Genome‐wide meta‐analysis identifies ancestry‐specific loci for Alzheimer's disease

Abstract

INTRODUCTION

Alzheimer's disease (AD) is a devastating neurological disease with complex genetic etiology. Yet most known loci have only identified from the late‐onset type AD in populations of European ancestry.

METHODS

We performed a two‐stage genome‐wide association study (GWAS) of AD totaling 6878 Chinese and 63,926 European individuals.

RESULTS

In addition to the apolipoprotein E (APOE) locus, our GWAS of two independent Chinese samples uncovered three novel AD susceptibility loci (KIAA2013, SLC52A3, and TCN2) and a novel ancestry‐specific variant within EGFR (rs1815157). More replicated variants were observed in the Chinese (31%) than in the European samples (15%). In combining genome‐wide associations and functional annotations, EGFR and TCN2 were prioritized as two of the most biologically significant genes. Phenome‐wide Mendelian randomization suggests that high mean corpuscular hemoglobin concentration might protect against AD.

DISCUSSION

The current study reveals novel AD susceptibility loci, emphasizes the importance of diverse populations in AD genetic research, and advances our understanding of disease etiology.

Highlights

• Loci KIAA2013, SLC52A3, and TCN2 were associated with Alzheimer's disease (AD) in Chinese populations.

• rs1815157 within the EGFR locus was associated with AD in Chinese populations.

• The genetic architecture of AD varied between Chinese and European populations.

• EGFR and TCN2 were prioritized as two of the most biologically significant genes.

• High mean corpuscular hemoglobin concentrations might have protective effects against AD.

1. BACKGROUND

Alzheimer's disease (AD), the most common type of dementia, currently affects more than 40 million people worldwide, and the number is projected to reach ≈139 million by 2050. ¹ The etiology of AD is complicated, with genetic factors playing critical roles. The heritability of late‐onset AD (LOAD) is estimated to be ≈60% to 80% according to family studies ² ; however, recent large‐scale genome‐wide association studies (GWASs) have reported much lower heritability estimates, ranging from 3% to 7%. ³ , ⁴ , ⁵ The same problem exists in early‐onset AD (EOAD), with the reported heritability estimated to be exceeding 90%, ⁶ whereas pathogenic mutations in the three famous causal genes—amyloid precursor protein (APP), presenilin 1 (PSEN1), and presenilin 2 (PSEN2)—account for only 5%≈10% of EOAD cases. ⁷ , ⁸ The missing heritability suggests that there is still much to uncover about the genetic basis of both EOAD and LOAD. Further discovery of the genetic determinates of AD is expected to enhance our knowledge of AD etiology and pave the way for developing new therapeutic strategies. ⁹ Moreover, understanding the genetic landscape of different clinical subtypes could help enable personalized medicine and healthcare. Nonetheless, no studies to date have systematically compared the genetic architecture of EOAD and LOAD.

Another significant gap in AD genetics research is that previous findings have been based mainly on individuals of European ancestry. ¹ The racial disparities have limited the generalizability of existing findings from a specific population to others. From the perspective of susceptibility to a certain gene, the genotype frequency and effect size were observed to vary significantly among cohorts from different ancestries. ¹⁰ , ¹¹ , ¹² For example, the R47H mutation in the triggering receptor expressed on myeloid cells 2 (TREM2) gene was strongly associated with LOAD in Europeans, but its minor allele frequency (MAF) is extremely low in Chinese individuals and thus might not play a major role in LOAD susceptibility in this population. ¹³ The predictive performance of polygenic risk score (PRS) also showed noticeable heterogeneity across multiple ancestries. ¹⁴ Hence, there is an urgent need to conduct cross‐ancestry comparisons of the AD genetic architecture and identify more AD susceptibility loci in non‐European populations. However, despite the emergence of non‐European genetic studies, their findings are inconsistent, and the reported novel loci are rarely replicable. ¹ This inconsistency can be attributed to varied demographics, non‐standard quality control procedures, insufficient sample sizes, and low statistical power.

To address existing gaps in the literature and enhance our understanding of AD etiology, we conducted a two‐stage AD GWAS involving 6878 Chinese and 63,926 European individuals. Our first objective was to compare the genetic basis of EOAD and LOAD. Thus, we began by performing GWASs for EOAD and LOAD using Chinese samples during the discovery stage. Then we presented a comprehensive genetic comparison of EOAD and LOAD. Our second goal was to identify additional AD susceptibility loci and investigate the potential differences for the identified significant loci between Chinese and European populations. To this end, we meta‐analyzed EOAD and LOAD due to the observed overlap in their genetic architecture of common variants. Next, we carried out replication and meta‐analysis in independent Chinese (Stage 2A) and European (Stage 2B) samples, facilitating a comparison of the results across diverse populations. Finally, we aimed to identify the most biologically significant AD‐associated genes and provide new insights into AD etiology. To accomplish this, we conducted functional follow‐up analyses, prioritized candidate AD genes, assessed the functions and druggability of the prioritized genes, and explored potential causal factors through phenome‐wide Mendelian randomization (phe‐MR) in Asian populations. The study workflow is illustrated in Figure 1.

FIGURE 1.

Study workflow. Top left, we performed a two‐stage GWAS covering 70,804 participants. In Stage 1, GWASs of EOAD and LOAD in the discovery sample were performed and subsequently meta‐analyzed due to their observed genetic overlap. In Stage 2, replication and meta‐analyses were conducted among all variants that passed the suggestive threshold (p < 1 × 10⁻⁵) in Stage 1 in independent Chinese and European samples, respectively. Top right, we conducted validation and genetic overlap analyses to compare the genetic architecture of EOAD and LOAD. Bottom left, ancestry‐specific AD loci were identified by replication analyses in Chinese and European individuals. We also investigated the allele frequencies of top‐associated variants in different populations. Bottom right, to gain more insights into AD etiology, we further performed functional exploration analyses, developed a scoring system to prioritize genes, and conducted a phenome‐wide MR to identify potential causal factors on AD. EOAD, early‐onset Alzheimer's disease; GWAS, genome‐wide association study; LOAD, late‐onset Alzheimer's disease; MR, Mendelian randomization; SNP, single nucleotide polymorphism.

2. METHODS

2.1. Samples

The Stage 1 discovery sample consisted of 1369 AD cases and 1639 age‐matched cognitively normal controls with Chinese Han ancestry, recruited from the Memory Clinic of the Huashan Hospital of Fudan University, Qingdao Municipal Hospital, and Daping Hospital. All participants underwent a medical history and neuropsychological evaluation. Structural imaging, such as magnetic resonance imaging (MRI) and biomarker assessments, including positron emission tomography (PET) and cerebrospinal fluid (CSF) analysis, were performed when necessary. Cases were diagnosed with probable AD by experienced neurologists following the diagnostic criteria of the National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer's Disease and Related Disorders Association (NINCDS‐ADRDA) ¹⁵ or the National Institute on Aging Alzheimer's Association workgroups (NIA‐AA). ¹⁶ , ¹⁷ Among the cases, 671 were categorized as EOAD with an age at onset (AAO) of younger than 65, and 698 were classified as LOAD with an AAO of 65 or older. Controls were removed if they were younger than 40; exhibited impaired cognition as confirmed by objective tests like the Mini‐Mental State Examination (MMSE); or had evidence of conditions such as stroke, epilepsy, encephalitis, major depressive disorder, or schizophrenia, which could impact cognition.

Stage 2A replication samples were collected from Xuanwu Hospital of Capital Medical University. A total of 2358 LOAD cases and 3116 controls with Chinese ancestry were recruited. Eligible AD cases were diagnosed following the NINCDS‐ADRDA or NIA‐AA criteria, ¹⁵ , ¹⁶ , ¹⁷ with an AAO of 60 or older and no family history of dementia. Qualified controls were individuals 60 years of age or older, with confirmed normal cognition through the MMSE score and Clinical Dementia Rating (CDR) Scale score, without subjective memory complaints or any evidence of diseases that could affect cognition. For the replication stage of European ancestry (Stage 2B), we obtained the GWAS summary statistics conducted by Kunkle et al., which include data from 21,982 clinically diagnosed AD cases and 41,944 controls from the International Genomics of Alzheimer's Project (IGAP). ⁴

2.2. Genotyping and quality control

All participants in Stage 1 were genotyped by the Infinium Asian Screening Array Kit on the Illumina BEADLAB platform (Illumina). We carried out standard quality control procedures before and after imputation using PLINK version 1.9. ¹⁸ , ¹⁹ Variants with call rate <0.95, Hardy‐Weinberg p‐value <1×10⁻⁶, or MAF < 0.01 were excluded before imputation. Individuals with a genotyping call rate < 0.95 or inbreeding coefficient > 0.1 were also removed. Haplotype phasing was performed with Beagle version 3.3.2, and imputation was carried out using IMPUTE5 software, with 1000G phase 3 CHB/CHS Asian samples as reference panel. ²⁰ , ²¹ , ²² The 1000G was selected as the reference panel for imputation because it is the most widely used panel in Asian GWAS studies and shows reliable performance in genotype imputation. ²² After imputation, variants were filtered by call rate > 0.99, MAF > 0.01, Hardy‐Weinberg p‐value > 1×10⁻⁶, and IMPUTE INFO score > 0.7. We also removed samples with self‐reported and genetic‐inferred sex mismatch, relatedness (PiHat > 0.1875), heterozygosity rate outliers (>3SD), and ethnic outliers via principal component (PC) analysis merged with 1000 Genomes data. These quality control criteria resulted in 2809 individuals and 7,676,018 autosomal variants for analysis. The genotyping, quality control, and imputation procedures of replication samples are described in the Supplementary Methods.

Research in context

1. Systematic review: We reviewed the relevant literature using the PubMed database. To date, most known susceptibility loci for Alzheimer's disease (AD) have been identified only from the late‐onset type of AD in populations of European ancestry. The sampling bias has limited the generalizability of existing findings to other populations.

2. Interpretation: We identified three novel AD susceptibility loci (KIAA2013, SLC52A3, and TCN2) and a novel variant within EGFR (rs1815157) through genome‐wide meta‐analysis in the Chinese population. Replication analyses revealed discrepancies in allele frequencies and genetic effects between Chinese and European samples. We prioritized EGFR and TCN2 as two of the most biologically significant genes through a ranking strategy.

3. Future directions: The novel loci and variants identified in this study provide valuable insights into AD etiology, warranting further experimental investigation. It is imperative to conduct larger studies encompassing all AD clinical subtypes across diverse populations to refine the AD genetic landscape.

2.3. GWAS in the discovery stage

We conducted GWASs for EOAD and LOAD using the fastGWA‐GLMM model, as implemented in GCTA version 1.94.0. ²³ To account for potential confounders, we regressed age (AAO for cases and age at last examination for controls), sex, and the first PC because only the first PC was significant (P < 0.05) in the twstats analysis. ²⁴ Subsequently, we carried out a standard error–weighted meta‐analysis for EOAD and LOAD subtypes with genomic control through the METAL software. ²⁵ A sensitivity analysis combining EOAD and LOAD individuals was also performed. The genomic inflation factor was estimated using the R package “QCEWAS.” ²⁶ We also conducted the gene‐based association analyses in GCTA with GWAS summary statistics of EOAD, LOAD, and Stage 1 meta‐analysis as input. ²⁷ Genes that passed the Bonferroni adjusted p‐value threshold (0.05/18866 =  2.65 × 10⁻⁶) in the mBAT‐combo test were considered significant.

2.4. Genetic comparison between EOAD and LOAD

We tested if the previously reported loci could be validated in EOAD and LOAD using GWAS results in Stage 1. The validation analysis was based on five recently published large‐scale AD GWASs, ³ , ⁴ , ⁵ , ²⁸ , ²⁹ with sample sizes ranging from 94,437 to 788,989. We identified 86 non‐overlapping loci from these GWASs, and the lead variants of 67 non‐MHC (major histocompatibility complex) loci could be found in our genetic data. We assessed whether the previously reported loci could be validated nominally and further tested if they could be validated after false discovery rate (FDR) correction. We also systematically examined the shared and unique common variants and genes between EOAD and LOAD at different significance levels (Supplementary Methods).

2.5. Replication and meta‐analyses

For all variants that passed the suggestive significant threshold in Stage 1 meta‐analysis (P < 1×10⁻⁵), replication was carried out in independent Chinese (Stage 2A) and European samples (Stage 2B). In Stage 2A, the genetic associations with AD status were tested using logistic regression under the additive model with PLINK 1.9 after adjusting for age (onset for case and last examination for control), sex, and the first two PCs. ¹⁹ For Stage 2B, we employed the meta‐analyzed summary statistic results of AD status conducted by Kunkle et al. ⁴ The original individual GWAS results constituting the meta‐analysis in Stage 2B were calculated using the additive model, with age, sex, and PCs included as covariates. ⁴

A variant was considered replicated if its effect estimates in the replication stage passed the Bonferroni correction and showed consistent effect directions compared to the discovery stage. After the above replication analyses, we conducted a meta‐analysis within Chinese samples (Stage 1 + 2A) to identify population‐specific AD susceptibility loci. Moreover, we performed a bi‐ancestry meta‐analysis, combining data from Chinese and European samples (Stage 1 + 2B) to discover additional genetic loci. The meta‐analyses of the discovery and replication stages were performed using METAL software. ²⁵ A gender‐stratified analysis of the novel loci was conducted to investigate the potential gender bias.

2.6. Identification and prioritization of candidate AD genes

Given that Stage 2A replicated a higher percentage of the original results than Stage 2B and identified novel loci following meta‐analyses, we proceeded to map genome‐wide significant loci based on the Stage 1 + 2A meta‐analyses results. This mapping was conducted by leveraging positional information (up to 10 kb apart), brain expression quantitative trait loci (eQTL) data, and chromatin interaction data from adult and fetal human brain samples. The position, QTL, and chromatin interaction mapping were conducted on the FUMA platform. ³⁰ We considered all genes identified by gene mapping or gene‐based analysis as candidate AD genes. We conducted enrichment analyses of cell types and pathways using all candidate AD genes to gain functional insights (Supplementary Methods).

In order to prioritize AD genes, we evaluated all candidate AD‐associated genes by differential expression in AD brains and controls using post‐mortem expression data via Agora. ³¹ This analysis was based on data from over 1100 individuals in the Religious Orders Study and Memory and Aging Project (ROSMAP), Mayo RNAseq (MAYO), and Mount Sinai Brain Bank (MSBB) across nine brain regions. Next, all candidate AD‐associated genes were assessed by their correlation with AD neuropathology. To this end, we inquired into two databases, the Alzdata and Agora. ³¹ , ³² The detailed descriptions of data and analytical models can be found in Supplementary Methods. Finally, we established a scoring system to prioritize candidate genes based on four categories: GWAS significance, gene‐based association, gene mapping, and gene expression. We assigned two or three points to each gene within each category, depending on their supporting evidence (Supplementary Methods).

2.7. Follow‐up of prioritized genes

Phenome‐wide association scanning was performed to explore the potential associations of the prioritized genes using the PhenoScanner database. ³³ , ³⁴ We inquired about all diseases and traits with genome‐wide significant associations with the target gene. Furthermore, we visualized the expression patterns of brain cell types associated with the prioritized genes using the AlzData database. ³⁵ In addition, we investigated the potential interactions of the prioritized genes in this study with established core AD genes (e.g., PSEN1, PSEN2, APP, and APOE) via the GeneMANIA database. ³⁶ The druggability of candidate genes was assessed by AMP‐AD and obtained from Synapse (Synapse ID syn13363443).

2.8. Phenome‐wide MR in Asian populations

For phenome‐wide MR analysis we extracted all available exposures in the BioBank Japan Project through the MR‐Base platform. ³⁷ , ³⁸ We selected genome‐wide significant (P < 5 × 10⁻⁸) and independent (r ² < 0.001) variants as instruments based on the 1000 Genomes Asian reference panel. The phenome‐wide MR analyses were conducted for 68 phenotypes with at least five instruments to reduce imprecise results. We employed four MR methods: Inverse Variance Weighting (IVW), MR‐Egger, Weighted Median (WMe), and Weighted Mode (WMo). The primary method was IVW, as recommended by guidelines, ³⁹ and results passing the FDR correction were considered significant. Heterogeneity was estimated by Q statistic in MR‐IVW, and pleiotropy was quantified by the intercept in MR‐Egger. We also performed a leave‐one‐variant‐out sensitivity analysis and an additional sensitivity analysis after excluding all instruments located within ± 250 kb of the apolipoprotein E (APOE) ɛ4 defining variant, rs429358, to assess the robustness of the original results. ⁴⁰

3. RESULTS

3.1. Cohort demographics

The discovery sample comprised 1286 clinically diagnosed AD cases and 1523 cognitively normal controls of Chinese ancestry. We divided the sample into two age groups, with similar sample sizes based on whether the AAO of cases or age at the last examination of controls was younger than 65. For the younger group, the mean (± SD) age was 56.4 ± 5.2 years in cases and 55.0 ± 6.7 years in controls. For the older group, the mean age was 72.9 ± 5.4 years in cases and 72.5 ± 6.2 years in controls.

Participants in the replication stages were of Chinese (Stage 2A) and European (Stage 2B) ancestry. All cases in replication stages were clinically diagnosed with LOAD (Stage 2A: mean AAO = 72.7 ± 9.6 years; Stage 2B: mean AAO = 72.9 ± 15.2 years). Further detailed demographics can be found in Table 1.

TABLE 1.

Demographics of participants in Stage 1 discovery, Stage 2A, and Stage 2B.

	Stage1 EOAD		Stage1 LOAD		Stage 2A replication		Stage 2B replication
	Case	Control	Case	Control	Case	Control	Case	Control
N	630	787	656	736	1595	2474	21982	41944
Age, years *	56.4 (± 5.2)	55.0 (± 6.7)	72.9 (± 5.4)	72.5 (± 6.2)	72.7 (± 9.6)	72.0 (± 13.1)	72.9 (± 15.2)	72.4 (± 12.7)
Female, %	61.0	49.0	57.9	41.3	58.0	56.7	61.3	57.1
APOE ε4 frequency, %	21.7	7.4	26.1	10.5	24.8	9.6	NA	NA
MMSE score	13.1 (± 6.5)	27.8 (± 2.5)	14.0 (± 6.4)	26.3 (± 2.8)	16.3 (± 6.8)	28.1 (± 1.3)	NA	NA

Notes: All statistics are mean (± SD) unless otherwise specified.

^*Age at onset for cases and age at the last examination for controls.

Abbreviations: APOE, apolipoprotein E; EOAD, early‐onset Alzheimer's disease; LOAD, late‐onset Alzheimer's disease; MMSE, Mini‐Mental State Examination.

3.2. Genetic overlap between EOAD and LOAD

We investigated the genetic overlap between EOAD and LOAD from two perspectives. First, using the established risk loci of LOAD from the five large European GWASs as a reference, we estimated whether these loci could be validated in both Chinese EOAD and LOAD. The validation of the common risk loci for EOAD was not performed due to the lack of published EOAD GWAS. Second, we conducted a comprehensive comparison of variants and genes between EOAD and LOAD in Chinese samples across various significance levels.

We conducted separate GWASs for Chinese EOAD and LOAD using data from the Stage 1 discovery sample (Tables S1, S2). It was found that 15% of the top variants of the known LOAD loci identified from European samples had at least nominal significance in Chinese LOAD (BIN1, CLNK, CD2AP, HS3ST5, ANK3, and APOE) and EOAD (NME8, PICALM, IQCK, KAT8, and APOE). The most significant signal was APOE rs429358, which showed similar associations with EOAD (odds ratio [OR] = 2.32, Pp = 3.57 × 10⁻¹⁵, p _adj  = 2.39 × 10⁻¹³) and LOAD (OR = 2.49, p = 4.34 × 10⁻²¹, p _adj  = 2.91 × 10⁻¹⁹). The second validated variant was BIN1 rs6733839. It was significantly associated with LOAD (OR = 1.31, p = 8.24 × 10⁻⁴, p _adj = 0.03) and was also associated with EOAD in the same direction but with less significance (OR = 1.15, p = 0.10). Of note, except for NME8, all of the nominally significant variants associated with Chinese EOAD or LOAD showed consistent directions with the original European LOAD GWAS, suggesting the remarkable genetic similarity of EOAD and LOAD. The different directions of NME8 observed in our study were consistent with a previous study, reporting that NME8 showed opposite effects in atypical and typical AD. ⁴¹

We then systematically explored the genetic overlap between EOAD and LOAD at different significance levels. At the genome‐wide significant threshold (p < 5 × 10⁻⁸), APOE emerged as the shared and only significant signal for EOAD and LOAD. We observed that the proportion of shared variants between EOAD and LOAD was generally high but gradually decreased when the p‐value threshold was relaxed. Specifically, the shared proportion of variants was 100% at p < 1 × 10⁻⁷, 92% at p < 5 × 10⁻⁷, 73% at p < 1 × 10⁻⁶, 38% at p < 5 × 10⁻⁶, and 26% at p < 1 × 10⁻⁵ (Table S3). The gene‐level analysis showed a similar trend, with the shared elements continuously decreasing as the p threshold was relaxed (Figure S1 and Table S4). To gain mechanistic insights, we further conducted gene‐set enrichment analysis and found the enriched ontologies of shared genes between EOAD and LOAD mainly involved in lipid metabolism (Table S5).

3.3. Genome‐wide meta‐analysis and replication identify new ancestry‐specific loci

Owing to the observed genetic overlap between EOAD and LOAD in the above analyses, we next combined EOAD and LOAD to enhance the discovery of AD susceptibility loci. We performed a meta‐analysis of EOAD and LOAD GWAS results in the discovery stage, and the genomic inflation factor after the meta‐analysis was modest (λ = 1.07). In addition to APOE, three novel loci (7q31.2, 12q23.3, and 22q12.2) reached the genome‐wide significance threshold in the Stage 1 meta‐analysis (Figure 2A). A sensitivity analysis combining EOAD and LOAD individuals yielded results similar to those of the meta‐analysis (Table S6). Of interest, we observed a striking difference in allele frequencies between Asian and European populations for the lead variants of these novel loci. According to the 1000G reference, rs147226119 in 12q23.3 and rs3804080 in 22q12.2 were common in East Asian populations (MAF = 0.01 and 0.05, respectively) but rare in European populations (MAF < 0.001). In addition, the MAF of rs757278 in 7q31.2 was 0.01 in 1000G East Asian populations but 0.35 in 1000G European populations.

FIGURE 2.

Manhattan plots of GWAS Stage 1 and Stage 1 + 2A. Manhattan plots display genome‐wide associations with AD in Stage 1 (A) and Stage 1 + 2A (B). The orange line denotes genome‐wide significance (P = 5 × 10⁻⁸), whereas the gray line represents the suggestive level (P = 1 × 10⁻⁵). The y‐axis is limited for better visualization of non‐APOE loci.

Next we conducted replication analyses in Chinese (Stage 2A) and European (Stage 2B) samples to assess the robustness of the results. It was found that 31% of all suggestive variants (P < 1 × 10⁻⁵) identified in Stage 1 could be validated in an independent Chinese sample after Bonferroni correction (Stage 2A, Table S7). Notably, the locus 22q12.2 (TCN2) was successfully replicated with a remarkably small p‐value (replication OR = 0.67, p  = 1.79 × 10⁻⁸), although the other two significant loci (i.e., 7q31.2 and 12q23.3) in Stage 1 were not replicated in Stage 2A. In contrast to Stage 2A, the replication in European samples (Stage 2B) failed to identify any Bonferroni‐corrected significant associations, with the exception of variants in the APOE region (Table S8).

Meta‐analysis of the results from the discovery samples (Stage 1) and the replication samples (Stage 2A or 2B) were subsequently performed. Through the Stage 1 + 2A meta‐analysis, we identified three new genome‐wide significant signals specifically for Chinese populations, including the loci 1p36.22, 7p11.2, and 20p13 (Figure 2B and Table 2). These three signals showed statistical significance in both Chinese cohorts (discovery p < 1 × 10⁻⁵ and replication p _adj < 0.05), and no gender bias was found in the stratified analysis (Figure S2). Loci 1p36.22 (KIAA2013) and 20p13 (SLC52A3) have not been reported previously in other GWAS, and the lead signal of EGFR, rs1815157 has been identified as a novel susceptibility variant for AD for the first time. In contrast with Stage 1 + 2A, the cross‐ancestry meta‐analysis of Stage 1 + 2B did not identify any additional genome‐wide significant signals. In addition to the different genetic associations, we again noted differences in allele frequencies between Chinese and European populations. For example, the rs75680863‐T (TCN2) frequency is 0.12 in our sample but only 0.001 in European populations (Table 2). These findings further support the hypothesis that the genetic determinants of AD vary among different populations.

TABLE 2.

Summary of results for identified loci reaching genome‐wide significance after Stages 1 and 2A.

							Stage 1 discovery				Stage 2A replication		Stage 1 + Stage 2A meta‐analysis
Locus	Lead SNP	Position	EA	OA	EAF CHN	EAF EUR *	EOAD P	LOAD P	Overall OR	Overall P	OR	P	OR	P
KIAA2013	rs6682554	1:11986621	T	C	0.45	0.46	1.83E‐03	5.56E‐04	0.76	5.99E‐06	0.81	3.90E‐06	0.79	8.65E‐10
EGFR	rs1815157	7:55237080	T	C	0.45	0.70	1.05E‐03	3.04E‐04	0.75	2.09E‐06	0.84	1.22E‐04	0.80	1.04E‐08
APOE	rs429358	19:45411941	C	T	0.17	0.16	3.57E‐15	4.34E‐21	2.41	6.60E‐33	3.17	1.15E‐41	2.69	1.83E‐67
SLC52A3	rs6117562	20:753310	A	G	0.40	0.26	5.55E‐04	1.91E‐03	1.30	6.48E‐06	1.28	1.16E‐07	1.29	1.82E‐11
TCN2	rs75680863	22:31007023	T	A	0.12	<0.01	2.62E‐05	6.91E‐04	0.63	1.69E‐07	0.67	1.79E‐08	0.66	1.38E‐14

Notes: This table displays the results of Stage 1 + 2A meta‐analysis, and the full results are available in Table S7. In contrast with Stage 1 + 2A, Stage 1 + 2B meta‐analysis did not yield any new significant locus other than APOE, and full results can be found in Table S8.

^*MAF in European is based on the 1000G European reference panel.

Abbreviations: CHN, Chinese; EA, effect allele; EUR, European; EAF, effect allele frequency; OA, other allele; OR, odds ratio; SNP, single nucleotide polymorphism.

3.4. Functional interpretation and prioritization of AD candidate genes

We next carried out positional mapping, brain eQTL mapping, and chromatin interaction mapping based on GWAS results from Stage 1 + 2A to identify new genes associated with AD. These three methods identified 79 unique candidate AD genes in total (Table S9). Furthermore, we performed gene‐based association tests, identifying 10 significant genes after Bonferroni correction (Table S10 and Figure S1). Collectively, gene mapping strategies and gene‐based association tests in Chinese samples yielded 83 unique candidate AD genes. Brain cell–type analysis revealed that these candidate AD genes exhibited differential expression in astrocytes, oligodendrocytes, and neurons when comparing AD cases with controls (Figure 3A). Furthermore, enrichment analyses of gene ontologies (GOs) for these candidate AD genes showed that the top enriched GO terms were related to lipid metabolism (Figure 3B). Of interest, these ontologies were consistent with those enriched from the shared genes between EOAD and LOAD (Table S5).

FIGURE 3.

Functional exploration of candidate genes and gene prioritization. (A) Enriched gene ontologies of all candidate AD genes. (B), Relative expression levels in brain cell types of all candidate AD genes. The p‐values represent the significance of differential brain cell type expression between AD and controls. (C) Prioritization of candidate AD genes. Only genes with a priority score ≥5 are shown in this figure, and the full results can be found in Table S13. †Gene expression in human post‐mortem dorsolateral prefrontal cortex is correlated with the distribution and severity of neurofibrillary tangles (BRAAK), neuritic plaque density (CERAD), or final consensus cognitive diagnosis (COGDX) in the Agora database. ‡Gene expression is correlated with AD pathology in Aβ/tau line AD mouse models according to the AlzData database. BP, biological process; CC, cellular component; eQTL, expression quantitative trait loci; GO, gene ontology; MF, molecular function.

Subsequently we conducted functional exploration analyses for all candidate AD‐associated genes. These analyses included examining differential expression in AD brains compared to healthy controls, as well as the correlation between gene expression and AD neuropathology in human and mouse brains. We observed that 23 genes had meaningful differential expression between AD and healthy controls after multiple testing adjustments (Table S11). In addition, the expression levels of 19 genes were correlated with amyloid beta (Aβ) or tau pathology in mouse models (Table S12). The expression levels of seven candidate genes also showed significant associations with AD hallmarks in human post‐mortem samples, including distribution and severity of neurofibrillary tangles (EGFR, APOC1, and PES1), neuritic plaque density (EGFR, APOC1, RNF215, and LIMK2), and consensus cognitive diagnosis at time of death (SEC14L6 and TCN2), although these genes except APOC1 have not been linked to CSF diagnostic biomarkers in vivo. ⁴² , ⁴³ Genes associated with AD hallmarks could be grouped into four clusters, characterized by their involvement in lipid metabolism, ribosomal ribonucleic acid, epidermal growth factor receptor (EGFR) binding, and mitogen‐activated protein kinase signaling (Figure S3).

Based on the above findings, we developed a priority score encompassing GWAS significance, gene‐based association, functional mapping, differential expression, and correlation with AD neuropathology. This score ranged from 0 to 10 points (Table S13). Notably, three genes—EGFR, APOC1, and TCN2—achieved the highest score and were, therefore, considered the most prioritized (Figure 3C).

EGFR encodes a receptor for epidermal growth factor, which participates in early brain development and brain atrophy. ⁴⁴ Variants within EGFR have been reported to be associated with AD and related dementias, ²⁹ as well as deaths due to amyloidosis and brain diseases in European studies (Table S14), although none of them were related to the lead variant in our Chinese AD GWAS, rs1815157. Furthermore, EGFR is expressed predominantly in astrocytes (Figure S4), and the gene was found to be overexpressed in multiple brain regions affected by AD (Figure 4). Furthermore, EGFR could interact with PSEN1, PSEN2, and APOE (Figure S5) and had favorable therapeutic potential and safety profiles, indicating that it is a promising drug target for AD. APOC1, although mapping to the APOE region, outscored APOE in our gene prioritization because APOC1 was significant in brain eQTL mapping, and its expression level was correlated significantly with AD neuropathology, such as the distribution and severity of neurofibrillary tangles. APOC1 has also been implicated in lipid metabolism traits (Table S15).

FIGURE 4.

Differential expression of prioritized genes between human AD brains and controls. Meaningful differential expression is considered to be a log2 fold‐change value > 0.263, or < −0.263 according to the Agora database.

TCN2, a novel gene identified in our Chinese meta‐analysis, encodes a protein belonging to the vitamin B12–binding protein family. The lead variant of TCN2 in our study, rs75680863, is a likely deleterious missense variant (Combined Annotation Dependent Depletion [CADD] score = 22.2) common in Chinese but rare in European populations. In addition, the eQTL in the dorsolateral prefrontal cortex, rs5997690, is in linkage disequilibrium (LD, r ²  = 0.59 in our sample) with rs75680863, and it is also nominally associated with AD in our sample (OR = 0.73, p = 4.48 × 10⁻⁵). TCN2 may be involved in vascular mechanisms as it is expressed in endothelial cells (Figure S4), and its variants have been reported to be associated with multiple cerebrovascular disorders (Table S16).

3.5. Phenome‐wide MR reveals a causal role of MCHC on EOAD

Finally, in order to provide causal insights into AD etiology in Asian populations, we performed phenome‐wide MR (or phe‐MR) using exposures from the BioBank Japan Project (Tables S17–S19). Only genetically determined mean corpuscular hemoglobin concentration (MCHC) showed a significant protective effect on EOAD after the multiple testing correction (IVW OR = 0.30, p _raw = 4.09 × 10⁻⁴, p _adj = 0.03). The effect direction was consistent in all sensitivity analyses, and no heterogeneity or pleiotropy was observed (Table S17). Moreover, the results remained stable in the leave‐one‐out analysis and were not biased by instruments within the APOE region (Figure S6).

4. DISCUSSION

The combination of genetic data from Chinese and European individuals enables us to provide novel insights into the genetic etiology of AD. We demonstrated genetic overlap between EOAD and LOAD. Through a meta‐analysis of two independent Chinese cohorts, we identified five genome‐wide significant loci associated with AD, three of which (KIAA2013, SLC52A3, and TCN2) were novel. We also identified a novel AD‐associated variant, rs1815157, exclusively through the meta‐analysis within Chinese samples. Its corresponding gene, EGFR, is considered a promising drug target for AD. TCN2 is another gene with the highest priority score and might be involved in vitamin B12 and vascular‐related mechanisms. TCN2 is an ancestry‐specific locus, with its lead variant common in Chinese but rare in European populations. The disparities in allele frequencies and association estimates between Chinese and European samples for these novel AD‐associated variants highlight racial differences. In addition, MR analysis also suggests that red blood cell indices might contribute to AD etiology.

About half of our cases in the discovery sample developed AD before age 65, facilitating easier comparison of the genetic architecture between EOAD and LOAD. We observed shared genetic determinants between EOAD and LOAD, indicating that they should not be considered as distinct disease entities. We confirmed that APOE plays a critical and comparable role in EOAD and LOAD, ⁴⁵ and that the two subtypes shared a range of variants and genes at less significant levels. The shared genetic architecture we observed is consistent with the shared pathologic features at the end stage of disease. ⁴⁵ In addition, we noted that the percentage of shared genetic determinants between EOAD and LOAD decreased as the significance threshold was relaxed, suggesting that more EOAD‐associated loci could be discovered with larger samples and more advanced sequencing techniques in the future. Although several suggestive loci associated with EOAD have been identified in this study, they require further confirmation, and the signals need to be enhanced in future studies until they reach genome‐wide significance.

A major finding of this work is the ancestry‐related differences in the genetic architecture of AD. Variations in allele frequencies among different populations could account for the observed discrepancies to some extent, as genetic analyses tend to identify significant variants that have higher allele frequencies within the specific population. ⁴⁶ Our results support this hypothesis because 12% of lead variants of the established loci identified from European studies were rare in Chinese (Table S1) and most of the novel loci identified in this study also exhibit different MAFs than those in other populations. In addition, the distinct linkage disequilibrium patterns and epigenetic differences resulting from various cultural environments might also influence the ancestry‐specific genetic effects of AD. The limited evidence of associations between variants within our novel genes and AD core biomarkers could be partially explained by racial disparity, ⁴² , ⁴³ although it is also possible that these variants influence AD risk through other mechanisms rather than of altering the core biomarkers. Moreover, despite the recent emphasis on the impact of racial disparities on AD genetics, ¹ non‐European studies are still not sufficient, thus limiting our understanding of AD etiologies and personalized treatment in non‐European populations. Our efforts could help refine the landscape of AD genetics from the East Asian perspective.

EGFR was initially identified as a risk gene by large‐scale GWAS in 2022, although the closest gene of the lead variant was not EGFR exactly. ²⁹ Prior to this study, EGFR polymorphisms had been found to be associated with AD by a case–control study of Chinese Han ancestry, ⁴⁷ and the results could also be validated in our Stage 1 meta‐analysis within Chinese (rs730437, p = 0.017; rs1468727, p = 7.12 × 10⁻⁶) but not European samples (p > 0.05 for both variants ⁴ , ²⁹ ). Because our lead variant of EGFR locus was also not significant in the European GWAS, we speculate that the genetic effect of EGFR might vary between Chinese and European populations, and this hypothesis needs further investigation. Previous literature has shown that the hyperactivation of the EGFR is prevalent in several neuroinflammatory conditions, including AD. ⁴⁴ , ⁴⁸ In AD brains, pathological proteins (e.g., Aβ42) might upregulate EGFR, subsequently initiating downstream signaling phosphorylation, and ultimately leading to extensive Aβ42 production and tau phosphorylation. ⁴⁴ The circulating EGFR protein was also revealed as a marker predicting dementia risk within a 15‐year follow‐up in middle‐aged adults. ⁴⁹ It is encouraging that recent experiments suggest that the inhibition of EGFR could effectively suppress the activation of reactive astrocytes, promote autophagy, ameliorate Aβ toxicity and neuroinflammation, and facilitate the regeneration of axonal degradation. ⁵⁰ , ⁵¹ EGFR inhibitors have been approved for cancer treatment, some of which (e.g., lapatinib and ibrutinib) are permeable to the blood–brain barrier (BBB), providing a unique drug‐repurposing opportunity for AD. ⁵⁰

In contrast with EGFR, another prioritized gene, TCN2, has less existing evidence, probably due to the low MAF of our significant variants in other populations. When transcobalamin II (encoded by TCN2) is attached to vitamin B12, it is referred to as holo‐transcobalamin (holo‐TC, also known as active vitamin B12), delivering vitamin B12 to cells. ⁵² Holo‐TC was reported to have an association with cognition and could predict AD risk. ⁵³ , ⁵⁴ Although previous efforts have attempted to link the holo‐TC concentration‐influencing variant in TCN2 (rs1801198) with AD risk in Caucasians, no significant association was found. ⁵⁵ We confirmed that rs1801198 was not associated with AD in our Chinese cohort (p > 0.05) and identified a novel genome‐wide significant variant in TCN2 (rs1815157), whereas its low MAF might have hindered the investigation of TCN2 in previous European‐based studies. The possible underlying mechanisms of TCN2 on AD include vitamin B12 deficiency, ⁵⁶ cross‐talk with amyloid or tau pathologies, ⁵⁷ , ⁵⁸ , ⁵⁹ , ⁶⁰ and endothelial damage. ⁶¹ Considering that 87% of patients diagnosed with probable AD exhibited vascular pathology on post‐mortem examination, ⁶² it would be unsurprising that the AD‐associated gene TCN2 is enriched in endothelial cells. Further exploration of TCN2 might provide a breakthrough for understanding the coexistence of vascular and Alzheimer's pathologies in AD cases.

Phe‐MR revealed that higher levels of MCHC have potential causal effects on lower risk of EOAD but not LOAD in East Asian populations. Our result aligns with a longitudinal observational study of 313,448 participants, ⁶³ which reported that the protective effect of MCHC on all‐cause dementia attenuated with age (interaction with age, p = 0.002). MCHC is an indicator of red blood cell and iron homeostasis. ⁶⁴ One possible explanation for the observed relationship is that iron accumulates in tissues (e.g., the neocortex) and is not adequately mobilized in individuals with a high risk of EOAD. ⁶⁵ Our findings support that the abnormal hemoglobin and iron regulation might contribute to EOAD etiology.

This study has some limitations that should be noted. First, patients with known pathogenic mutations (PSEN1, PSEN2, and APP) that could cause EOAD had not been excluded before GWAS analysis. Nonetheless, considering that 90%–95% of EOAD cases did not carry these pathogenic mutations, we consider the influence on results to be minimal. Second, most AD patients were diagnosed based on NINCDS‐ADRDA clinical criteria, so the results warrant further validation in biologically defined AD. Third, we were not able to quantify the heritability and polygenicity of EOAD and LOAD due to the relatively small sample size of the discovery sample. Fourth, it may be better to use large‐scale deep‐sequencing cohorts such as TOPMed as the imputation reference panel, which could further enhance the imputation accuracy for low‐frequency variants. ⁶⁶ , ⁶⁷ Finally, our prioritization scoring system also has limitations, including not assessing the directional consistency and causal relationships between variants' effects on eQTL and gene expression, and not incorporating other genetic regulatory mechanisms such as protein isoforms and epigenetic modifications.

In summary, our results take a step further in elucidating the complex etiology of AD. We revealed the extensive genetic overlap between EOAD and LOAD; however, differences exist. Using data from two populations, we were able to identify ancestry‐specific variants and genes, such as TCN2. Three genes—EGFR, APOC1, and TCN2—have been biologically prioritized, with EGFR emerging as a promising target for AD treatment. Future larger studies covering all AD clinical subtypes using GWAS, epigenomic sequencing, and more advanced technologies in all populations can help enhance the AD genetic landscape and provide more valuable insights into AD etiology.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the supporting information.

CONSENT STATEMENT

The study was approved by the institutional review boards of all participating institutions. All participants provided written informed consent in accordance with the Declaration of Helsinki prior to study enrollment.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Ge Y‐J, Chen S‐D, Wu B‐S, et al. Genome‐wide meta‐analysis identifies ancestry‐specific loci for Alzheimer's disease. Alzheimer's Dement. 2024;20:6243–6256. 10.1002/alz.14121