between genome-wide PRS and circulating lipid and metabolite biomarkers to detect shared risk biomarkers with heterogeneous associations between respiratory and cardiometabolic diseases. We utilized blood nuclear magnetic resonance (NMR) biomarker data (Nightingale Health Metabolic Biomarkers) from 51,612 EAS individuals registered in the BBJ1 and those from 245,349 EUR individuals from the UKB. After quality control and the removal of technical variation using ukbnmr R package^[144]49 (Supplementary Data [145]3), we assessed 325 metabolites for the EAS and EUR samples (Fig. [146]4, Supplementary Fig. [147]13, and Supplementary Data [148]4). In the EAS respiratory and autoimmune diseases, the genome-wide PRS presented the opposite directions of association against dyslipidemia (positive correlations with VLDL-C and negative ones with HDL-C in EAS respiratory and autoimmune diseases). Fig. 4. Results from genome-wide PRS association analysis for circulating NMR lipid and metabolite markers. [149]Fig. 4 [150]Open in a new tab We assessed the associations between genome-wide PRS and NMR metabolome, adjusting for age, sex, and the top ten genetic PCs. Heatmaps are colored based on standardized regression coefficients (β) calculated from 325 biomarkers and genome-wide PRS association analysis. We categorized circulating lipid and metabolite markers based on the classification defined in ukbnmr R package, a toolkit for quality control and removing technical variation for NMR metabolome data. As positive controls for the analysis, we found positive correlations of dyslipidemia PRS with VLDL-C-related markers and negative ones with HDL-C-related markers in both populations. CAD: coronary artery disease; COPD: chronic obstructive pulmonary disease; HF: heart failure; ILD: interstitial lung disease; RA: rheumatoid arthritis; and T2D: type 2 diabetes. Based on the consistent negative genetic association between asthma and dyslipidemia in the EAS individuals, we further evaluated correlations of β[_biomaker] of PRS obtained from asthma and COPD analyses with those from dyslipidemia analyses (Supplementary Fig. [151]14). The PRS β[_biomaker] for asthma and COPD negatively correlated with those for dyslipidemia only in the EAS individuals (asthma–dyslipidemia in EAS: Spearman’s rank correlation coefficient [r[S]]=−0.45, P = 3.6 × 10^−6; asthma–dyslipidemia in EUR: r[S] = 0.31, P = 1.9×10^-7; COPD–dyslipidemia in EAS: r[S] = −0.29, P = 0.0062; and COPD–dyslipidemia in EAS: r[S] = 0.41, P = 7.7 × 10^−12). Our cross-trait analysis for genome-wide PRS and NMR metabolome revealed the decreased genetic risks of dyslipidemia in the EAS individuals with respiratory diseases. We further analyzed the association of genome-wide PRS with circulating blood proteins to detect biomarkers affected by the genetic risks of respiratory diseases. We measured protein concentrations using Olink platform from 2071 EAS individuals registered in the BBJ1 and those from 53,058 EUR individuals from the UKB. After bridging sample normalization and applying rank-inverse normal transformation, we assessed 2917 proteins for the EAS and EUR samples. We investigated the association of protein concentrations with genome-wide PRS for asthma, COPD, and dyslipidemia for all individuals with the proteomics data (Supplementary Fig. [152]15 and Supplementary Tables [153]4 and [154]5). We assessed the association of β[_protein] for the two respiratory disease PRSs and identified the negative association of β[_protein] for genome-wide PRS for the two respiratory diseases and dyslipidemia. Among eleven (asthma–dyslipidemia) and four (COPD–dyslipidemia) proteins that satisfied P < 0.05 in both phenotypes and both populations, we identified proteins associated with airway hypersensitivity (STC1) and dysfunction (EZR) in asthma^[155]50,[156]51. Our comprehensive multi-omics analyses showed the genetic negative associations of asthma with dyslipidemia and pinpointed the biomarkers with heterogeneous variations in the EAS individuals. Detecting biological processes responsible for the heterogeneity of multimorbidity with pathway polygenic risk scores Since the dysregulated immune system and lipid metabolism influence the onset of respiratory and cardiometabolic diseases, we assessed biology underlying the global negative genetic correlations by aggregating the effect sizes of variants to the gene sets of immune and metabolic pathways. In detail, we analyzed 335 pathways related to the immune system and lipid metabolism registered in the Reactome subset of curated gene sets in MSigDB^[157]52. With PRS-CSx, we constructed Bayesian pathway PRS to improve the predictive performance. For cross-trait pathway enrichment analyses using Bayesian pathway PRS, we performed three quality control steps: (1) performing Bayesian (PRS-CSx) and clumping + thresholding (C + T) pathway PRS (PRSet) methods in parallel to confirm the consistency of associations, (2) pre-analysis filtering: exclusion of pathways without associations with any base phenotypes using MAGMA^[158]53 pathway enrichment analysis, and (3) post-analysis filtering: exclusion of pathways with false positive associations and those with small predictive performances by setting the thresholds of P-value and Nagelkerke’s R^2 for Bayesian pathway PRS analysis. We described the details of Bayesian pathway PRS in the [159]Supplementary Methods. Briefly, we confirmed that pathway PRS analysis using PRS-CSx showed predictive performances consistent with C + T pathway PRS in phenotypes with sufficient heritability (Supplementary Figs. [160]16–[161]18). For cross-trait Bayesian pathway PRS analyses targeting the immune system and lipid metabolism, we set the P-value threshold using Bonferroni method and Nagelkelke’s R^2 threshold to 0.001 (Supplementary Figs. [162]19 and [163]20). Lastly, we selected 18 pathways significantly associated with at least one phenotype in MAGMA pathway enrichment analyses (P < 0.05/335) from the 335 immune or lipid pathways (Supplementary Data [164]5). For all combinations of the 18 pathways and 132 phenotype pairs (2376 combinations), we conducted a cross-trait association analysis adjusted for the same covariates used in the primary analysis of genome-wide PRS. Applying Bonferroni-corrected P-value threshold of 2.1 × 10^−5 (0.05/2,376) and PRS R^2 threshold of 0.001 to the results of Bayesian pathway PRS analysis, we identified five pathways with multicategorical associations (Table [165]1 and Supplementary Data [166]6) and eight with intracategorical associations (Supplementary Data [167]7). We observed the well-known associations of dyslipidemia with CAD in pathways regulating lipid metabolism. In the EAS analysis alone, pathways regulating lipid metabolism exhibited significant negative associations of dyslipidemia with asthma, COPD, and RA. To examine the heterogeneity in the directions of pathway associations, we further analyzed all 18 MAGMA-selected pathways for the two pairs of phenotypes (asthma–dyslipidemia and COPD–dyslipidemia). All the five lipid metabolism pathways with significant cross-trait association had a negative value of β in the analyses of asthma for EAS population. However, only one of the three significant lipid metabolism pathways showed negative β in the asthma analysis of EUR population (Fig. [168]5a and Supplementary Data [169]8). We examined the five pathways with multicategorical associations that passed the P-value and PRS R^2 thresholds to illustrate the genetic heterogeneity at the pathway level (Fig. [170]5b and Supplementary Figs. [171]21 and [172]22). Both in the individuals with and without multimorbidity, we identified a pathway regulating lipid metabolism by peroxisome proliferator-activated receptors (PPAR) α (Supplementary Data [173]9), a fatty acid-activated transcription factor of nuclear hormone receptor that regulates thermogenesis by stimulating adipocytes and represses interferon-γ production in T cells^[174]54. Furthermore, sex-stratified analyses for the PPARα pathway (Supplementary Fig. [175]23) identified the negative association of asthma with dyslipidemia in the EAS males (males: β = −0.080, P = 4.3 × 10^−9; and females: β = −0.059, P = 1.3 × 10^−4). We demonstrated that the Bayesian cross-trait pathway PRS analysis workflow enables the functional annotations of the genetics underlying the heterogeneous cross-trait associations. Table 1. Significant pathways in Bayesian pathway PRS analyses investigating the associations between immune-mediated and cardiometabolic diseases Population Parental pathway Pathway Phenotype Standardized regression coefficient SE P-value Nagelkerke’s R^2 Base Target EAS Metabolism Regulation of lipid metabolism by PPARα Asthma Dyslipidemia −0.070 0.010 5.5 × 10^−12 0.0013 Transport of small molecules Assembly of active LPL and LIPC lipase complexes Asthma Dyslipidemia −0.088 0.010 4.2 × 10^−18 0.0021 Chylomicron remodeling Asthma Dyslipidemia −0.097 0.010 2.2 × 10^−21 0.0025 COPD Dyslipidemia −0.081 0.010 8.3 × 10^−16 0.0017 Plasma lipoprotein assembly, remodeling, and clearance Asthma Dyslipidemia −0.079 0.010 1.1 × 10^−14 0.0016 Plasma lipoprotein remodeling Asthma Dyslipidemia −0.078 0.010 2.6 × 10^−14 0.0016 COPD Dyslipidemia −0.064 0.010 2.0 × 10^−10 0.0011 RA Dyslipidemia −0.064 0.010 1.8 × 10^−10 0.0011 [176]Open in a new tab We performed logistic regression analyses between pathway PRS of base phenotypes and binary target phenotypes, adjusting for age, sex, and top 10 genetic principal components. P values in the table are uncorrected and two-sided. LPL lipoprotein lipase, LIPC hepatic triacylglycerol lipase, PPARα Peroxisome proliferator-activated receptor α. Fig. 5. Association analyses of immune and lipid metabolism pathways in asthma and COPD. [177]Fig. 5 [178]Open in a new tab a Results from logistic regression analyses investigating the pathway associations of dyslipidemia with asthma and COPD. P-values of the two-sided tests are adjusted using Bonferroni corrections. For the phenotype pairs of dyslipidemia−asthma and dyslipidemia−COPD, we analyzed all 18 pathways with significant associations in MAGMA gene-set analysis. The bar plots provide standardized regression coefficients of the analyzed pathways. Arrows on the bar plots pointed PPARα pathway. *: P[uncorrected] < 0.05/2,376. b Forest plots of logistic regression analyses that assess the pathway regulating the lipid metabolism by PPARα. For pathways regulating lipid metabolism by PPARα, we analyzed cross-trait associations of asthma and COPD with other phenotypes. We generated pathway PRS of asthma and COPD from the training datasets to test the associations with target phenotypes in the testing datasets. In the forest plots, dots indicate standardized regression coefficients, and whiskers represent 95% confidence intervals. P-values of the two-sided tests are adjusted using Bonferroni corrections. ●: P[uncorrected] < 0.05/2376; ○: P[uncorrected] ≥ 0.05/2376. Cell type specificity of traits and pathways in the lungs To acquire further biological insights into the genetic basis of multimorbidity in the lungs, we conducted cell type-specific analyses using the human lung scRNA-seq dataset derived from the Human Lung Cell Atlas (HLCA)^[179]55. The scRNA-seq dataset of the lungs consisted of 50 cell types, classified broadly into endothelial, epithelial, immune, and stromal cells. After calculating gene-level Z-scores through MAGMA gene analyses for all combinations of phenotypes and populations based on the GWAS meta-analysis summary statistics, we selected the top 1000 genes representing the polygenic risk in each phenotype-population combination. For cell type enrichment analysis, we calculated disease scores using scDRS^[180]36 based on the Z-scores of the selected genes. We performed cell type enrichment analyses with the default settings and investigated the differences in cell type enrichment between EAS and EUR populations^[181]56. Among the 600 pairs of 50 cell types × 12 phenotypes, we focused on the combination with prominent cross-population differences in scDRS disease scores between EAS and EUR populations (“Methods”). We showed an overview of the cross-population cell type enrichment analysis in Fig. [182]6. To assess the cross-population differences in cell type enrichment, we started to denote the baseline cell type enrichment per population. Our results pointed to the cell type specificity with known clinical relevance to each phenotype. For instance, we saw the enrichment of immune cells in RA and asthma and endothelial cells and fibroblasts in CAD (Supplementary Fig. [183]24). Among the cell types with significant enrichment, we identified 15 population-specific enriched cell types (Fig. [184]7 and Supplementary Table [185]6). The remarkable enrichment of fibroblasts reflected the large proportion of emphysema in EAS^[186]23,[187]57, a COPD subtype characterized by alveolar destruction (Fig. [188]7c). Fibroblasts repair alveolar damage caused by smoking, but their functional impairment leads to emphysema^[189]58. Additionally, we identified EAS-specific enrichment of dendritic cells in RA (Supplementary Fig. [190]25) and fibroblasts in dyslipidemia. In contrast, two cell types enriched specifically in respiratory diseases in EUR: goblet cells in ILD (Fig. [191]7d) and B cells in asthma (Fig. [192]7e). Our findings suggest that analyzing the differences in cell type enrichment between populations can help explain the heterogeneity of diseases. Fig. 6. Overview of cross-population cell type enrichment analyses. [193]Fig. 6 [194]Open in a new tab Disease associations of and the mean differences in scDRS disease scores at the human lung cell type level. To investigate the disease-associated cell types, we analyzed a scRNA-seq dataset with 50 cell types obtained from the lungs. We applied scDRS with the default settings to calculate disease scores for the twelve complex traits using the GWAS summary statistics and to perform downstream analysis. The left and middle heatmaps show the cell type enrichment in EAS and EUR populations, respectively. Each tile in the left and middle heatmaps is colored based on the proportion of significant cells in the tested cell type. The right heatmap is colored based on the mean differences in scDRS disease scores between EAS and EUR populations. □: FDR < 0.05; ×: FDR of heterogeneity <0.05; +: FDR of t-tests assessing the significance of the scDRS disease score differences across all pairs of the cell types and phenotypes. Fig. 7. Cell types with different enrichment between EAS and EUR populations. [195]Fig. 7 [196]Open in a new tab a A UMAP plot of the scRNA-seq dataset from the Human Lung Cell Atlas colored by coarsest annotations. b Mean differences in scDRS disease scores of population-specific cell types. From all phenotype-cell type pairs identified as significant in EAS or EUR cell type enrichment analysis, the plots described the pairs with prominent differences in the mean scDRS disease scores between EAS and EUR populations. The bar plots provide the differences in disease scores of EAS and EUR populations. c UMAP plots of fibroblasts colored by the enrichment in COPD. The leftmost UMAP plot is colored based on the class of tested cell types. The two UMAP plots in the middle are colored based on the disease scores for EAS (left) and EUR (right) populations. The rightmost UMAP plot is colored based on the differences in disease scores between EAS and EUR populations. d UMAP plots of goblet cells colored by the enrichment in ILD. e UMAP plots of B cells colored by the enrichment in asthma. We then analyzed pathway-level cell type enrichment for pathways with 100–2000 genes and significant associations because the original paper on scDRS validated the statistical power of scDRS for gene sets within the range^[197]36. As the PPARα pathway newly showed the concordance of enrichment between genetic and biological findings which regulates both the immune system and lipid metabolism, we investigated its pathway-level cell type enrichment in immune cells of the lungs. We calculated pathway-level disease scores using Z-scores of genes in the PPARα pathway (Fig. [198]8). Despite the cross-population differences in MAGMA gene analyses between EAS and EUR populations (Fig. [199]8a), we found a similar suggestive cell type enrichment of T cells in asthma and that of macrophages in dyslipidemia (Fig. [200]8c and [201]d). CD4^+ (EAS: P = 0.035 and EUR: P = 0.096) and CD8^+ (EAS: P = 0.041 and EUR: P = 0.083) T cells enriched suggestively in EAS asthma. On the other hand, elicited macrophages (EAS: P = 0.10 and EUR: P = 0.0070) and non-classical monocytes (EAS: P = 0.22 and EUR: P = 0.039) exhibited suggestive enrichment in EUR dyslipidemia. To detect genes contributing to the pathway-level enrichment, we investigated the expressions of genes in PPARα pathway in the immune cells. The expressions of top five genes associated with the traits in MAGMA gene analyses aligned with the cross-trait differences in the enrichment patterns of asthma and dyslipidemia (Supplementary Fig. [202]26). Among the top five genes significant in MAGMA gene analyses for asthma, RORA contributed to the pathway-level cell type enrichment in T cells (Supplementary Fig. [203]27). The pathway-level cell type enrichment analysis pinpointed cell types associated with the genetic risks of asthma and dyslipidemia and gained insights into the diverse biology of PPARα pathway in the diseases. Fig. 8. PPARα pathway enrichment analysis of immune cells for asthma and dyslipidemia. [204]Fig. 8 [205]Open in a new tab a MAGMA gene analysis targeting the pathway regulating the lipid metabolism by PPARα. We plotted the logarithm of uncorrected P-values obtained from the gene analysis of asthma and dyslipidemia in EAS and EUR populations. The plots label the top five significant genes in EAS or EUR populations. P-values of the one-sided tests are adjusted using Bonferroni corrections. The corrected significance thresholds are shown as purple dashed lines. b Coarse annotations of immune cells in the lung tissue. c Enrichment analyses of the PPARα pathway in asthma. After selecting gene-level Z-scores included in the PPARα pathway, we calculated the pathway-level disease scores for eight cell types using scDRS. The UMAP plots are colored based on the disease scores of the PPARα pathway. d Enrichment analyses of the PPARα pathway in dyslipidemia. Discussion In this study, we dissected the genetics underlying the biology of multimorbidity in twelve complex traits relevant to inflammation and the immune system using the three large-scale biobank resources. Global and local genetic correlation analyses revealed the potential effects of pleiotropy on the negative genetic correlations between respiratory and cardiometabolic diseases in EAS population. In the cross-trait genome-wide PRS analysis, the negative genetic association between asthma and dyslipidemia was consistent in the individuals with and without multimorbidity. We observed the negative association between asthma and dyslipidemia, especially in the EAS males. Genome-wide PRS and metabolome association analyses revealed the negative association of asthma PRS with circulating lipid and metabolite biomarkers, supporting the negative genetic correlation between asthma and dyslipidemia in the EAS individuals. We then successfully constructed the Bayesian pathway PRS to analyze the biology underlying multimorbidity in the complex traits. This identified the biological processes of the negative correlations between asthma and dyslipidemia, one of which regulated lipid metabolism via PPARα. Cell type specificities of the traits and pathways with the lung scRNA-seq dataset highlighted the epidemiology and biology of traits. The enrichment of fibroblasts in COPD corresponded to the dominant phenotypic distribution of emphysema in EAS population, characterized by the dysfunction of alveolar repairs. Our pathway enrichment analyses using the lung scRNA-seq dataset revealed the enrichment of the PPARα pathway in immune cells (T cells in asthma and macrophages in dyslipidemia), highlighting cell types associated with biology of the diseases. Among the genes in PPARα pathway, RORA contributed to the enrichment of T cells in asthma for EAS population. Our results demonstrated the negative genetic risk association of the PPARα pathway between asthma and dyslipidemia. The heterogeneity of genetic correlations between asthma and lipid metabolism-related traits was implicated in preceding studies^[206]11,[207]15,[208]59. While dyslipidemia is associated with increased phenotypic and genetic risks of asthma^[209]15,[210]60, a preceding study showed a local genetic correlation analysis found LD blocks with negative genetic correlations of asthma with blood levels of triglycerides and cholesterols^[211]15. Another multi-ancestry genome-wide PRS study suggested a negative association of pediatric asthma with dyslipidemia in a multi-population cohort^[212]61. As the former study focused on EUR population and the latter was for mixed populations, the genetic and biological heterogeneity underlying the negative associations has remained unclear across populations. In EAS population, our results identified the negative genetic correlations between asthma and dyslipidemia from global (Fig. [213]2 and Supplementary Fig. [214]5), local (Supplementary Figs. [215]3 and [216]4), and pathway levels (Table [217]1) with consistency. Our study suggests that the different effect sizes of the cross-trait asthma–dyslipidemia association in the PPARα pathway may drive the heterogeneity of multimorbidity in EAS and EUR populations. As PPARα exerts not only anti- but pro-inflammatory effects via multiple regulatory mechanisms^[218]62, the variety in the effect sizes of the PPARα pathway may be associated with the dissimilar balance of the anti- and pro-inflammatory states. For instance, the PPARα pathway represses interferon-γ production in human T cells in response to androgen^[219]63, potentially leading to the sex and population difference in the association of asthma with dyslipidemia (Supplementary Fig. [220]12). Because cross-population genetic analysis needs larger sample sizes to gain statistical power^[221]43, we anticipate that further research on large-scale non-EUR individuals will validate our findings and provide the basis for personalized medicine for multimorbidity in asthma and dyslipidemia. In EAS population, our analysis revealed the high enrichment of fibroblasts in COPD and obesity (Fig. [222]7). COPD has two main subtypes, one with emphysema (alveolar destruction) and loss of body weight and another with chronic bronchitis and obesity^[223]64. COPD in EAS population is characterized epidemiologically by the dominance of emphysema and weight loss^[224]22,[225]23, unlike in EUR population^[226]57,[227]65. As functionally impaired fibroblasts have a reduced capacity to repair alveolar destruction caused by smoking^[228]58, our results suggested that the significant enrichment of fibroblasts aligns with the high prevalence of emphysema in EAS population (Fig. [229]7). In addition, dysregulating lipolysis in white adipose tissue can lead to obesity, such as via the resistance to fibroblast growth factors^[230]66. Therefore, our results imply that the complex interactions of fibroblasts and adipose tissues may introduce the heterogeneous phenotypic correlation of COPD and obesity, leading to the different prevalence of emphysema and weight change between EAS and EUR populations. This study has several potential limitations. First, we leveraged the five biobank resources to explore the population differences of multimorbidity. The differences in sample sizes can bias our study, especially increasing the false negative results in the traits with small sample sizes. Despite the detailed investigation of our findings using the independent datasets, genotyping platforms, imputation procedures, phenotyping definitions, and biobank-specific confounders between the cohorts can bias our study. In addition, the biobank-scale resources currently do not have detailed information about phenotypes, for example, the onset of asthma and the subtypes of asthma and COPD in the BBJ. Since dyslipidemia has phenotypic associations with COPD and asthma subtypes (allergic and non-allergic asthma)^[231]67, enlarging non-EUR sample sizes with subtype information may help illustrate the complex associations of multimorbidity in subtypes of respiratory diseases. Therefore, future studies using various pipelines and resources will validate our findings and delve into the biology underlying multimorbidity. Finally, as pathways used in the PRS and cell type enrichment analyses are defined based on the known biology, genes and SNPs with unknown biology were beyond the scope of this study. Collectively, our study provided a piece of evidence that biobank-scale GWAS could highlight the heterogeneous polygenicity of multimorbidity across EAS and EUR populations, detect the biology driving the associations of multicategorical traits, and contribute to elucidating the global landscape of heritable multimorbidity risks. Our results demonstrate that exploring diverse populations better promotes understanding the genetic basis of multimorbidity. Methods East Asian samples in BioBank Japan EAS population analysis included samples from BBJ, a hospital-based registry with multi-omics data from genotypes to multitudes of phenotypes. All the participants in BBJ provided written, informed consent approved by ethics committees of the Institute of Medical Sciences, the University of Tokyo and RIKEN Center for Integrative Medical Sciences^[232]68. BBJ1 recruited approximately 200,000 individuals with at least one of 47 target diseases from twelve Japanese medical institutions and collected DNA, serum samples, and clinical information between 2003 and 2007^[233]33,[234]69. BBJ2 is an additional cohort including ~67,000 independent individuals recruited independently of BBJ1 with at least one of 38 target diseases and registered between 2013 and 2018. For meta-analysis and PRS analysis, we included samples derived from the BBJ1 and BBJ2. Tohoku Medical Megabank We used 99,561 TMM samples to validate the cross-trait PRS associations between respiratory cardiometabolic diseases in EAS population. TMM is a population-based prospective cohort that enrolled participants from Miyagi and Iwate Prefectures in the Tohoku region, in the northeastern part of Japan^[235]39. Genotyping was conducted using a custom SNP array for the Japanese population (Japonica Array v.2). Details of imputation and quality control criteria for samples and variants are described elsewhere^[236]70. After imputation, variants with INFO score of <0.3 or minor allele frequency <0.005 were excluded. Quality control of the study participants was performed with the following exclusion criteria: (1) outliers from East Asian ancestry clustering based on the projection PCA with samples of 1KGP3 data; (2) One of pairs within the third degree of kinship;(3) genotype call rate of <95%; (4) without phenotype or covariate information. UK Biobank We obtained the genomic data of UKB, a population-based registry with approximately 500,000 individuals aged between 40 and 69 recruited in the UK^[237]34. The individual registration process is described elsewhere^[238]71. Briefly, the UKB individuals were genotyped using the Applied Biosystems UK BiLEVE Axiom Array or the Applied Biosystems UK Biobank Axiom Array. After quality control, genotype data were imputed with the Haplotype Reference Consortium data and the merged UK10K and 1000 Genomes Project (1KG) Project Phase 3 reference panels using IMPUTE4^[239]34. We analyzed EUR individuals tagged “Caucasian” in UKB Data-Field 22006 and confirmed that all individuals were classified into EUR ancestry using principal component analysis (PCA). FinnGen FinnGen is a large public-private genome research project that collects and analyzes genome and health data from Finnish biobanks and digital health record data from Finnish health registries, with its original phenotypes defined mainly using International Classification of Diseases (ICD) and Anatomical Chemical Therapeutic classification codes^[240]35. We used the GWAS summary statistics of FinnGen Data Freeze 8 (released on December 1st, 2022), for which association tests were conducted using SAIGE (v.0.35.8.8). The datasets we used for the analyses were listed in Supplementary Table [241]2. Phenotype definition According to ICD-10 codes, we defined cases and controls for twelve phenotypes of GWAS and downstream analyses in the BBJ and UKB. In brief, cases were individuals with the twelve phenotypes (asthma, COPD, ILD, RA, obesity, dyslipidemia, smoking, hypertension, T2D, CAD, HF, and stroke). To disentangle the genetic effects of multiple risk factors on cardiometabolic diseases, we analyzed both cardiometabolic disorders and the risk factors in parallel. We excluded individuals with target and related phenotypes from the controls (Supplementary Table [242]2). Because autoimmune disorders share immune genetic backgrounds with asthma^[243]13 and can bias the results of GWAS and post-GWAS analyses for asthma and COPD, we conservatively excluded individuals with autoimmune diseases from the controls for asthma and COPD to gain the robustness of the study. In addition, we included RA, an autoimmune disease available in the datasets, because our previous study showed the positive genetic correlation between asthma and RA in both populations^[244]13. Genotyping and imputation of autosomal chromosomes in BioBank Japan We genotyped the Japanese samples in BBJ1 with the Illumina HumanOmniExpressExome BeadChip or a combination of the Illumina HumanOmniExpress and HumanExome BeadChips. Quality control of samples and genotypes was conducted as described elsewhere^[245]72. We included individuals identified as EAS ancestry based on PCA. We used Eagle (v.2.3) for haplotype phasing of the genotype data and imputed genotype dosages using Minimac3 with the combined reference panel of 1KG Phase 3 version 5 genotype data (n = 2504) and Japanese whole-genome sequencing (WGS) data obtained from the BBJ1 (n = 1037). BBJ2 (~67,000) and a part of BBJ1 (~12,000) individuals were genotyped using Illumina Asian Screening Array-24 v1.0 BeadChip. Quality control and genotype data of the BBJ2 were described elsewhere^[246]68. Using Minimac4, we imputed genotype dosages with the combined reference panel of 1KG Project Phase 3 and Japanese WGS data. GWAS and meta-analysis We conducted GWAS for each phenotype in a single population using a saddle point approximation implemented in Regenie (v.3.1.1)^[247]73 to adjust for case-control imbalance. As covariates for GWAS, we included age, sex, and top ten genetic PCs. We excluded variants with an imputation quality Rsq < 0.7 or minor allele frequency (MAF) < 0.005. For all downstream analyses, we excluded the MHC regions (chromosome 6: 25–34 Mb) due to their complex and strong LD structure^[248]29. We applied UCSC liftOver^[249]74 to convert the genome builds of BBJ and UKB datasets from GRCh37/hg19 to GRCh38/hg38 and annotated variants using SnpSift^[250]75. For usage in downstream analyses, we filtered out variants with imputation quality Rsq < 0.9 or MAF < 0.01 and meta-analyzed each phenotype per population using a standard fixed-effect approach in RE2C^[251]37. We included 4,612,828–4,622,861 SNPs in the EAS GWAS meta-analyses and 7,057,994–7,061,744 SNPs in the EUR GWAS meta-analyses. Global genetic correlation analysis Given the differences in LD structures of diverse populations, we used two software to estimate global genetic correlations that account for genome-wide SNPs and to avoid the influence of mismatched top variants in the GWAS between phenotypes and among populations. For a single-population analysis, we calculated heritability of each phenotype and genetic correlations among phenotype pairs using LDSC^[252]38 (v.1.0.1). We confirmed the concordance of heritability obtained from Popcorn^[253]43 (v.1.0) and performed a cross-population analysis. Based on the standard protocol of each software, we used 1KG Phase 3 reference panels for matched populations. The significance levels of genetic correlation analysis were adjusted by Bonferroni correction (LDSC: P < 0.05/66; and Popcorn: P < 0.05/276). We then performed local genetic correlation and genome-wide PRS association analyses to understand the negative genetic associations between respiratory and cardiometabolic diseases using different genetic analysis methods, as described in the latter sections. Local genetic correlation analysis We applied SUPERGNOVA (v.1.0) to estimate local genetic correlations in the LD-independent segments^[254]41. First, we partitioned the genome into the LD-independent segments using LAVA^[255]15 supplementary program (v.1.0.0) to update the reference panels from 1KG Project Phase 1 to 3. As reference panels for LD estimation, we used the preprocessed EAS (n = 504) and EUR (n = 503) subsets of 1KG ([256]https://ctg.cncr.nl/software/magma). We assessed the significance of the local genetic correlations based on the significance of local genetic covariances, as in the SUPERGNOVA original article. The significance level for local genetic correlation analysis was a false-discovery rate (FDR) < 0.05 adjusted for the Benjamini-Hochberg method. Because the LD structure was different between EAS and EUR populations, the positions of LD blocks were not aligned. Therefore, we conservatively avoided the concordance of significant LD blocks between the populations. Alternatively, we compared the proportion of significant blocks with negative correlations between the two populations. For further validation of the local genetic correlation between asthma and dyslipidemia, we utilized KoGES dyslipidemia GWAS summary statistics to assess the local genetic correlation between asthma and dyslipidemia in EAS population^[257]40. Genome-wide PRS analysis To assess whether the results from cross-trait genetic correlation analyses are not derived from biases in genotyping, imputation, and phenotyping, we performed cross-trait PRS association analyses using the independent datasets generated from different genotyping and imputation platforms and phenotyping procedures. We used training GWAS datasets (BBJ1 for EAS and FinnGen for EUR) to construct PRS and applied the PRS weights to the testing datasets (BBJ2 for EAS and UKB for EUR) to perform association analyses. Based on the original paper^[258]32, we excluded variants with Rsq < 0.8 or MAF < 0.01. After excluding variants located in the MHC region, we ran PRS-CSx (released on July 29th, 2021) to calculate genome-wide PRS for each population. We set φ = 0.01 for analyzing highly polygenic traits and used the default settings for other parameters. We evaluated the predictive performance of the PRS for matched phenotype and population with logistic regression models, adjusting for age, sex, and top ten genetic PCs as covariates. Nagelkerke’s pseudo-R^2 was used to evaluate the predictive performance for all PRS analyses. In addition, we calculated liability-scale R^2 to assess the proportion of heritability genome-wide PRS explained based on the disease prevalence per population^[259]76. We showed the methods and prevalences of traits used to calculate liability-scale R^2 in [260]Supplementary Methods and Supplementary Data [261]10. Because we hypothesized that cross-trait genetic correlations might be in the same direction in individuals with and without multimorbidity, we analyzed individuals with and without multimorbidity separately. For the analysis of the individuals without multimorbidity, we excluded samples overlapping base and target phenotypes from the testing datasets for each phenotype pair. We then investigated the cross-trait associations per population. The significance level for cross-trait analysis was set to P < 0.05/132, adjusted by Bonferroni correction. Next, to adjust for the interaction of smoking behavior and sex on the associations of phenotype pairs, we conducted an interaction analysis. In this analysis, we tested the cross-trait associations of genome-wide PRS, including age, sex, smoking amount (pack-years), top ten genetic PCs, and sex * smoking amount as the interaction term in the logistic regression models. We used lmtest R package (v0.9.40) to perform likelihood ratio tests and calculate the P-values of the interaction term. To evaluate the cross-trait associations of genome-wide PRS for males and females, we performed a sex-stratified PRS association analysis. Next, we assessed the associations of genetic risks between phenotypes in individuals with multimorbidity using genome-wide PRS. For multimorbid individuals with base and target phenotypes in the testing datasets, we tested the associations between the PRS for base and target phenotypes, adjusting for age, sex, and top ten genetic PCs based on the linear regression models. Finally, we assessed the cross-trait associations between respiratory and cardiometabolic diseases across EAS biobanks. We constructed genome-wide PRS using EAS GWAS meta-analysis (BBJ1 + BBJ2) as a training dataset. We then tested the cross-trait associations between respiratory and cardiometabolic diseases in the TMM as a testing dataset, adjusting for the same covariates used in the genome-wide PRS association analysis. Genome-wide PRS association analysis for metabolome and proteome We used targeted high-throughput NMR metabolomics from Nightingale Health Ltd (biomarker quantification version 2020) to measure 249 circulating lipid and metabolite biomarkers from 54,250 serum samples obtained from the BBJ1 participants. For PRS association analysis in EUR population, we obtained the UKB NMR metabolomics data comprising 291,003 samples. We performed quality control for the two metabolome datasets separately to remove technical variation using ukbnmr R package (v.2.2). The details of technical variation adjustments using ukbnmr are described elsewhere^[262]49. We analyzed 325 markers in total generated from ukbnmr. After excluding duplicates, we used samples derived from EAS (BBJ1: 51,612 individuals) and EUR (UKB: 245,349 individuals) populations based on genotype PCA criteria as the testing datasets. All metabolites were subject to inverse rank normalization transformations before association analysis. To remove the sample overlap between training and testing datasets, we performed GWAS after excluding samples measured metabolome or proteome for the BBJ1 and UKB (Supplementary Table [263]7). We meta-analyzed the GWAS summary statistics for each population (BBJ1 and BBJ2 in EAS and UKB and FinnGen in EUR) to construct single-population genome-wide PRS using PRS-CSx. We conducted linear regression analyses for metabolome and PRS, adjusting for age, sex, and top 10 genetic PCs^[264]77. As the BBJ1 NMR metabolomics data were derived from the two batches (47,355 and 4257 individuals), whose genotyping and imputation protocols were different, we first analyzed genome-wide PRS–metabolome associations per biobank and then performed inverse-variance fixed-effect meta-analysis using metagen function implemented in meta R package (v.7.0-0). We measured expressions of circulating proteins using Olink Explore 3072 platform from 2700 individuals from the BBJ1 across three batches. The expression levels in a normalized scale (Normalized Protein eXpression: NPX) were bridge-normalized using OlinkAnalyze R package and subsequently rank-inverse normal transformed. We excluded proteins with missing data in >80% of the samples. For EUR analysis, we obtained bridge-normalized Olink Explore 3072 proteomics data from the UKB measured from 53,058 individuals^[265]78. We selected samples from 2071 EAS individuals in the BBJ1 and 45,631 EUR individuals in the UKB based on the genotype PCA criteria. We assessed the association of genome-wide PRS with the processed expression levels of 2917 proteins measured in both the BBJ1 and UKB, adjusting for the same covariates used in the genome-wide PRS–metabolome association analysis. MAGMA gene-set analysis As one of the three quality control steps for pathway PRS analysis, we investigated the enrichment of pathways related to the immune system and lipid metabolism using MAGMA (v.1.10)^[266]53. First, we selected 335 pathways tagged “Immune system”, “Transport of small molecules”, and “Metabolism” from the curated gene sets of Reactome in MSigDB (released in January 2023)^[267]52. Then, we analyzed the pathway enrichment using MAGMA gene-set analysis function and 1KG Project Phase 3 reference data for matched populations. We used the meta-analyzed GWAS summary statistics for the MAGMA gene-set analysis. The significance level for the analysis was defined as P < 0.05/335, adjusted by Bonferroni correction. In total, 18 pathways passed the significance level and were used for downstream analyses. Pathway PRS analysis Because pathway PRS aggregates risk alleles per pathway, we performed cross-trait pathway PRS analysis to detect functions shared between respiratory and cardiometabolic diseases. We used training GWAS datasets (BBJ1 for EAS and FinnGen for EUR) to construct PRS and applied the PRS weights to the testing datasets (BBJ2 for EAS and UKB for EUR) to perform association analyses. We described the details of pathway PRS analysis described in [268]Supplementary Methods. To investigate the predictive performance and concordance of PRS-CSx with PRSet (implemented in PRSice v.2.3.5), we compared the statistics calculated from the curated gene sets of Reactome in MSigDB^[269]29,[270]32,[271]52. As PRSet targets pathways with 10–2000 genes, we analyzed 1,319 pathways satisfying the requirement. We adjusted the size of gene boundaries in PRS-CSx analyses to that in other MAGMA-based analyses. In detail, we extended SNPs located within the gene coordinates ±10 kilobases (kb) of each gene to include potential regulatory elements selectively. For the PRSet analyses, we included SNPs selectively located within the gene coordinates, encompassing 35 kb upstream and 10 kb downstream of each gene. We evaluated the predictive performance of the two PRS tools based on the single-trait analysis per population. The results from logistic regression analysis were adjusted for age, sex, and the top ten genetic PCs. Next, we investigated the pathway-level cross-trait associations for the 18 significant pathways in MAGMA gene-set analyses. We excluded samples overlapping base and target phenotypes from the testing datasets and analyzed cross-trait associations for each population, adjusting for the same covariates in other PRS analyses. We defined the significance level as P < 0.05/2,376 in the analyses, adjusted by Bonferroni correction. We filtered out pathways with false positive associations and those with small predictive performance by setting Nagelkerke’s R^2 threshold to 0.001 in the cross-trait Bayesian pathway PRS analysis. Cell type enrichment analysis For cell type enrichment analysis, we focused on relevant cell types to avoid overly strict multiple-testing corrections derived from cell types irrelevant to or weakly correlated with respiratory diseases. Because the lungs are the organs responsible for asthma, we analyzed the human lung scRNA-seq dataset derived from the Human Lung Cell Atlas^[272]55, the 50 cell types of which included stromal, immune, epithelial, and endothelial cells. We calculated gene-level Z-scores from the GWAS meta-analysis summary statistics using MAGMA. For the cell type enrichment analysis, we selected the top 1000 genes based on the gene-level Z-scores as a set of putative disease genes. Using the “compute_score” function in scDRS (v1.0.2), we calculated a disease score of each cell in the scRNA datasets by aggregating the expression of the putative disease gene sets and computed 1000 sets of control scores using a random gene set^[273]36. Then, we normalized the raw disease and control scores for each cell. We used the “compute_downstream” function in scDRS with the default settings to associate the putative gene sets with the cell types. The significance level in the association and heterogeneity of cell types was a FDR < 0.05 adjusted using the Benjamini-Hochberg method across all the 600 pairs of the cell types and phenotypes. For cross-population analysis of cell type enrichment, we compared the mean differences in scDRS disease scores between the EAS and EUR populations. We calculated a mean scDRS disease score per cell type for each individual in the scRNA-seq dataset. We then compared the mean differences in scDRS disease scores generated from the EAS and EUR GWAS using a two-sided paired t-test for all the combinations of 12 phenotypes and 50 cell types (adjusted for multiple comparisons using the Benjamini-Hochberg method across all pairs of the cell types and phenotypes)^[274]56. Since many pairs of cell types and phenotypes satisfied the significance level, we wanted to focus on the cell type–phenotype pairs with prominent differences in the mean scDRS disease scores between EAS and EUR populations. Among the 600 pairs of 50 cell types × 12 phenotypes, we defined the combinations satisfying three requirements as population-specific enriched cell types: (1) the combinations were significant in the EAS or EUR cell type enrichment analyses; (2) the mean differences in the scDRS disease scores generated from the EAS and EUR GWAS were significant (FDR < 0.05); and (3) the mean differences in scDRS disease scores generated from the EAS and EUR GWAS exceeded the 95% confidence interval of the distribution collecting all the 600 mean differences (Supplementary Fig. [275]28). For pathway-level cell type enrichment analyses, we used a sequence of gene sets identified in the cross-trait pathway PRS analyses. Because the original paper on scDRS validated the statistical power of scDRS for gene sets with 100–2000 genes^[276]36, we conservatively analyzed pathways with more than 100 genes. Among the five pathways with significant multicategorical associations in respiratory and cardiometabolic diseases, the PPARα pathway was eligible for the analysis. As in the cross-population analysis, we used the default settings of scDRS. We focused on enrichment in immune cells to assess the association of lipid metabolism with the immune system. To further focus on genes contributing to the pathway-level enrichment identified in scDRS, we compared the expression of top genes defined by MAGMA gene analysis to the pathway-level disease scores. Because 25–42 genes in the PPARα pathway satisfied the suggestive threshold (P < 0.05) in MAGMA gene analysis for asthma and dyslipidemia, we calculated the average expressions of the top 5, 10, 15, 20, and 25 genes in MAGMA gene analysis using “scanpy.get.aggregate” function implemented in scanpy (v.1.9.8)^[277]79. We found that the top five genes were representative of the pathway-level enrichment of PPARα pathway in the immune cells, especially in the analysis of asthma (Supplementary Fig. [278]26). Thus, we assessed the expressions of top five genes in the pathway-level cell type enrichment analysis. Reporting summary Further information on research design is available in the [279]Nature Portfolio Reporting Summary linked to this article. Supplementary information [280]Supplementary Information^ (8.8MB, pdf) [281]41467_2025_58149_MOESM2_ESM.pdf^ (222.9KB, pdf) Description of Additional Supplementary Files [282]Supplementary Data 1^ (56.9KB, xlsx) [283]Supplementary Data 2^ (19.9KB, xlsx) [284]Supplementary Data 3^ (23.4KB, xlsx) [285]Supplementary Data 4^ (268KB, xlsx) [286]Supplementary Data 5^ (12.2KB, xlsx) [287]Supplementary Data 6^ (12.2KB, xlsx) [288]Supplementary Data 7^ (13.5KB, xlsx) [289]Supplementary Data 8^ (17.9KB, xlsx) [290]Supplementary Data 9^ (13KB, xlsx) [291]Supplementary Data 10^ (11.5KB, xlsx) [292]Reporting Summary^ (2.5MB, pdf) [293]Transparent Peer Review file^ (5MB, pdf) Acknowledgements