INTRODUCTION
Educational attainment (EA), defined as the highest level of education an individual completes, is influenced by both environmental and genetic factors
1. It is closely related to social outcomes such as income, civic engagement, and overall health, making it a subject of significant interest in social and genetic research
2,3. Genome-wide association study (GWAS) is a widely used study design to explore the genetic architecture of complex traits such as EA. By analyzing large-scale population data, GWAS can associate specific genetic polymorphisms with variation in educational level, which helps elucidate the biological underpinnings and relevant molecular pathways
4.
The genetic architecture of EA has been extensively studied in individuals of European (EUR) ancestry, with thousands of associated loci identified. These studies have revealed that EA heritability is enriched in functional genomic regions active in the central nervous system, particularly neuronal cell types, highlighting the neurobiological basis of the trait
5,6. Beyond its social and behavioral implications, the polygenic risk score (PRS) for EA has significant predictive power and can improve the prediction of complex diseases, including cardiometabolic and psychiatric disorders
7. Despite these advances, the genetic basis of EA in non-EUR ancestries remains underexplored. This is a substantial limitation because pronounced divergence in ancestry-specific allele frequencies and linkage disequilibrium (LD) patterns and potential heterogeneity in gene-by-environment interactions across global education systems mean that relying solely on EUR-based GWAS constrains a comprehensive understanding of the genetic components of EA. The consequences of this limitation are evident in the low cross-ancestry transferability of EA PRS
7,8, which can occur even in the absence of ancestry differences in causal genetic effects. Although a previous EA GWAS in East Asian (EAS) ancestry demonstrated improvements in fine-mapping and PRS power by incorporating multi-ancestry data
9, it failed to identify novel loci, and a comprehensive understanding of the ancestry differences in the genetic architecture of EA is still lacking. Considering the substantial disparity in sample sizes and the limited number of associations discovered in EAS GWAS, significant gaps persist in understanding how the genetic architecture of EA diverges across ancestries.
To address this gap, we conducted a large-scale meta-analysis of EA GWAS comprising 448,558 individuals of EAS ancestry (Fig. 1). We identified a total of 33 independent loci, including novel loci not previously reported in EAS. We also defined EAS-specific association signals by conditioning on EUR-reported signals. Leveraging genetic correlation and Mendelian randomization (MR) analyses, we investigated genetic sharing between EA and other complex traits. Multiple gene prioritization methods and functional annotation datasets were used to further uncover the functional genes and potential mechanisms underlying these associations. Finally, we mapped region-specific cell type associations with EA and the spatial distribution of EA-associated cells in the brain by incorporating single-cell and spatial transcriptomics data.
RESULTS
Within- and cross-ancestry meta-analysis
We collected EA GWAS summary statistics from seven cohorts in mainland China (CKB, BGI-lcWGS, WGEG, WGEA, 23mofang, WeGut, and WGEC) and meta-analyzed GWAS data from six cohorts in Japan (denoted as JPN-meta), totaling 283,326 EAS individuals (Fig. 1; Supplementary Table S1). To assess and control for potential inflation caused by population stratification, we calculated the attenuation ratio using linkage disequilibrium score (LDSC) regression
10 for each cohort (Supplementary Table S2). For large cohorts with an LDSC ratio > 10% (BGI-lcWGS and JPN-meta), we corrected their test statistics using the LDSC intercept. We further incorporated the Chen
et al.'s EA GWAS data from EAS ancestry, comprising 165,232 individuals from the Taiwan Biobank (TWB) and the Korean Genome and Epidemiology Study (KoGES)
9. The genetic correlation between the meta-analysis of our eight EAS cohorts and that of Chen
et al. is 0.994 (
), confirms that the two datasets capture the same underlying genetic signals for EA. After quality control, we performed an inverse-variance-weighted meta-analysis
11 of all EAS datasets, representing the largest EA GWAS in EAS ancestry to date (
). The LDSC ratio indicated no significant evidence of inflation from confounding bias in the test statistics (LDSC intercept = 1.019, s.e. = 0.014 ; attenuation ratio = 0.031, s.e. = 0.023; Supplementary Fig. S1). Of the 9,837,352 tested variants, we identified 732 common variants that passed the genome-wide significance threshold (
; Fig. 2a). The standardized effect sizes of these variants ranged from 0.012 to 0.036 in absolute value, corresponding to approximately 2.6 to 7.9 weeks of schooling per allele. The SNP-based heritability was estimated at 6.1% by LDSC, which is lower than the previously reported estimate in the EAS cohort (9.0%), potentially reflecting increased cross-cohort heterogeneity from the more diverse datasets included in our meta-analysis, as well as cohort-specific characteristics and unmeasured gene-by-environment effects. LD-based clumping
12 of the GWAS summary statistics revealed 33 independent variants (Supplementary Table S3), including 25 variants located outside the seven genome-wide significant loci reported by Chen
et al.9 (Materials and Methods). Statistical fine-mapping at the 33 loci (1-Mb genomic regions centered on the lead variants) using CARMA
13 identified 29 credible sets (CSs) with a 99% probability of including a causal variant, and 50 genome-wide significant variants with a posterior inclusion probability (PIP) larger than 0.1 (Supplementary Tables S4, S5). Conditional and joint analysis (COJO)
14 revealed 31 independent lead variants, including two secondary signals that were masked in single-variant analysis (Supplementary Table S6). Conditioning on genome-wide significant variants reported in the previous EAS GWAS of EA, we identified 521 associated variants that remained genome-wide significant (23 independent loci after clumping; Supplementary Table S36), illustrating the increased statistical power and unique findings achieved by incorporating new datasets.
We calculated the genetic correlation of EA between EAS and EUR using the largest publicly available EUR summary statistics from Lee
et al. (
; EA3)
6. The estimated genetic correlation was 0.768 (
). Although the cross-ancestry genetic correlation reported by the previous EAS study using a different method
9 is higher than ours, the 95% confidence intervals overlap. We also performed age-, sex-, country-, and sampling-characteristic-stratified meta-analyses in our EAS cohorts, and the cross-ancestry genetic correlations varied across stratified groups (Supplementary Note 2 and Tables S44–S46). Overall, these cross-ancestry genetic correlation estimates are not significantly different from one another given the large estimation uncertainty, consistent with a generally shared genetic architecture of EA between ancestries. Because full summary statistics for the more recent EUR meta-analysis by Okbay
et al. (
; EA4)
7 were unavailable, we utilized EA3 for a cross-ancestry meta-analysis (
;
), which identified 809 independent signals. We subsequently compared these signals with the 3,952 reported lead variants from EA4, revealing 23 potential novel signals located more than 500 kb away from, and showing LD
in both EAS and EUR with, any previously reported lead variant in the largest EUR meta-analysis (Fig. 2b; Supplementary Table S7). Furthermore, we found that 73.6% of the genome-wide significant variants discovered only in the EAS GWAS exhibited a higher minor allele frequency (MAF) in EAS than in EUR (EA3). The mean ratio of expected heterozygosity between EAS and EUR across the EAS-identified variants was 1.45, indicating significantly higher genetic variation in EAS at these loci. Of these variants, 98.7% exhibited a consistent direction of effect with EUR, and 76.3% had larger absolute effect sizes in EAS; moreover, all variants discovered only in the EAS GWAS had larger absolute effect sizes in EAS than in EUR (Supplementary Fig. S2). Although we observed some EUR-only signals with larger effect sizes in EAS, none of these variants passed the genome-wide significance threshold in our EAS GWAS (Supplementary Table S40). This demonstrates that, due to the increase in the non-centrality parameter (NCP) of the test statistic driven by both MAF divergence and ancestry-specific effect sizes, our GWAS data can uncover novel associations that would otherwise remain undetectable in EUR studies despite their much larger sample sizes. These results underscore the significance of ancestry diversity in achieving a more comprehensive understanding of EA.
To evaluate the predictive power of our GWAS, we constructed PRS using our EAS GWAS data, the EUR GWAS data (EA3), and the tuned cross-ancestry ensemble of these two datasets (Materials and Methods). We quantified their performance via the incremental in a holdout sample of EAS individuals from the UK Biobank (UKB). The EAS-only and EUR-only PRS explained 1.58% and 0.70% of the phenotypic variance, respectively. The cross-ancestry PRS reached 1.59% (Supplementary Fig. S3 and Tables S8, S9). While the 95% confidence intervals of the prediction overlapped, reflecting the constraints of the test cohort’s sample size, the nominal advantage of the EAS PRS over the EUR PRS, coupled with the marginal gain of the cross-ancestry PRS over the EAS PRS, is consistent with the influence of ancestry differences in LD and allele frequencies on PRS transferability, with possible additional contributions from ancestry-specific genetic architecture. Sensitivity analyses, including comparison with the previous EAS study, leave-one-cohort-out validation, and an alternative UKB-based LD reference, yielded consistent results, with the cross-ancestry PRS showing the clearest advantage in the leave-one-cohort-out analyses (Supplementary Note 4 and Table S37).
Identification of EAS-specific genetic associations
To identify EAS-specific genetic associations, we performed conditional analyses by adjusting the EAS GWAS summary statistics for genetic associations reported in EUR. For this analysis, we utilized the 3,952 independent lead variants reported by Okbay
et al. (EA4), which represent the largest and most contemporary set of EA associations identified in EUR to date. Given the challenges posed by cross-ancestry LD differences, the potential absence of EUR lead variants in the EAS data, and the risk of collinearity among covariate variants when modeled using EAS LD, we implemented a two-step refinement process. First, we expanded the lead variants into a dense list using a EUR LD reference panel
15 to include all variants in LD with the lead variants (
), ensuring that proxy variants could be used even if certain EUR lead variants were absent in the EAS data. Second, to prevent collinearity in the conditional analysis model, we employed model selection implemented in GCTA-COJO to identify jointly significant variants from this expanded set based on EAS GWAS and LD data. This process was designed to maximize the inclusion of EUR signals as covariates in the conditional analysis model while accounting for two key factors: the unavailability of full summary statistics from EA4 and the inherent differences in LD structure between EAS and EUR (Materials and Methods).
We identified 88 genome-wide significant variants after conditioning on EUR signals (Fig. 2a). Notably, 18.8% of these variants were located in coding regions, and 28.3% were located in functional regions such as promoters and enhancers (Materials and Methods). Compared to all tested variants, these variants showed significant enrichment in coding regions (). LD-based clumping of the conditionally significant associations identified 11 independent EAS-specific associations (Supplementary Table S10). One of the associations, a secondary signal in the original EAS meta-analysis (Supplementary Table S6), also reached genome-wide significance after conditioning on the EUR associations. Four associated variants were located more than ± 500 kb from any EUR-reported lead variant (specifically rs10893782 on chr11, rs1844434 on chr3, rs4410534 on chr4, and rs3809297 on chr12). The remaining seven associated variants mapped within established EUR-reported loci, representing EAS-specific signals distinct from the EUR-reported signal at the same locus (Supplementary Fig. S4). These findings indicate potential ancestry-specific genetic architecture underlying EA.
Prioritization of high-confidence EA-associated genes
To dissect the function of EA-associated loci identified in our EAS GWAS, we integrated multiple gene prioritization methods, functional annotation datasets, and molecular quantitative trait loci (xQTL) datasets to identify EA-associated genes and generated gene rankings based on the number of methods by which each gene was prioritized (Materials and Methods), thereby enhancing the robustness and interpretability of the prioritized genes. Our framework incorporates six gene discovery methods based on four rationales
16: 1) gene-based association analysis (MAGMA
17 and mBAT-combo
18); 2) variant-to-gene (V2G) mapping; 3) integrative analysis of GWAS and xQTL (SMR
19 and COLOC
20); and 4) network-based prioritization (PoPS
21). The V2G mapping incorporated four types of cis-regulatory element (cCRE)-gene connection annotations across various tissues and cell types, while the integrative analysis of GWAS and xQTL leveraged four layers of EAS xQTL datasets from multiple tissues and cell types.
Using this integrative framework, we identified 111 protein-coding genes underlying the genome-wide significant loci, nearly half of which (46.8%) were supported by multiple gene prioritization methods (Fig. 3a; Supplementary Tables S11–S17). While COLOC identified a high number of unique genes, the overall intersection across all methods was limited, highlighting the inherent challenges in achieving convergence among distinct prioritization methods. Consequently, we defined a ‘high-confidence’ set of 28 genes, supported by three or more independent methods (Fig. 3b), as such cross-method convergence provides robust evidence for their functional association with EA. For example,
MED27 was prioritized by all six methods and has not been reported by previous EA studies. The EA association signal at the
MED27 locus colocalized with the
MED27 eQTL signal (COLOC PP4 = 0.85), and the lead variant was linked to
MED27 by an active enhancer–promoter interaction detected by PCHiC in brain, suggesting a potential regulatory mechanism by which EA-associated variants impact its expression, implicating this gene in pathways relevant to EA (Fig. 3c).
MED27 is a known component of the mediator complex that activates transcription, while loss-of-function mutations in
MED27 have recently been reported to cause neurodevelopmental disorders
22-24, indicating its critical role in neurological function.
Moreover, we prioritized four high-confidence genes through EAS-specific associations, including
DCAKD,
NMT1,
PLCD3, and
CBLN2, which represent novel biological targets derived from the incorporation of EAS data.
DCAKD, a gene predicted to be involved in the coenzyme A biosynthetic process, was also prioritized by all six methods. The lead variant (rs4525538 on chromosome 17) of its eQTL in a Chinese brain dataset suggested a potential mechanism in which increased
DCAKD expression contributes to higher EA (
;
). The
DCAKD locus also exhibits an EAS-specific association (the lead variant, rs7502512, was linked to
DCAKD by a potential enhancer detected via the activity-by-contact interaction; Supplementary Fig. S14), with a similar MAF but a much larger effect size in EAS than in EUR (
,
;
,
). Another one of these prioritized targets,
CBLN2 has been experimentally demonstrated to promote prefrontal spinogenesis, an essential process in neuron development
25, and has been shown to maintain cognitive function and task performance in mice
26, strengthening its functional association with EA. However, neither the
CBLN2 gene nor the GWAS locus has been reported in previous GWAS studies in EUR (Fig. 3d), likely due to the pronounced difference in allele frequencies of variants at this locus. For instance, a coding variant, rs8094389, in
CBLN2 showed a large divergence in its MAF between ancestries (
expected-heterozygosity ratio
).
The remaining high-confidence genes can also be categorized into two major biological domains: neurobiology and transcriptional regulation. For example,
NRXN1,
MCTP1, and
LRFN2 are essential components of neurotransmission, with
NRXN1 and
MCTP1 involved in calcium signaling and
LRFN2 involved in synaptic adhesion. The critical nature of these genes is underscored by previous studies that linked
NRXN1 to several psychiatric disorders
27,28 and
LRFN2 to learning disabilities
29. Additionally, we identified several genes involved in cellular regulation, including
CUL3 and
AUTS2, both of which are established candidate genes for diverse neurodevelopmental disorders
30,31, as well as several zinc finger protein-coding genes, including
ZKSCAN4 and
ZSCAN9.
These high-confidence genes demonstrate the power of multiple-method integration and ancestry diversity for gene discovery, suggesting that genetic variants may influence EA by regulating genes that mediate neurotransmission, neurodevelopment, and transcriptional pathways. Collectively, this list of high-confidence genes, prioritized through multiple independent lines of evidence, provides a strong foundation for downstream experimental validation of causal mechanisms.
Genetic correlation and causal inference across ancestries
We used LDSC to estimate the genetic correlation (
) between EA and 259 complex traits in EAS (Supplementary Table S18), identifying 48 traits with significant correlations (
), spanning four domains: biomarkers, diseases, drug usage, and nutrition intake (Fig. 4). We observed a significant inverse genetic correlation between EA and the risk of metabolic and cardiovascular diseases, including type 2 diabetes and heart failure. These results align with the negative correlations between EA and its corresponding biomarkers such as body mass index (BMI), fasting glucose, and triglycerides, and the positive correlation with high-density lipoprotein-cholesterol (HDL-C). Moreover, EA showed negative correlations with corresponding drug usage, such as diuretics and calcium channel blockers, and positive correlations with nutrition intake. However, the risks of myopia (indicated by negative values of refractive error measures
32), schizophrenia, and prostate cancer (PRCA) were observed to be positively correlated with EA. Most of these significant genetic correlations were consistent with those observed in EUR, with age at menarche, schizophrenia, pulmonary tuberculosis, and glucose in urine significantly genetically correlated with EA in EAS only. Notably, the risks of glaucoma and uterine fibroids showed significant but opposite signs of genetic correlations with EA in different ancestries (Supplementary Table S18).
To investigate putative causal relationship between EA and these traits, we performed GSMR
33 analyses using EA as the exposure. Because there is potential sample overlap between some traits and EA, we incorporated two approaches for sensitivity analyses (Materials and Methods), one for traits with an independent GWAS dataset and another for traits without an independent dataset, and excluded traits that were significant in neither approach. We identified putative causal associations between EA and 28 traits in EAS (Supplementary Table S19), with the estimated effect sizes (
) consistent in direction with the pairwise genetic correlations. BMI (FDR-adjusted
; FDR-adjusted
using an independent GWAS) exhibited the strongest evidence for both genetic correlation and causal relationship with EA. Triangulation across multiple MR methods yielded concordant effect directions; while not all individual causal estimates achieved statistical significance, this directional consistency across multiple models supports the robustness of the identified putative causal trends (Supplementary Figs. S5, S6). We also performed MR analyses using the same methods and the same traits in EUR, and identified 32 traits with putative causal relationships with EA (Supplementary Table S20). All traits significant in both ancestries exhibited the same direction of causal effects. Among the six traits with ancestry-dependent genetic correlations with EA,
i.e., opposite in sign between ancestries or significant only in EAS, age at menarche and uterine fibroids showed evidence of a causal relationship with EA only in EAS, with signs of effect sizes consistent with the genetic correlation results. We observed a potential reverse causal effect of BMI on EA across ancestries (Supplementary Tables S33, S34), indicating a bidirectional relationship between them.
Mapping EA-associated cell types and spatial distributions
As in previous studies, we conducted stratified-LDSC (S-LDSC
34,35) to characterize the biological contexts in which the genetic variants identified in EAS manifest their effects on EA. Partitioning heritability using S-LDSC with baseline annotations (Materials and Methods) revealed significant enrichment of heritability (
) in non-synonymous variants and gene-related regulatory elements (Supplementary Fig. S7a and Table S21), such as chromatin regions marked by monomethylated lysine 4 on histone H3 (H3K4me1). Analysis leveraging annotations of tissue-specific activating histone marks further demonstrated significant heritability enrichment in brain-specific regulatory elements (Supplementary Fig. S7b and Table S22). Specifically, chromatin regions marked by H3K4me1 in the male fetal brain showed the strongest enrichment among all annotations (Enrichment
), while those marked by acetylated lysine 27 on histone H3 (H3K27ac) in the dorsolateral prefrontal cortex showed the strongest enrichment compared to annotations of other brain regions (Enrichment
). These heritability enrichment patterns in brain tissues were highly consistent with previous results in both EAS and EUR ancestries, reinforcing the central nervous system as the most relevant tissue domain for EA.
To better understand how brain function mediates genetic effects on EA, we mapped EA-associated cell types using high-resolution brain single-cell RNA sequencing data (scRNA-seq) and spatial transcriptomics (ST) data. While the EAS-specific genetic associations demonstrate the unique contribution of our data, the high genetic correlation observed between EAS and EUR () and the concordant enrichment of biological annotations strongly support the hypothesis that the underlying molecular and cellular architecture of EA is conserved across ancestries. Moreover, the cross-ancestry genetic correlation between EA and IQ was 0.595 (), estimated by Popcorn, similar to the within-ancestry genetic correlation between EA and IQ in EUR (, ) by LDSC, suggesting that EA shares core biological mechanisms with IQ across ancestries. To maximize statistical power and deepen biological insights, we leveraged the cross-ancestry meta-analysis summary statistics (integrating our EAS GWAS with the EA3 GWAS) for cell mapping. By applying S-LDSC to the cross-ancestry meta-analysis, we confirmed the gain in power: it recovered the same enrichment profile as the EAS-only analysis (Pearson's r = 0.790 for the enrichment P-values of significantly associated functional genomic regions and 0.677 for tissue-specific open chromatin regions), while identifying more associated regions and tissues with smaller enrichment P-values (Supplementary Tables S38, S39). We then used the cross-ancestry meta-analysis of EA for cell mapping. Acknowledging that these cell-trait associations are driven largely by the larger EUR sample within the cross-ancestry meta-analysis, we validated them using ancestry-stratified summary statistics, which showed highly concordant cell-type association patterns across ancestries, with the cross-ancestry data offering increased power (Supplementary Note 5, Figs. S9–S11 and Tables S28–S32).
We first integrated scRNA-seq data from the human prefrontal cortex (PFC)
36 with GWAS summary statistics using scDRS
37. Excitatory neurons, inhibitory neurons, and oligodendrocyte precursor cells (OPCs) were significantly associated with EA (
; Fig. 5a–c). Excitatory neurons exhibited the highest trait score (referred to as ‘disease score’ in scDRS) and the most significant association among all cell types (
; Supplementary Table S23), suggesting that EA-relevant genes have elevated expression in excitatory neurons. These results confirm findings from previous studies that EA heritability is more enriched in neurons than in glial cells, while also refining associations with specific neuronal cell types and identifying additional associations with OPCs, consistent with established evidence that OPCs play a role in neural circuits by forming synapses with neurons
38. To validate this finding using an alternative method, we performed cell-type enrichment analysis via S-LDSC (Materials and Methods). Consistently, EA heritability was significantly enriched in excitatory and inhibitory neurons (
; Supplementary Table S24). However, OPCs were not identified as a significantly associated cell type in the S-LDSC analysis. This discrepancy likely stems from differing mapping resolutions between the two frameworks: while scDRS preserves transcriptional heterogeneity within cell types and may capture specialized subpopulations or states of OPCs, the aggregation of cellular-level expression into broad cell-type annotations in S-LDSC may attenuate these signals.
To characterize the spatial distributions of cells associated with EA, we integrated ST data from the mouse brain
39 and GWAS summary statistics using gsMap
40. We confirmed excitatory neurons as the most relevant cell type for EA (
; Supplementary Table S25), and astrocytes exhibited the most significant association with EA among non-neuronal cell types in this dataset (
). We also observed heterogeneous association strengths among excitatory neurons across brain regions (Fig. 5d), with the cerebral cortex harboring the most significantly associated excitatory-neuron subpopulations (Supplementary Table S26). These results were highly consistent with findings from additional human scRNA-seq data across 10 brain regions
41, confirming that the cerebral cortex is the primary associated brain region and demonstrating that neuronal subpopulations exhibit region-dependent association strengths (Supplementary Fig. S8 and Table S27). Furthermore, visualizing the spatial distribution of EA-associated cells based on ST data revealed that this distribution was not only heterogeneous across brain regions but also varied within brain regions (especially within the cortex), and that this spatial pattern was independent of predefined cell types (Fig. 5e, f; Supplementary Fig. S10b). Similar findings have been reported in a recent study in the mouse brain, where heterogeneous functions of neuronal subpopulations throughout the cortex were observed in the processes of decision-making
42. These results strongly suggest that the neurobiological process underlying EA is influenced by spatially dependent cellular mechanisms and local microenvironmental factors.
DISCUSSION
This study represents the largest EA GWAS in individuals of EAS ancestry to date, providing a systematic exploration of the ancestry-specific and conserved genetic architecture of EA across EAS and EUR ancestries. By increasing the EAS sample size, we achieved a threefold increase in identified genetic associations compared to previous efforts, with 11 signals independent of EUR-reported signals. The heterogeneity in the genetic basis of EA across ancestries led to the identification of novel genes and ancestry-specific genetic correlations between EA and other complex traits. Leveraging the conserved biological mechanism of EA across ancestries, we mapped the specific subtypes and spatial locations of cells across brain regions that were significantly associated with EA. These results provide a more comprehensive landscape of the genetic component of EA.
One of the major findings of this study is the discovery of ancestry-specific genetic associations. By conditioning on the reported signals from the largest published EUR GWAS (EA4), we identified 11 lead variants that are statistically independent of the established EUR associations. We hypothesize that this independence arises from two primary scenarios: 1) Low MAF in EUR: This is supported by the fact that 9 of the 11 EAS-specific signals had a higher MAF in EAS than in EUR, with a mean ratio of expected heterozygosity of 2.27 between EAS and EUR, consistent with findings from recent cross-ancestry studies suggesting that most novel associations in non-EUR ancestries can be attributed to a higher ancestry-specific MAF
43. With this increased discovery power, our data lay a foundation for detecting genetic associations that were underpowered in other ancestries due to a small MAF and facilitate novel discoveries in subsequent analyses. 2) Differences in tagging variants between ancestries: this is supported by several EAS-specific signals identified at EUR-reported loci, where the causal variant may be weakly tagged by the EUR lead variant but strongly tagged by a different variant in EAS owing to its distinct LD pattern. These loci benefit from cross-ancestry fine-mapping, as exploiting divergent LD patterns improves the resolution in separating tag variants from causal variants. For example, cross-ancestry fine-mapping results at a locus on chromosome 2 yielded five CSs (Supplementary Fig. S4a), suggesting the presence of five distinct putative causal signals in this region. CS 1, 4, and 5 each comprised a single variant with a PIP close to 1 in the center of the locus. These variants exhibited moderate LD (
) with the top SNP in EAS but weak LD (
) with the top SNP in EUR, indicating they represent shared causal variants that are tagged differently across ancestries. Compared to within-ancestry fine-mapping results at this locus, cross-ancestry fine-mapping identified more CSs and reduced the CS sizes in EUR. By applying cross-ancestry fine-mapping to the loci significantly associated with EA in both EAS and EUR, we identified more credible sets and more high-PIP variants than EUR-only fine-mapping (Supplementary Table S41), suggesting additional candidate causal variants and reduced uncertainty.
Through the integration of multiple gene prioritization methods, functional annotations, and xQTL data, we identified 28 high-confidence genes at the significant loci of our EAS GWAS. We defined confidence based on convergence across six methods, a stringent criterion given the distinct underlying assumptions and datasets of each approach. The functions of these high-confidence genes converge on three key biological domains: neurotransmission, neurodevelopment, and transcriptional regulation. Genes involved in neurotransmission (e.g.,
NRXN1,
CBLN2,
MCTP1) suggest that synaptic communication efficiency is a key biological factor underlying variation in EA, potentially by modulating complex brain functions such as working memory
44. Similarly, neurodevelopmental genes (e.g.,
MED27,
CUL3,
AUTS2) likely influence early-life cognitive performance, thereby contributing to divergent educational outcomes. Our high-confidence gene set provides a strong foundation for future experimental validation of the biological mechanisms underlying EA. Furthermore, we identified novel target genes through EAS-specific associations. A notable example is
CBLN2; despite its established role in synapse regulation, its association signal for EA was detected in EAS but not in EUR (Fig. 3d). In contrast,
NRXN1, a conserved postsynaptic cell adhesion molecule in the same signaling pathway (
CBLN2-
NRXN1-
GLUD45), was prioritized in both ancestries. This contrast reinforces the hypothesis that the absence of a
CBLN2 signal in EUR is primarily due to limited statistical power driven by allele frequency differences, suggesting that while the biological pathway is conserved, the specific variants are ancestry-dependent.
We further investigated the shared and divergent genetic architecture between EA and other complex traits across ancestries. Consistent with findings in EUR, we observed robust genetic correlations between EA and favorable metabolic profiles, biomarkers, and pharmacotherapeutic usage. This suggests that the genetic liability for higher EA is intrinsically linked to better health outcomes, highlighting the potential utility of integrating EA-related genetic factors into health equity frameworks. We also investigated the distinct genetic correlations between EA and diseases across ancestries (Supplementary Table S18). For instance, while we observed a significant positive genetic correlation between EA and the risk of uterine fibroids (UF) in EAS (
, FDR-corrected
), the corresponding estimate was negative in EUR (
, FDR-corrected
, though we caution against over-interpreting this cross-ancestry difference given its modest magnitude. Although the high cross-ancestry genetic correlation of UF (
,
) suggests that its core biology is shared across ancestries, its fluctuating SNP-based heritability
46 across ancestries indicates that it is prone to ancestry-specific gene-by-environment interactions, similar to EA. We performed local genetic correlation tests using LAVA
47 to further investigate the potential mechanism. Only one locus with a significant univariate signal was eligible for testing, and no significant local correlation was observed. Nevertheless, we observed a potential positive causal effect of EA on UF in EAS. The putative causal relationship between EA and UF was negative and significant in the GSMR result in EUR, but it did not pass the sensitivity analysis for sample overlap using MRlap
48. In another case, the negative genetic correlation and putative causal inference between EA and age at menarche (AAM) were observed exclusively in EAS. The potentially overlapping biological mechanism underlying these two traits is supported by a recent study reporting that biological pathways, including head development, synaptic signaling, and behavior, are enriched among high-confidence AAM genes
49. Overall, given the shared biology across ancestries for these traits (as illustrated by cross-ancestry genetic correlation results), we think the distinct genetic correlations between EA and complex traits such as UF and AAM across ancestries may reflect unmeasured gene-by-environment effects. For the putative causal inference, EA could be an important exposure, but we suggest being cautious about this result regarding collider bias, and replication analyses using different datasets will be required in the future.
Moreover, we mapped the distinct subtypes and spatial locations of cells significantly associated with EA across brain regions by integrating scRNA-seq and ST data with the EA GWAS meta-analysis summary statistics. The convergence of multiple datasets and methods revealed that cortical neuronal cell types, particularly excitatory neurons, exhibited a stronger association with EA than other cell populations. The increased resolution provided by single-cell data enabled the identification of not only refined neuronal cell types, such as intratelencephalic excitatory neurons and distinct medium spiny neurons (MSNs), which express dopamine receptors and are implicated in reward-based behaviors (Supplementary Fig. S8 and Table S27), but also subpopulations of nonneuronal cells, including astrocytes and OPCs (Fig. 5b, d). By linking the EA-associated genetic variants to the specific expression patterns of these cell populations, we suggest that the genetic component of EA may involve the modulation of synaptic signaling and plasticity
50. This process is known to be centered on excitatory neurons
51 and is precisely modulated by GABA released from inhibitory neurons
52, by the clearance of neurotransmitters by astrocytes
53, and by the direct interactions between OPCs and neurons
38. The incorporation of multiple brain regions and spatial information indicated that the cellular mechanisms underlying EA depend jointly on cell type, brain region, and the spatial location of cells within regions. These results provide potential cellular mechanisms that influence EA, enhancing our understanding of its neurobiological basis.
The SNP-based heritability estimate of EA in our GWAS was lower than previously observed. Given the observed heterogeneity in EA genetic architecture between EAS and EUR, we expect differences in EA heritability across ancestries. Here, we mainly discuss potential factors underlying the difference in SNP-based heritability between our EAS GWAS and the previous EAS GWAS by Chen
et al. We hypothesize that cross-cohort heterogeneity, cohort-specific characteristics, and unmeasured gene-by-environment effects could contribute to the difference in SNP-based heritability. Compared with the previous EAS GWAS (a meta-analysis of two prospective cohorts: TWB and KoGES), our new datasets included samples from prospective studies, clinical centers, and commercial genetic testing platforms. Multiple factors vary across these datasets and can contribute to cross-cohort heterogeneity and decreased heritability estimates in the meta-analysis, including genotyping processes (Supplementary Table S1), phenotype distributions, and cohort characteristics (Supplementary Table S35). For example, the BGI-lcWGS cohort leveraged an ultra-low-coverage whole-genome sequencing technique for genotyping and imputed genotypes to Chinese samples from the 1000 Genomes Project Phase 3 (1KGP3), while the JAPAN_2 to JAPAN_5 cohorts leveraged the Affymetrix Axiom Japonica arrays for genotyping and imputed genotypes to a reference panel built from 3,552 within-cohort samples. Phenotypic heterogeneity was also observed: mean years of schooling (
EduYears) ranged from 8.8 (CKB) to 16.9 (WGEA), with distinct phenotype distributions across ISCED levels. Average birth year differed by up to 47 years across cohorts (1938 in JAPAN_1 vs 1985 in WGEA), and average
EduYears and birth year were strongly correlated across the new cohorts (Pearson’s
,
), indicating that era-dependent environmental factors, such as the expansion of educational opportunities across successive generations, contribute to the phenotypic differences among cohorts. Even among cohorts with similar birth years (1952−1956), average years of schooling differed between CKB (8.8 years) and the JAPAN_2 to JAPAN_5 cohorts (12.1–12.9 years), possibly reflecting differences in education systems between China and Japan. These factors contribute to cross-cohort heterogeneity and imperfect genetic correlations, thereby lowering the SNP-based heritability of the GWAS meta-analysis, consistent with a recent study showing that heritability estimates decrease as the number of cohorts in a meta-analysis increases when genetic correlations between cohorts are less than one or when cohort-specific heritability is heterogeneous
54. Furthermore, cohort-specific characteristics and unmeasured gene-by-environment effects may also contribute to lower cohort-level heritability estimates. For the new datasets from the genetic testing platforms, we found their phenotypic standard deviations to be smaller than those of samples from TWB and KoGES (Supplementary Fig. S12a), which may indicate that these samples are more clustered. For the BGI-lcWGS cohort, samples were from pregnant women recruited at a hospital in Wuhan, which could lead to greater homogeneity among samples and an attenuated estimate of heritability. The WeGut cohort had characteristics similar to those of the KoGES in the previous EAS study. Its participants, who were born around 1957, were recruited from prospective studies and had
EduYears distributions similar to those of KoGES participants. Consistent with these similarities, we observed a higher SNP-based heritability estimate in the WeGut data, though the relatively large standard error warrants conservative interpretation. For large cohorts such as CKB and JPN-meta, we found heterogeneous heritability estimates across age- and sex-stratified groups. For example, the SNP-based heritability estimate from rural samples in CKB was 0.0333 (
), whereas the estimate from urban samples in CKB was 0.0826 (
). The SNP-based heritability estimate from samples with an enrollment age (defined as birth year plus 6) before 1945 in JPN-meta was 0.0447 (
), whereas the estimate from samples enrolled between 1946 and 1965 in JPN-meta was 0.0815 (
). For samples enrolled in the same period (1946–1965), the SNP-based heritability estimate in CKB (
,
) differed from that in JPN-meta (Supplementary Table S47). These observations may indicate interactions between environmental and genetic factors that were not captured in the current model. Although the SNP-based heritability estimated from our EAS GWAS was lower than in previous studies, the high genetic correlation between our newly collected cohorts and those of Chen
et al., together with the increased power of the EAS meta-analysis, strongly supports the robustness of our data.
Several other limitations exist in our study. First, although we attributed the novel associations discovered only in EAS primarily to differences in MAF and tagging variants, we also observed heterogeneity in variant effect sizes across ancestries. For example, the lead variant at the
DCAKD locus on chromosome 17 showed a similar MAF but heterogeneous effect sizes across ancestries and was identified as an EAS-specific association in our conditional analysis (Supplementary Fig. S14). Furthermore, all the genome-wide significant variants discovered only in EAS showed consistent direction and larger absolute effect sizes than in EUR (Supplementary Table S40). These findings reflect heterogeneity in the genetic architecture of EA across ancestries, leading to differences in heritability estimates, imperfect genetic correlations, and a partial loss of PRS relative accuracy
7,8. However, this factor is closely intertwined with MAF differences and correlated with unmeasured gene-by-environment interactions, which could be further investigated in future studies. Second, although our GWAS achieved a higher nominal predictive
than the previous EAS study in predicting the phenotypes of EAS samples in the UKB (Supplementary Note 4), the overall limited SNP-based heritability and the large standard error of the prediction accuracy estimated from the small test cohort preclude the use of our EA PRS for polygenic prediction of other complex traits. Third, we primarily incorporated blood xQTL datasets in the integrative analyses for gene prioritization, given the scarcity of brain xQTL datasets in EAS. This limitation may have led to the omission of functional genes regulated by brain-specific genetic effects that colocalize with EA GWAS signals.
In summary, this study provides a deeper understanding of the genetic and biological components of EA, specifically within EAS ancestry. Our findings demonstrate that while the neurobiological architecture of EA is largely shared across ancestries, novel loci, gene targets, and genetic correlations with other complex traits emerged only through the inclusion of EAS data, underscoring the critical importance of ancestry diversity for a comprehensive understanding of the genetic landscape.
MATERIALS AND METHODS
Ethical approval
This study was approved by the Ethics Committee of Westlake University (approval no. 20200722YJ001).
Phenotype definition
We adopted a phenotype mapping approach similar to that of previous studies. Individual educational information collected from different types of questionnaires was first classified using the International Standard Classification of Education (ISCED) 2011 levels. These ISCED levels were subsequently converted to years of schooling (EduYears) using official country-specific mapping files from the ISCED website, with each cohort allowed the flexibility to adjust the exact level-to-year mapping within the range provided by those files. The resulting years of schooling were standardized () and used as phenotypes. Detailed mapping standards are provided in Supplementary Note 1 and Tables S49, S50, and cohort-specific ISCED mappings are provided in Supplementary Table S35.
Cohort descriptives
Our study incorporated seven Chinese cohorts representing diverse geographic regions: (1) China Kadoorie Biobank (CKB): A prospective cohort of 80,466 adults aged 30–79 from 10 geographic regions across China was genotyped using a customized Affymetrix Axiom array
55. Variants were phased with SHAPEIT2 and imputed to the 1KGP3 EAS reference panel. CKB Data from baseline are available under the CKB Open Access Data Policy to
bona fide researchers; full details of the CKB Data Sharing Policy are available at
ckbiobank.org. The CKB study was approved by the Ethical Review Committee of the Chinese Center for Disease Control and Prevention (Beijing, China), the Institutional Review Board of Peking University (Beijing, China), and the Oxford Tropical Research Ethics Committee at the University of Oxford (UK). All participants provided written informed consent prior to data collection. (2) BGI low coverage WGS (BGI-lcWGS): A cohort of 69,788 pregnant women from Wuhan Children’s Hospital was genotyped using the BGISEQ-500 WGS with ultra-low sequencing depth during their non-invasive prenatal testing
56,57. Variants were imputed to 1KGP3 using STITCH. The BGI-lcWGS study received approval from the Bioethics and Biosafety of BGI (BGI-IRB 21088-T2). Each participant provided informed consent before enrollment. (3) WeGene cohorts (WGEG, WGEA, WGEC): Three cohorts comprising 15,126, 12,389, and 4,078 individuals recruited through the genetic testing service
58 from WeGene, Shenzhen Zaozhidao Technology Co., Ltd. were genotyped using Illumina Global Screening, customized Affymetrix WeGene V1, and Illumina Chinese Genotyping arrays, respectively. Variants were imputed via the China Metabolic Analytics Project (ChinaMAP) online server
59. All WeGene cohorts were approved by the Ethical Committee of Shenzhen WeGene Clinical Laboratory. Informed consent for online research was obtained from all participants. (4) 23mofang: 9,155 samples recruited by the genetic testing service from 23mofang, Inc. were genotyped using customized Axiom Precision Medicine or Illumina Infinium arrays. Variants were phased with SHAPEIT4 and imputed to 1KGP3 using Minimac4. The 23mofang cohort was reviewed and approved by Ethical and Independent Review Services, a private institutional review board. All participants granted their informed consent and responded to online surveys according to the approved 23andMe human subjects protocol. (5) Westlake Gut (WeGut) consortium cohorts: including 4,254 participants from 4 cohorts, including Guangzhou Nutrition and Health study, Zhoushan Metabolic Syndrome Cohort, Westlake Precision Birth Cohort, and Tongji-Shuangliu Birth Cohort
60, which were genotyped using the Illumina Global Screening array. Variants were imputed on the ChinaMAP online server. The WeGut Project protocol was approved by the Ethics Committee of Westlake University (20220429ZJS001), and all participants provided written informed consent. We also incorporated the EA GWAS meta-analysis of six Japanese cohorts: (1) Cohort 1: 6,956 individuals recruited by the National Center for Geriatrics and Gerontology were genotyped using the Illumina Asian Screening array. Variants were phased with Eagle v2.4.1 and imputed to the JEWEL 3k and 1KGP3 reference panels with Minimac4. This study was approved by the Human Research Ethics Committee of the National Center for Geriatrics and Gerontology (approval No. 990-16). (2) Cohort 2−5: Four cohorts with 71,045 individuals recruited through the Tohoku Medical Megabank Organization (ToMMo) community-based cohort study were genotyped using a customized Affymetrix Axiom array. Variants were phased with SHAPEIT2 and imputed to the cohort-specific reference panel using IMPUTE2, comprising 3,552 samples. These studies were approved by the Ethics Committee of RIKEN (RIKEN-Y-2022-002) and the ToMMo (2021-0032-1). (3) Cohort 6: 10,069 individuals recruited through the Nagahama Prospective Genome Cohort for Comprehensive Human Bioscience (the Nagahama Study) by Kyoto University were genotyped by Illumina Asian Screening or Japanese Screening arrays. Variants were phased with SHAPEIT4 and imputed to the TOPMed reference panel Version r2 on the online server
61. This project is a collaboration between Kyoto University and Nagahama City
62, and this study was approved by the Ethics Committee of Kyoto University (Identifier: G1289). Written informed consent was obtained from all participants in each Japanese cohort. Further details on cohort-level genotyping, phasing, imputation, and association analyses can be found in Supplementary Table S1.
We incorporated the EA GWAS data from our study with the publicly available EAS EA GWAS summary statistics from Chen
et al. for the within-ancestry meta-analysis, which included 165,232 individuals of EAS ancestry from the TWB and KoGES
9. To ensure a compatible scale for phenotype and variant effect sizes across all cohorts, we performed
z-score standardization of the summary statistics from Chen
et al. Furthermore, a publicly available EUR GWAS summary statistics dataset comprising 766,345 individuals of EUR ancestry from the Lee
et al. study (EA3)
6 was incorporated into the cross-ancestry meta-analysis. Since the available summary statistics from Lee
et al. had already been standardized, the variant effect sizes were on the same scale as our data, and we performed the meta-analysis without further adjustment.
Meta-analysis
We conducted cohort-level quality control using the EasyQC R package
63, excluding variants with low imputation accuracy or invalid estimates. For each GWAS, we estimated SNP-based heritability and potential inflation using LDSC
10 (Supplementary Table S2). The BGI-lcWGS cohort (
) and the Japanese meta-analysis (JPN-meta;
) showed potential inflation from population stratification (
). To address this, we adjusted the standard errors of variant effects for these cohorts by the square root of the LDSC intercept (
for BGI-lcWGS and
for JPN-meta).
We performed an inverse-variance-weighted (IVW) fixed-effects meta-analysis of nine quality-controlled GWAS results using OmicS-data-based Complex trait Analysis (OSCA
11). We used OSCA because it performs allele checking before the meta-analysis, provides per-SNP sample sizes, and produces output in a format directly usable for our downstream analyses. We performed a sensitivity analysis using METAL
64 (the software used by previous studies) and validated the consistency of results between the two approaches using the Pearson’s correlations of estimated effect sizes and association
P-values for the 723 genome-wide significant variants (
; Supplementary Fig. S13). We also estimated the SNP-level heterogeneity using METAL. We found that 4.11% of variants had a
P-value less than 0.05 in Cochran’s Q test, and 6.46% of variants exhibited an
larger than 50%, with a mean
of around 10.5%, suggesting that variant-level heterogeneity across cohorts is low and that the fixed-effect model is compatible with our EAS GWAS meta-analysis of EA. The meta-analysis of EAS GWAS data showed no evident inflation. We calculated the approximate number of weeks of schooling per allele by multiplying the standardized effect sizes by the combined standard deviation of
EduYears (4.19 years) across all EAS cohorts. We used consistent thresholds, as in previous EA studies, to identify independent genetic associations. We used an LD
cutoff of 0.1 and a distance cutoff of 10 Mb to perform the LD-based clumping. The 10-Mb distance was set based on the same assumption used in GCTA-COJO, which assumes that variants more than 10 Mb away from each other are in complete linkage equilibrium. We also leveraged the GCTA-COJO model selection to identify conditionally independent signals and potential secondary signals. Using the
1 Mb distance from the clumping lead variants as fine-mapping windows, we performed statistical fine-mapping analyses using CARMA
13 on our EAS meta-analysis data. The genotypes from EAS samples in the 1KGP3
65 were used as an out-of-sample LD reference for all the analyses above. We compared the genetic associations identified in our EAS GWAS to those from the previous EAS GWAS by Chen
et al. using two approaches. First, we reported the clumping-based lead variants located outside the previously identified genomic loci. For this analysis, we mapped the physical positions of our 33 clumping-based lead variants to the seven genomic regions defined in Table 1 of Chen
et al.’s paper. Second, we used --cojo-slct in GCTA to select, from the variants reported in Supplementary Table S3 of Chen
et al., those that were jointly associated with EA in our EAS GWAS (using the 1KGP3 EAS samples as the LD reference), and then performed conditional analysis (--cojo-cond) with these variants as covariates. We reported the variants that remained genome-wide significant after conditioning, thereby identifying EA-associated variants statistically independent of the previously reported genome-wide significant associations in EAS. As a sensitivity analysis, we additionally performed Multi-Ancestry Meta-Analysis (MAMA
66) using our EAS meta-analysis and the EUR meta-analysis from Lee
et al. Compared with Chen
et al., we also identified more independent loci in the EAS-specific meta-analysis using our data (Supplementary Table S42).
We used Popcorn
67 to analyze the genetic correlation of EA between our EAS GWAS and the GWAS summary statistics from Lee
et al. (EA3). An IVW cross-ancestry meta-analysis was then performed using OSCA to combine these two datasets. We used the genotypes of EUR samples from 1KGP3 as the LD reference for clumping the cross-ancestry meta-analysis. To ensure this did not introduce bias due to LD mismatch between the reference panel and the GWAS, we evaluated the LDSC intercept and attenuation ratio for the meta-analysis summary statistics using the EUR reference. The LDSC intercept was close to 1, and the attenuation ratio was less than 10% (LDSC intercept = 1.0721, s.e. = 0.0147; attenuation ratio = 0.0397, s.e. = 0.0081), indicating that using the EUR reference did not cause significant bias due to LD misspecification, consistent with previous studies
68,69. Moreover, we performed sensitivity analyses leveraging a mixed EAS and EUR LD reference with a sample ratio consistent with our cross-ancestry meta-analysis. The resulting LDSC intercept and attenuation ratio values were biased upward (LDSC intercept = 1.1317, s.e. = 0.0164; attenuation ratio = 0.0723, s.e. = 0.0090), potentially due to both LD structure and differences in variant inclusion (Supplementary Note 3 and Table S48). We further evaluated LD-based clumping using the mixed LD reference. While all lead variants identified by the mixed reference were successfully captured by the EUR LD reference, a substantial number of additional lead variants were exclusively identified by the EUR LD reference. This discrepancy occurred because numerous variants present in our meta-analysis were absent from the mixed EAS and EUR LD reference. Given these practical advantages, we utilized the 1KGP3 EUR samples as the primary LD reference for cross-ancestry analyses. Potential novel signals were defined as the clumping lead variants of the meta-analysis that were located more than 500 kb away from, and exhibiting LD
below 0.1 in both EAS and EUR with, any previously reported clumping lead variant from Okbay
et al. (EA4)
7. LD
values were calculated using the EAS and EUR samples from 1KGP3.
Polygenic prediction
The EA PRS was evaluated in a holdout cohort of 1,615 individuals with self-reported EAS ancestry from the UKB
15 who passed quality control (excluding F-coefficient outliers, related samples, and samples without phenotypes). We used our EAS GWAS and the EUR GWAS from Lee
et al. (EA3) as discovery GWAS, and the 1KGP3 EAS and EUR samples as LD references (formatted on the PRS-CS
70 website) to estimate the ancestry-specific posterior effect sizes of variants by PRS-CS and PRS-CSx
71. We used the default values for the optional parameters in both software packages. PLINK2
72 was used to construct the PRS for each individual. We regressed the test cohort phenotypes on birth year, sex, the interaction between birth year and sex, and the first 10 PCs as the control model, and evaluated prediction accuracy through the difference in
between the regressions with and without the EA PRS (incremental
). To determine the optimal ancestry-specific weights for building the cross-ancestry EA PRS via PRS-CSx, we employed a tuning approach based on 1,000 resampling iterations. For each iteration, we randomly partitioned the holdout cohort into a validation set (
) for weight tuning and a test set (
) for performance evaluation. We estimated the ancestry-specific posterior effect sizes and constructed the ancestry-specific EA PRS for each set. Then, we regressed the validation samples’ phenotypes on their ancestry-specific PRS to obtain the regression coefficients for each ancestry. Finally, we used the linear combination of the ancestry-specific PRS from the test set as the cross-ancestry PRS, with the regression coefficients from the validation set as weights. The predictive accuracy was evaluated by the mean incremental
across all 1,000 iterations to ensure robustness against sampling variance. Considering that the sample size of the hold-out UKB EAS cohort was small, we performed the following sensitivity analyses. First, we employed the same cross-ancestry EA PRS construction approach as in the previous EAS study, performing an inverse-variance-weighted meta-analysis of the posterior effect sizes across ancestries (using --meta in PRS-CSx). This approach bypasses the need for an independent tuning set and evaluates the prediction accuracy in the full UKB EAS cohort (
). Second, we employed two EAS cohorts, 23mofang (
) and WeGut (
), in our study for leave-one-cohort-out PRS analyses. Third, we switched the LD reference for PRS construction to EAS and EUR samples from UKB (provided by the PRS-CSx authors), which have much larger sample sizes (
;
) than the 1KGP3 reference, but which may cause over-fitting since our test cohort is from the same population (Supplementary Note 4 and Table S37).
Cross-ancestry conditional analysis pipeline
To identify associations independent of the largest EUR EA GWAS from Okbay et al. (EA4), we developed a rigorous conditional analysis pipeline, addressing the missingness of lead variants and collinearity when modeled in EAS LD.
(1) Variant set expansion. Due to the unavailability of the full summary statistics of EA4, we manually expanded their 3,952 lead variants by including LD proxies () of these variants using a LD reference panel comprising 50,000 randomly sampled individuals of EUR ancestry from the UKB. For lead variants missing from the reference panel, we took a conservative approach and included all variants located less than 1 Mb of the lead variant. This step generated a dense variant list encompassing all genetic associations identified in EUR.
(2) Covariate variant selection. To avoid collinearity in conditional analysis, we selected independent variants within EAS LD patterns that were jointly significant in both EAS and EUR GWAS using the GCTA-COJO model selection with LD reference from 1KGP3 EAS samples.
(3) Conditional analysis. We finally performed conditional analysis on the EAS GWAS by adjusting for the covariate variants in step (2), ensuring that the remaining signals were independent of the EUR-reported associations.
Cross-ancestry fine-mapping
We leveraged SuSiEx to compare the cross-ancestry and within-ancestry fine-mapping results
73 at loci with genome-wide significant associations in both the EAS and EUR GWAS of EA. We incorporated our EAS GWAS, the EUR GWAS from Lee
et al., and the 1KGP3 LD references of EAS and EUR samples into this analysis. Loci used for analysis were first defined as 1-Mb genomic regions around the 33 LD-based clumping lead variants from our EAS GWAS. Then, only the 15 loci with genome-wide significant associations in the EUR GWAS from Lee
et al. were included. We set the credible set coverage level to 95%, the maximum number of causal signals to 10, and the maximum number of iterations to 1,000 when applying SuSiEx. Other options were kept at their default values. We compared the number of credible sets, the median credible set size, and the maximum SNP PIP within each credible set across cross-ancestry, EUR-only, and EAS-only fine-mapping at the 15 common loci.
Gene prioritization
We incorporated six gene prioritization methods, four epigenetic annotations, and four layers of EAS xQTL datasets to identify EA-associated genes. For all methods, we included only protein-coding genes with significant results after multiple testing correction at GWAS-significant loci. We used MAGMA
17 and mBAT-combo
18 for gene-based analyses, and genes with Bonferroni-corrected
P were prioritized. For V2G mapping, we mapped genome-wide significant variants with CARMA PIP
to their linked genes using the following annotations: the transcription start site (TSS) and exons of each protein-coding gene (using the annotation from GENCODE
74 v40), promoter capture Hi-C (PCHiC) annotations of two published studies
75,76, activity-by-contact (ABC) scores of two published studies
77,78, and epigenomics annotations from RoadMap
79 and EpiMap
80 consortia. For coding variants, we directly mapped them to their targeted genes. For TSS annotation, we mapped each variant to its closest TSS gene. For other annotations, we selected the gene exhibiting the strongest interaction with the enhancer in which each variant resided. We used two integrative analysis methods (SMR
19 and COLOC
20) and four layers of omics data (mQTL, sQTL, eQTL, and pQTL) to integrate GWAS and xQTL data. We incorporated 54 xQTL datasets in this analysis, including eQTL datasets of 28 immune cell types and their meta-analysis
81, eQTL datasets of whole blood and brain tissue
82,83, sQTL datasets of 19 immune cell types and their meta-analysis
84, a serum pQTL dataset
85, a plasma pQTL dataset
82, and the meta-analysis of four blood mQTL datasets
86,87. Most of the xQTL datasets were collected from pure EAS samples. Although the Asian Immune Diversity Atlas (AIDA) sQTL dataset includes East Asians, South Asians (SAS), and Southeast Asians (SEA), 74% of the samples are of EAS ancestry, and the association signals were mainly driven by EAS samples. The potential impact of using the AIDA sQTL dataset is a moderate loss of power due to AF differences. No splice sites or genes were identified through this dataset in combination with our EAS GWAS. Genes identified by any xQTL dataset with a Bonferroni corrected SMR
P or SMR-multi
P , as well as a HEIDI test
P , were prioritized by SMR. Genes with a COLOC PP4
were prioritized by COLOC. We also used PoPS
21, a network-based method, for gene prioritization. The gene with the highest PoP score at each genome-wide significant locus was prioritized. Finally, we counted the number of methods by which each gene was prioritized, ranked them in descending order, and selected the genes prioritized by at least three independent methods as high-confidence genes associated with EA.
Genetic correlation and MR
We assessed genetic correlations between EA and 259 publicly available GWAS summary statistics of other complex traits from EAS using LDSC. We included traits from consortium GWAS papers using EAS samples, as well as traits collected from East Asian biobanks, including Biobank Japan (BBJ), the KoGES, the TWB, and the Taiwan Precision Medicine Initiative (TPMI). For traits shared among these biobanks, we first conducted a meta-analysis to increase statistical power and then performed the genetic correlation analysis. For traits shared between TWB and TPMI, we used only the TPMI GWAS due to its larger sample size. All data resources are listed in Supplementary Table S18. For traits showing significant genetic correlation with EA after FDR correction, we performed comparison analyses using the corresponding EUR data and EA3. The data sources were listed in Supplementary Table S18. We conducted bi-directional generalized summary-data-based Mendelian randomization (GSMR
33) to evaluate causal inference between EA and the complex traits significantly correlated with EA, with FDR correction to control for false-positive discoveries. Furthermore, we leveraged two approaches to conduct sensitivity analyses, accounting for potential sample overlap between our EAS GWAS and the complex-trait GWAS from TWB, KoGES, or TPMI, and between the EUR GWAS from Lee
et al. and the complex-trait GWAS datasets that included UKB samples. We first performed GSMR using complex-trait GWAS from independent datasets without overlapping samples with the EA GWAS. We used BBJ datasets for the analyses in EAS, and EUR datasets from the Million Veteran Program (MVP) or FinnGen for the analyses in EUR. Second, we used the MRlap R package, which performs IVW-MR and adjusted the results using the bivariate LD score regression intercept. All results were corrected for multiple testing using FDR. We also applied 10 commonly used MR methods to ensure the robustness of our findings using a previously developed pipeline
88, including IVW, Robust, weighted median, simple median, mode, MR-Egger, and contamination mixture method implemented in the MendelianRandomization package (v0.4.2,
CRAN.R-project.org/package=MendelianRandomization), MR-PRESSO (v1.0,
github.com/rondolab/MR-PRESSO), MRMix (v0.1.0,
github.com/gqi/MRMix), and RAPS (v0.2,
github.com/qingyuanzhao/mr.raps) package. Since the GSMR software specifically accounted for LD by leveraging a generalized least squares (GLS) regression model if the instrument variants are not fully independent, we retained its default LD threshold of 0.05. For MRlap and the other MR methods we used for triangulation, we used a stringent LD threshold of
across instrument variants. We used the default values for all the other optional parameters.
Heritability enrichment across tissues
We used S-LDSC for EA heritability enrichment analyses in functional genomic regions incorporating the baseline annotation (v2.2). We also used the baseline annotation to annotate the function of the 88 genome-wide significant variants after adjusting for EUR signals. We used LDSC-SEG
35 and annotations of tissue-specific open chromatin regions to assess enrichment of heritability for EA across tissues. Significance in multiple testing was determined using FDR correction.
Mapping EA-associated cells through brain single-cell and spatial transcriptomics data
To identify cell populations associated with EA, we integrated human brain single-nucleus transcriptomics data from ROSMAP
36 (
) with the multi-ancestry meta-analysis results using the single-cell disease relevance score (scDRS
37). The cell-type-level association was calculated by comparing the top 5% quantile trait score of each cell type to the top 5% quantile control scores of all cell populations. We identified cell types with a significant Monte Carlo
P-value after FDR correction as EA-associated. We also performed S-LDSC to ensure the robustness of scDRS findings using the established pipeline from previous studies
89,90. We computed gene expression specificity for each gene in each cell type as the expression level of a gene within a specific cell type divided by its total expression level across all cell types. Then, we ranked the proportional expression of all genes in each cell type and used the gene expression specificity of the top 10% genes as the annotation for that cell type. Finally, we used the baseline LDSC annotation as background and calculated the heritability enrichment for each cell type. We used the EUR samples in 1KGP3 as the reference panel.
For spatial mapping of EA-associated cells, we integrated mouse brain Stereo-seq data
39 and the cross-ancestry meta-analysis via gsMap
40, using the EUR samples in 1KGP3 as the reference panel and transforming mouse genes into human homologs. Cauchy’s combination of the single-cell association
P-values was employed to evaluate the cell-type association significance with EA across brain regions. We also used single-nucleus data from ten human brain regions (
)
41 to perform cell-type association analyses across brain regions using scDRS for validation. FDR correction was applied to avoid false positive results.
DATA AVAILABILITY
CODE AVAILABILITY
The following software and packages were used for data analyses: EasyQC (v23.8,
uni-regensburg.de/medizin/institut-fuer-epidemiologie-und-praeventivmedizin/genetische-epidemiologie/software); PLINK 1.9 (
cog-genomics.org/plink/1.9/); PLINK 2.0 (
cog-genomics.org/plink/2.0/); OSCA (v0.46.1,
yanglab.westlake.edu.cn/software/osca/#Download); METAL (
github.com/statgen/METAL); LDSC (v1.0.1,
github.com/bulik/ldsc); CARMA (
github.com/ZikunY/CARMA/tree/master); GCTA (v1.94.1,
yanglab.westlake.edu.cn/software/gcta/#Download); bedtools (v2.30.0,
github.com/arq5x/bedtools2); mama (
github.com/JonJala/mama); PRS-CS (
github.com/getian107/PRScs); PRS-CSx (
github.com/getian107/PRScsx); MAGMA (v1.10,
cncr.nl/research/magma/); SMR (v1.03,
yanglab.westlake.edu.cn/software/smr/#Download); COLOC (
github.com/chr1swallace/coloc); PoPS (v0.2,
github.com/FinucaneLab/pops/tree/master); MRlap (v0.0.3.3,
github.com/n-mounier/MRlap); LAVA (v0.1.5,
github.com/josefin-werme/LAVA); scDRS (v1.0.2,
github.com/martinjzhang/scDRS); gsMap (
github.com/JianYang-Lab/gsMap); and SuSiEx (
github.com/getian107/SuSiEx). The custom code for the core analyses presented in this manuscript is publicly available at:
github.com/RL-m/EAS_EA_GWAS_Ma_et_al.
The Author(s) 2026. Published by Higher Education Press. This is an Open Access article distributed under the terms of the CC BY license (https://creativecommons.org/licenses/by/4.0/).