Prediction of Alzheimer's disease using multi-variants from a Chinese genome-wide association study

Abstract

Previous genome-wide association studies have identified dozens of susceptibility loci for sporadic Alzheimer’s disease, but few of these loci have been validated in longitudinal cohorts. Establishing predictive models of Alzheimer’s disease based on these novel variants is clinically important for verifying whether they have pathological functions and provide a useful tool for screening of disease risk. In the current study, we performed a two-stage genome-wide association study of 3913 patients with Alzheimer’s disease and 7593 controls and identified four novel variants (rs3777215, rs6859823, rs234434, and rs2255835; P_combined = 3.07 × 10⁻¹⁹, 2.49 × 10⁻²³, 1.35 × 10⁻⁶⁷, and 4.81 × 10⁻⁹, respectively) as well as nine variants in the apolipoprotein E region with genome-wide significance (P < 5.0 × 10⁻⁸). Literature mining suggested that these novel single nucleotide polymorphisms are related to amyloid precursor protein transport and metabolism, antioxidation, and neurogenesis. Based on their possible roles in the development of Alzheimer’s disease, we used different combinations of these variants and the apolipoprotein E status and successively built 11 predictive models. The predictive models include relatively few single nucleotide polymorphisms useful for clinical practice, in which the maximum number was 13 and the minimum was only four. These predictive models were all significant and their peak of area under the curve reached 0.73 both in the first and second stages. Finally, these models were validated using a separate longitudinal cohort of 5474 individuals. The results showed that individuals carrying risk variants included in the models had a shorter latency and higher incidence of Alzheimer’s disease, suggesting that our models can predict Alzheimer’s disease onset in a population with genetic susceptibility. The effectiveness of the models for predicting Alzheimer’s disease onset confirmed the contributions of these identified variants to disease pathogenesis. In conclusion, this is the first study to validate genome-wide association study-based predictive models for evaluating the risk of Alzheimer’s disease onset in a large Chinese population. The clinical application of these models will be beneficial for individuals harbouring these risk variants, and particularly for young individuals seeking genetic consultation.

Jia et al. identify novel Alzheimer’s disease-related variants in a two-stage genome-wide association study in a Chinese population, and use the variants to build 11 predictive models. Validation of the models in a separate longitudinal cohort confirms that they can predict Alzheimer's disease risk.

Introduction

Alzheimer’s disease is the most common type of dementia and is genetically complex with an estimated heritability of 60–80% (Gatz et al., 1997). Previous genome-wide association studies (GWASs) of Alzheimer’s disease in Caucasian, African-American, and Asian populations have identified genetic risk variants in ABCA7, BIN1, CASS4, CD2AP, CD33, CDK5RAP2, CELF1, CLU, COBL, CR1, ECHDC3, EPHA1, EXOC3L2, FERMT2, HLA-DRB5, HLA-DRB1, HS3ST1, INPP5D, KANSL1, MEF2C, MS4A, NME8, PICALM, PM20D1, PTK2B, SLC10A2, SLC24A4, SORL1, TREM2, and ZCWPW1 (Harold et al., 2009; Lambert et al., 2009, 2013a; Seshadri et al., 2010; Hollingworth et al., 2011; Naj et al., 2011; Guerreiro et al., 2013; Miyashita et al., 2013; Reitz et al., 2013; Desikan et al., 2015; Jun et al., 2016; Lacour et al., 2017; Miron et al., 2018; Sanchez-Mut et al., 2018; Kunkle et al., 2019). These variants affect several Alzheimer’s disease-related processes, such as lipid metabolism, inflammation, innate immunity, production and clearance of amyloid-β, and endosomal vesicle recycling (Selkoe and Hardy, 2016). However, few of the variants reported in Caucasians have been identified in the Chinese population (Wang et al., 2016). A recent whole genome sequencing study in a Chinese population identified variants in GCH1 and KCNJ15, in addition to the well-known apolipoprotein E (APOE) locus; however, the sample size of this study was relatively small (Zhou et al., 2018).

Recently, genetic predictive models have been established for predicting the onset of Alzheimer’s disease using a polygenic risk score approach, which was used to reveal polygenetic contributions to Alzheimer’s disease risk of common single nucleotide polymorphisms (SNPs) that show a disease association but fail to meet the accepted P-value threshold for genome-wide significance (Escott-Price et al., 2015, 2017a, b, 2019; Chouraki et al., 2016; Stocker et al., 2018; Leonenko et al., 2019). These studies showed variable results. Specifically, Escott-Price et al. reported that the area under the curve (AUC) of their predictive models, which included APOE, >80 000 SNPs, age, and sex as predictors, was 0.78, whereas in their other study, the AUC of their models including >20 000 SNPs and APOE as predictors increased to 0.84 as the included individuals were pathologically but not clinically confirmed (Escott-Price et al., 2015, 2017a). However, despite the high predictive accuracy of these polygenic risk score-based models, it may not be easy to use these models in a clinical setting because an individual may not carry so many risk variants. Thus, simple and effective Alzheimer’s disease predictive models are needed for use as tools to screen for the genetic risk of Alzheimer’s disease, particularly in young individuals who carry the risk variants.

The current study aimed to investigate novel Alzheimer’s disease-related genetic variants in a GWAS, to establish predictive models based on these variants, and to validate the models in a longitudinal cohort. This approach can be applied for early intervention in individuals who are at a risk of developing Alzheimer’s disease.

Materials and methods

Subjects

The two-stage GWAS study involved 3913 patients with Alzheimer’s disease and 7593 controls from a Chinese population. The cohorts used in the two stages were independent of each other. Patients with Alzheimer’s disease were recruited from the outpatient memory clinics at the Department of Neurology, Xuanwu Hospital, Capital Medical University, Beijing, China and 46 other participating hospitals across China from 2013 to 2018. All diagnoses of Alzheimer’s disease in this study were based on the recommendations of the National Institute on Aging and the Alzheimer’s Association workgroup (McKhann et al., 2011) or National Institute of Neurological and Communicative Disorders and Stroke and the Alzheimer’s Disease and Related Disorders Association criteria (McKhann et al., 1984), with an age-at-onset ≥60 years and no family history of dementia. Controls were recruited from the aforementioned medical centre hospitals. All controls were ≥60 years of age, cognitively normal (without subjective memory complaints, a Mini-Mental State Examination score of 26–30, and Clinical Dementia Rating Scale score of 0), and free of any general or laboratory evidence of diseases that could impact cognition. Demographic information was collected from each subject using a structured questionnaire.

Furthermore, using associated SNPs from the GWAS data, predictive models of Alzheimer’s disease were generated by combining risk variants. To estimate the effectiveness of the predictive models, participants from a longitudinal cohort of the China Cognition and Aging Study (China COAST) (Jia et al., 2014) were selected. China COAST was a longitudinal study established in 2008 as a multicentre cohort study comprising normal, mild cognitive impairment-, and Alzheimer’s disease-affected individuals across 30 of 34 provinces in China with yearly follow-up. The inclusion criteria were as follows: (i) the individual was cognitively normal 10 years ago at baseline with indicative blood samples; (ii) the individual developed Alzheimer’s disease at the time of sample collection for the present study 10 years later; and (iii) the individual had a detailed clinical data profile including psychometric evaluation every year during follow-up. Finally, 5474 participants were recruited, from among which 2358 developed Alzheimer’s disease and 3116 were cognitively normal in 2019 (Supplementary Table 1). The study was approved by the Ethical Committees of Xuanwu Hospital, Capital Medical University. Written informed consent was obtained from either the subjects or their legal guardians according to the Declaration of Helsinki.

GWAS study

First stage

Genomic DNA was extracted from peripheral blood samples using a modified salting-out procedure (Nasiri et al., 2005). In the first stage, we performed genome-wide genotyping of 1679 patients with Alzheimer’s disease and 2508 controls using Illumina HumanOmniZhongHua-8 Bead Chips (Illumina). After genotyping, systematic quality control analyses were conducted using PLINK 1.90 software (http://www.cog-genomics.org/plink2) (Purcell et al., 2007; Chang et al., 2015). First, 118 samples (84 patients with Alzheimer’s disease and 34 controls) were omitted because of sample duplicates or cryptic relatedness (PI_HAT > 0.1875, which is the identity-by-descent expected between third- and second-degree relatives) (Ellingson and Fardo, 2016), or low individual call rate (<0.95). The remaining samples were assessed for population outliers and stratification in principal component analysis using EIGENSTRAT (Patterson et al., 2006). All non-autosomal variants were excluded from statistical analyses, as well as SNPs with a call rate <98%, minor allele frequency <0.01, and/or significant deviation from Hardy-Weinberg equilibrium in controls (P < 1.0 × 10⁻⁴) (Supplementary Table 2). Following quality control processing, the genotypes of 765 144 SNPs in 4069 Chinese individuals (1595 patients with Alzheimer’s disease and 2474 controls) were further analysed.

Phasing and imputation were performed by SHAPEIT (Delaneau et al., 2011) and IMPUTE2 (Howie et al., 2009), respectively, and version 3 of the 1000 Genomes Project data was used as the reference set (Genomes Project et al., 2012). Variants with r² values < 0.80 or impute information measures < 0.50 from IMPUTE2, missing frequency > 0.02, deviation from Hardy-Weinberg equilibrium (P < 1.0 × 10⁻⁴), and minor allele frequency < 0.01 were excluded from post-imputation quality control analysis. Logistic regression analysis of GWAS data was conducted before and after imputation to test the differences in allele dosage between cases with Alzheimer’s disease and controls under an additive genetic model, adjusted for sex, APOE status, age (defined as age-at-onset for cases and age-at-last exam for controls), and population substructure using the first two principal components with PLINK 1.90 software. Manhattan and quantile-quantile plots of the first stage before and after imputation and adjustments for sex and APOE status were generated using the R qqman package (Version 3.4.2, https://www.r-project.org/). Regional association plots were generated via LocusZoom (http://locuszoom.sph.umich.edu/locuszoom/) (Pruim et al., 2010). Linkage disequilibrium plots of variants in chromosome 19 were generated using Haploview software (https://www.broadinstitute.org/haploview/haploview). Conditional analysis was performed to assess the independence of the novel associations of the genotyped SNPs. In addition, stratified analysis was performed by gender and disease status.

Power calculations with Quanto software were applied to calculate the power of the results from the discovery stage (Gauderman et al., 2006). Alzheimer’s disease prevalence was set to 3.21% in accordance with epidemiological studies of Alzheimer’s disease in Chinese subjects aged ≥65 years (Jia et al., 2014). Parameters included outcome (disease), design (unmatched case-control ratio of 1:1.5), hypothesis (gene only), sample size (n = 1679 cases), significance (1.0 × 10⁻⁵, two-sided), mode of inheritance (log-additive), and population risk (0.0321).

Second stage

To replicate the first stage association results, the top 34 variants showing an association with a P < 1.0 × 10⁻⁵ after adjusting for age, sex, APOE status and the first two principal components were selected and analysed as part of an independent cohort of 7319 Chinese individuals consisting of 2234 cases with Alzheimer’s disease and 5085 controls (Table 1 and Supplementary Table 1). These 34 SNPs were genotyped at BioMiao Biological Technology Beijing Co. using the MassArray System (Agena iPLEXassay).

Table 1.

Association of 13 SNPs with Alzheimer's disease in the first and second stages

Chr.	SNP	Nearest gene	Position	Minor allele	First stage				Second stage				P com	ORcom (95% CI)	P het
					(1679 cases, 2508 controls)				(2234 cases, 5085 controls)
					MAF_CA, %	MAF_CL, %	OR	P gwas	MAF_CA, %	MAF_CL, %	OR	P re
5	rs3777215	RHOBTB3, GLRX	95122000	A	12.9	17.1	0.69	9.34 × 10−8	15.5	20.9	0.69	1.31 × 10−13	3.07 × 10−19	0.69 (0.64–0.74)	0.72
5	rs6859823	CTC-278L1.1	105554384	T	29.2	35.8	0.74	3.28 × 10−9	27.1	33.9	0.72	5.32 × 10−15	2.49 × 10−23	0.74 (0.69–0.78)	0.79
14	rs234434	CTD-2506J14.1	97821020	G	36.6	25.8	1.42	3.70 × 10−11	37.2	25.4	1.75	1.07 × 10−44	1.35 × 10−67	1.71 (1.61–1.82)	0.87
19	rs11668861	NECTIN2	45380970	G	33.1	22.7	1.76	1.75 × 10−10	34.9	22.4	1.86	2.99 × 10−52	4.13 × 10−63	1.81 (1.69–1.94)	0.32
19	rs6859	NECTIN2	45382034	A	40.9	31.6	1.55	2.92 × 10−8	43.7	32.6	1.61	9.67 × 10−35	3.70 × 10−42	1.58 (1.48–1.69)	0.48
19	rs3852860	NECTIN2	45382966	C	33.5	23.6	1.72	1.87 × 10−9	35.6	24.1	1.75	3.07 × 10−44	4.20 × 10−54	1.72 (1.60–1.84)	0.48
19	rs71352238	TOMM40	45394336	C	23.3	9.6	3.13	5.64 × 10−15	21.8	8.7	2.91	2.40 × 10−100	1.16 × 10−131	2.90 (2.66–3.17)	0.69
19	rs157580	TOMM40	45395266	A	53.9	45.4	1.48	1.10 × 10−6	54.9	46.6	1.40	1.60 × 10−19	2.99 × 10−25	1.40 (1.31–1.49)	0.86
19	rs2075650	TOMM40	45395619	G	23.3	9.4	3.27	3.05 × 10−18	25.2	9.4	3.25	5.30 × 10−132	2.57 × 10−164	3.17 (2.91–3.45)	0.36
19	rs157582	TOMM40	45396219	T	32.0	18.9	2.33	2.53 × 10−10	33.7	18.9	2.19	6.57 × 10−76	2.97 × 10−95	2.15 (2.00–2.31)	0.62
19	rs439401	APOE	45414451	C	50.9	43.3	1.50	2.06 × 10−6	52.8	42.1	1.54	2.08 × 10−31	1.17 × 10−35	1.50 (1.40–1.59)	0.14
19	rs4420638	APOC1	45422946	G	27.2	12.2	2.98	2.62 × 10−17	28.5	11.0	3.20	3.40 × 10−140	8.32 × 10−171	3.07 (2.83–3.32)	0.13
21	rs2255835	CHODL	19491664	C	33.6	27.7	1.33	8.61 × 10−6	26.2	23.0	1.19	6.78 × 10−05	4.81 × 10−9	1.23 (1.16–1.31)	0.51

Chr. = chromosome; MAF_CA = minor allele frequency of cases; MAF_CL = minor allele frequency of controls; Minor allele = minor allele of the controls; OR = odds ratios calculated according to the minor allele; OR_com = the odds ratio of the combined analysis; P_com = the P-value of the combined stage using meta-analysis; P_gwas, the P-value of the first stage adjusted for sex, APOE status, age (defined as age at onset for cases and age at last exam for controls) and two ancestry principal components; P_het = the P-value of heterogeneity; Position = base pair position according to hg 19; P_re = the P-value of the second stage.

Combined analysis of the first and second stages

To improve statistical power, a meta-analysis was applied to combine the associated results from the first two stages using METAL (Willer et al., 2010) with an inverse variance-based model. Heterogeneity tests between the two groups were performed using the Breslow-Day test (Higgins and Thompson, 2002), and the extent of heterogeneity was assessed using the I² and P-values of the Q statistics calculated by METAL (Higgins et al., 2003).

Single nucleotide polymorphism annotation

SNPnexus was used for SNP annotation (https://www.snp-nexus.org/v4/) (Chelala et al., 2009; Dayem Ullah et al., 2012, 2013, 2018). For co-localization of SNPs with significant associations in both stages of the study, we conducted expression quantitative trait loci (eQTL) analysis using the dataset presented by Ramasamy et al. (2014), COLOC analysis (http://coloc.cs.ucl.ac.uk) (Giambartolomei et al., 2014) using the brain-eQTL datasets (Trabzuni et al., 2011; Ramasamy et al., 2013), and summary Mendelian randomization-Heidi analysis (https://cnsgenomics.com/software/smr/) (Zhu et al., 2016) using summary eQTL data from the brain and blood (Westra et al., 2013; Lloyd-Jones et al., 2017; Qi et al., 2018). The expression of novel Alzheimer’s disease-associated genes was analysed using data from the National Center for Biotechnology Information Gene Expression Omnibus dataset (http://www.ncbi.nlm.nih.gov/geo). These included the expression of genes in the frontal cortex, hippocampus, and temporal cortex of controls and patients with Alzheimer’s disease (Supplementary Table 3). Prism software (version 8.0.0, GraphPad Software, Inc., CA, USA) was used to compare gene expression between cognitively normal and Alzheimer’s disease groups (unpaired t-test and Welch's t-test) and to generate figures. For the Alzheimer’s disease-associated genes in this study, STRING analysis was performed to evaluate protein-protein interactions (Szklarczyk et al., 2015). Medium confidence (0.400) was used as the minimum required interaction score and no more than 50 interactors were shown in the first shell. The exported network was analysed using the bioinformatics software platform Cytoscape (Version: 3.7.1, https://cytoscape.org/). In addition, we exported the Gene Ontology information, including molecular function, biological process, and cellular component, as well as the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways for the network, from STRING analysis.

APOE genotyping

The APOE genotypes for haplotypes derived from rs7412 and rs429358 in samples from both stages of the study were determined using the Sanger sequencing method (Sanger et al., 1977).

Validation of gene associations in Caucasian populations

To determine whether there is underlying heterogeneity in the contributors of genetic risk between Chinese and Caucasian populations, the eligible novel SNPs were examined in the data from ‘The International Genomics of Alzheimer’s Project summary statistics from stage 1 data’ (Lambert et al., 2013b). The genetic correlation between the Chinese GWAS and publicly available International Genomics of Alzheimer’s Project (IGAP) summary statistics was estimated using linkage disequilibrium score regression implemented in the online software LD Hub (http://ldsc.broadinstitute.org/) (Zheng et al., 2017).

Predictive model study

We performed predictive modelling using the polygenic risk score based on SNP significance in combined analysis and APOE status as predictor variables, based on the data of the first stage. The individual polygenic risk scores were generated as sums of the risk variants weighted by effect sizes derived from logistic regression. We also ran the predictive analyses on second-stage data using the same factors. Furthermore, we tested different predictive models with different combinations of SNPs in a population negative for APOE ε4. Areas under the receiver operating characteristic curve were calculated by comparing the observed case/control status and polygenic risk score calculated using PRSice2 (Choi and O'Reilly, 2019) profiling in a standard weighted allele-dose manner.

To confirm the capacity of the models to predict Alzheimer’s disease, we applied the models to individuals who were recruited in a longitudinal study from 2009 to 2019. To estimate the effectiveness of the GWAS-based predictive models, by measuring the fraction of individuals living without Alzheimer’s disease for a certain amount of time from baseline, survival curve analyses were performed using the follow-up data in this longitudinal cohort.

Data availability

The data that support the findings of this study are available on request from the corresponding author.

Results

Demographics of three cohorts

In the GWAS, a total of 11 506 individuals participated in this two-stage study, including 3913 patients with Alzheimer’s disease and 7593 controls (Supplementary Table 1). In the first stage, the mean age at onset of 1679 patients with Alzheimer’s disease was 72.07 ± 6.4 years, of which 949 (56.5%) were female, whereas the mean age at examination of 2508 controls was 72.95 ± 19.5 years, of which 1398 (55.7%) were female. In the second stage, the mean age at onset of 2234 patients with Alzheimer’s disease was 72.35 ± 7.2 years of which 1253 (56.1%) were female, whereas the mean age at examination of 5085 controls was 69.22 ± 7.0 years, of which 2931 (57.6%) were female. In the longitudinal cohort, 5474 individuals were recruited, comprising 2358 patients with Alzheimer’s disease and 3116 controls (Supplementary Table 1). The mean age at onset of 2358 patients with Alzheimer’s disease was 71.70 ± 6.8 years, of which 1351 (57.3%) were female, whereas the mean age at examination of 3116 controls was 74.70 ± 7.5 years, of which 1676 (53.8%) were female. The flow chart of the current study is shown in Fig. 1.

Figure 1.

Study flow chart. AD = Alzheimer’s disease.

Results from genome-wide association studies

In the first stage, 1679 patients with Alzheimer’s disease and 2508 controls were genotyped (Supplementary Table 1). Overall, 765 144 autosomal SNPs passed the quality control standards and were included for further analysis. Principal component analysis confirmed that the patients with Alzheimer’s disease and controls were well-matched (Supplementary Fig. 1). A quantile-quantile plot indicated that population stratification had negligible effects on the statistical results (λ_GC = 1.08; Supplementary Fig. 2). After adjusting for age, sex, and APOE status along with the first two principal components, several markers on various chromosomes exhibited genome-wide significance, with the sentinel markers occurring on chromosome 19 (Supplementary Fig. 3). Power calculations indicated that the sample size used in this GWAS provided sufficient statistical power to detect Alzheimer’s disease-associated variants. In the second stage, the top 34 SNPs with evidence of associations with Alzheimer’s disease (Supplementary Table 4) were selected for genotyping as part of an independent Chinese cohort of 2234 patients with Alzheimer’s disease and 5085 controls (Supplementary Table 1). Of these 34 Alzheimer’s disease-associated SNPs, 13 surpassed the Bonferroni correction threshold (P < 1.47 × 10⁻³) with no detectable heterogeneity between stages. This included nine SNPs on chromosome 19 in the APOE region (Jun et al., 2012) (APOE rs439401, APOC1 rs4420638, TOMM40 rs2075650, TOMM40 rs71352238, TOMM40 rs157582, TOMM40 rs157580, NECTIN2 rs11668861, NECTIN2 rs3852860, and NECTIN2 rs6859; Table 1). The other four SNPs were not located on chromosome 19 (rs3777215, rs6859823, rs234434, and rs2255835; P = 1.31 × 10⁻¹³, 5.32 × 10⁻¹⁵, 1.07 × 10⁻⁻⁴⁴, and 6.78 × 10⁻⁰⁵, respectively). In combined analysis, all 13 of these SNPs showed associations exceeding the genome-wide significance threshold (P < 5.0 × 10⁻⁸; Table 1). Furthermore, all four novel variants remained significant after adjusting for the effect of APOE ε4. Stratified analyses indicated that the 34 top SNPs in the first stage remained significant after stratification by gender (Supplementary Table 5). Conditional analyses indicated that the four novel SNPs we identified were independent signals (Supplementary Table 6).

Functional annotation of the four novel single nucleotide polymorphisms

Variant rs3777215 was significantly associated with Alzheimer’s disease [P_combined = 3.07 × 10⁻¹⁹, odds ratio (OR) = 0.69] and was located on chromosome 5 in the intron regions of RHOBTB3 and GLRX (Fig. 2A, Table 1 and Supplementary Table 7). As shown in Fig. 3, RHOBTB3 was significantly upregulated in the frontal cortex, hippocampus, and temporal cortex of patients with Alzheimer’s disease compared to cognitively normal individuals (P < 0.05, 0.01 or 0.0001; Fig. 3A). GLRX showed lower expression levels in the brain tissues of patients with Alzheimer’s disease than in that of cognitively normal individuals (P < 0.05, 0.01 or 0.0001; Fig. 3B). Another SNP detected on chromosome 5 was rs6859823 (P_combined = 2.49 × 10⁻²³, OR = 0.74; Fig. 2B and Table 1). Variant rs6859823 was intergenic and located between RNA5SP189 and CTC-278L1.1 (Supplementary Table 7). RNA5SP189 and CTC-278L1.1 were both identified as pseudogenes according to GeneCards and SNPnexus, and neither have been previously reported to be associated with Alzheimer’s disease or any other disease. SNP rs234434 (P_combined = 1.35 × 10⁻⁶⁷, OR = 1.71; Fig. 2C and Table 1) was intergenic and was between two long intergenic non-coding RNAs known as RP11-359N5.1 and CTD-2506J14.1. SNP rs2255835 was located on chromosome 21 in the intron region of CHODL and showed genome-wide significance (P_combined = 4.81 × 10⁻⁹, OR = 1.23; Fig. 2D and Table 1). CHODL expression was higher in the hippocampus (P < 0.05; Fig. 3C) and lower in the temporal cortex of patients with Alzheimer’s disease compared to that in normal individuals (P < 0.05; Fig. 3C). Co-localization analyses indicated that the four SNPs were related to various genes being expressed in the blood and different brain regions (Supplementary Tables 8–11). Roadmap epigenomics showed that rs3777215 and rs2255835 were related to transcriptional activation (H3K36me3, H3K4me1, and H3K14ac) in neurons or neuronal progenitor cells, whereas rs234434 was associated with transcriptional repression (H3K27me3) in neurons (Supplementary Table 7). STRING analysis demonstrated that the proteins encoded by the variant genes, except for CHODL, interact with APOE (Supplementary Fig. 4). Gene Ontology enrichments of the genes in the STRING network suggested that these genes are involved in several biological processes. For example, RHOBTB3 was suggested to be involved in the ‘establishment of localization’ and ‘transport’ (Supplementary Table 12). Genes enriched in Gene Ontology cellular components and molecular functions, as well as KEGG pathways, are listed in Supplementary Tables 13–15.

Figure 2.

Regional association plots. (A–D) Association results are shown for the analysed SNPs with recombination rates in the four loci associated with genome-wide significance at chromosome 5 (A and B), 14 (C), and 21 (D). The −log10 (P-values) (y-axis) of SNPs within the ±500 kb region centred on each marker SNP are presented according to the chromosomal positions of the SNPs (x-axis; NCBI Build 37). Purple diamonds represent the most significantly associated SNP (marker SNP) in the combined analysis. SNPs are coloured according to their linkage disequilibrium with the marker SNP. Linkage disequilibrium values were based on the 1000 Genome Project Asian data. Blue lines represent the estimated recombination rates based on the 1000 Genome Project samples. Arrows depict genes in the regions of interest annotated from the UCSC Genome Browser.

Figure 3.

Differential expression of the annotated genes in Gene Expression Omnibus datasets. (A–C) shows the differential expression of RHOBTB3 (A), GLRX (B), CHODL (C) in frontal cortex, hippocampus, and temporal cortex. The bold red line indicates the median of each group, and the black dotted lines show the quartiles. AD = Alzheimer’s disease; CN = cognitively normal; FC = fold change; GSE = Gene Expression Omnibus Series; ns = no significance; *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001.

Validation in Caucasian genome-wide association studies datasets

Overall, the APOE ε4 allele frequency in Chinese subjects in the present study was lower than that in Caucasians for both, patients with Alzheimer’s disease and controls (Supplementary Table 1). Despite these differences, nine SNPs with significant associations in the APOE region in the present study were either reported previously or found in strong linkage disequilibrium with nearby SNPs in Caucasians (Supplementary Fig. 5). One of the four novel SNPs outside chromosome 19 reached genome-wide significance in the IGAP stage 1 data for non-Asian populations (rs6859823; Supplementary Table 16). Moreover, linkage disequilibrium score regression analysis of Chinese GWAS and publicly available IGAP summary statistics revealed a genetic correlation of −0.14 (P = 0.73).

Predictive models

We tested 11 predictive models; four models were used to analyse all populations (Fig. 4A–D) and seven models were used to analyse subjects who were negative for APOE ε4 (Fig. 4E–K) in this study. The number of SNPs in our models was relatively low, and even for the maximum, the number of SNPs for model A3 was only 13 with an AUC of 0.73 [95% confidence interval (CI): 0.70–0.75] in the first stage. The AUCs of all 11 models were significant (P < 0.05; range of AUC: 0.63–0.73), and the specific AUC values for the 11 models are presented in Fig. 4. For model A1, with the four novel SNPs found in this study and APOE ε4 status as predictors, training on the first-stage data, prediction accuracy AUC = 0.69 (95% CI: 0.67–0.71) was achieved based on a logistic regression model. However, the prediction accuracy AUC reached 0.73 (95% CI: 0.71–0.74) when using second-stage data (Fig. 4A). Model B1 (with our four novel SNPs found in this study as predictors) and model B4 (with two novel SNPs and two SNPs in the APOE region in this study as predictors) covered more individuals who were negative for APOE ε4 in our study, and the number of model B2 was maximum in B models with four novel SNPs and three APOE-region SNPs. For model B1, the prediction accuracy AUC values were 0.63 (95% CI: 0.61–0.66) and 0.66 (95% CI: 0.63–0.69) based on the first and second data, respectively (Fig. 4E); whereas, they were 0.71 (95% CI: 0.68–0.73) and 0.68 (95% CI: 0.65–0.70) for model B4 (Fig. 4H), and 0.72 (95% CI: 0.69–0.74) and 0.68 (95% CI: 0.65–0.70) for model B2, based on the first and second data, respectively (Fig. 4E).

Figure 4.

ROC curves for 11 predictive models with different predictors in the three cohorts and survival curves in a longitudinal cohort. The factors included in the 11 models are as follows. (A) A1: APOE ε4 status, rs3777215, rs6859823, rs234434 and rs2255835; (B) A2: APOE ε4 status, rs3777215, rs6859823, rs234434, rs2255835, rs11668861, rs71352238 and rs4420638; (C) A3: APOE ε4 status, rs3777215, rs6859823, rs234434, rs2255835, rs11668861, rs6859, rs3852860, rs71352238, rs157580, rs2075650, rs157582, rs439401 and rs4420638; (D) A4: rs3777215, rs6859823, rs234434, rs2255835, rs11668861, rs6859, rs3852860, rs71352238, rs157580, rs2075650, rs157582, rs439401 and rs4420638; (E) B1: rs3777215, rs6859823, rs234434 and rs2255835; (F) B2: rs3777215, rs6859823, rs234434, rs2255835, rs11668861, rs71352238 and rs4420638; (G) B3: rs3777215, rs6859823, rs234434, rs71352238 and rs4420638; (H) B4: rs3777215, rs234434, rs71352238 and rs4420638; (I) B5: rs6859823, rs234434, rs71352238 and rs4420638; (J) B6: rs3777215, rs6859823, rs234434 and rs71352238; (K) B7: rs3777215, rs6859823, rs234434 and rs4420638. (L) Survival curves of the longitudinal cohort, *P < 0.001. AUC1 indicates AUC of the first stage; AUC2 indicates AUC of the second stage; AUC3 indicates AUC of the longitudinal cohort; ROC = receiver operating characteristic curve.

Longitudinal study

In the longitudinal cohort, all 13 SNPs identified in the combined analysis reached significance (P < 3.85 × 10⁻³; Supplementary Table 17). The AUC values of the 11 predictive models were also significant using the data from the longitudinal cohort (P < 0.05; range of AUC: 0.67–0.73; Fig. 4). To confirm the capacity of the models to predict Alzheimer’s disease, we performed survival curve analysis. First, the predictive models were applied to individuals (2358 Alzheimer’s disease, 3116 controls, 5474 in total) in the longitudinal study. Models A1–A4 were generated based on an APOE ε4-positive or APOE ε4-negative population, and models B1–B7 were based on an APOE ε4-negative population only. Because the same populations were included in models A1–A4, and models B1–B7, we combined them into model A and model B for survival analysis. Individuals were divided into three groups as follows: model A positive (model A+), model B positive (model B+), and model A or B negative (model A− or B−). In total, the number of subjects in model A+ was 213, in model B+ was 1035, and in model A− or B− was 4226. Data from individuals from each group were plotted into the survival curve according to the follow-up data comprising the incidence and onset of Alzheimer’s disease from 2009 to 2019. Kaplan-Meier survival curve analysis revealed significant differences between the model A+, model B+, and model A− or B− groups (P < 0.001; Fig. 4L), indicating that individuals with model A+ and model B+ had a shorter latency and a higher proportion of Alzheimer’s disease during the follow-up period. Our data suggest that the 11 predictive models have sufficient capacity for predicting Alzheimer’s disease risk.

Discussion

We found four novel variants in addition to nine APOE-region variants that were correlated with Alzheimer’s disease risk in a Chinese population. Using different combinations of these variants and the APOE status, we established 11 predictive models with significant AUC values. Validation of these models in a longitudinal cohort indicated the genetic power of SNPs for Alzheimer’s disease prediction. Overall, these findings may improve the understanding of how genetic variants impact the initiation of Alzheimer’s disease.

The novel variants identified are functionally involved in the pathogenesis of Alzheimer’s disease. One of the annotated genes for the novel SNP rs3777215 was RHOBTB3. Reported as a candidate Alzheimer’s disease vulnerability gene through transcription analysis, this gene was found to be overexpressed in the CA1 region following Alzheimer’s disease progression (Miller et al., 2013). RHOBTB3 encodes Rho-related BTB domain-containing protein 3 (RHOBTB3), which is involved in endosome-to-Golgi transport and retrograde transport (Espinosa et al., 2009). Thus, RHOBTB3 may affect APP processing and provide a pathological basis for the development of Alzheimer’s disease. The other annotated gene for the novel SNP rs3777215 was GLRX, which encodes glutaredoxin-1 (GRX1), an oxidoreductase that contributes greatly to the antioxidant defence system and internal environment homeostasis. Functionally, GRX1 and thioredoxin-1 (TRX1) are antioxidants and their reduced forms can inhibit apoptosis signal regulating kinase (ASK1). It has been reported that amyloid-β can oxidize GRX1 and TRX1, resulting in apoptosis induction via ASK1 (Akterin et al., 2006). Another study found that increasing GRX1 levels in the brain of an Alzheimer’s disease mouse model can reverse synaptic dysfunction and cognitive deficits, suggesting GRX1 as a target for Alzheimer’s disease intervention (Kommaddi et al., 2019). For the novel SNP rs2255835, the annotated gene was CHODL. This gene encodes chondrolectin (CHODL), which is involved in the endocytosis of glycoproteins and exogenous sugar-bearing pathogens (Zelensky and Gready, 2005). CHODL affects cell survival and neuronal outgrowth in animal models (Sleigh et al., 2014). Interestingly, in some early-onset patients with Alzheimer’s disease induced by APP duplication, the duplicated region also contains CHODL, as well as eQTL-associated genes of rs2255835, such as BTG3, C21orf91, and TMPRSS15, which may participate in neurogenesis and/or APP metabolism (McNaughton et al., 2012; Wiseman et al., 2015).

Using different combinations of variants identified in the current study, we established 11 predictive models. Building predictive models based on GWAS data to distinguish asymptomatic population at a high-risk of developing Alzheimer’s disease from ‘normal’ individuals has gained attention (Escott-Price et al., 2015, 2017a, b, 2019; Chouraki et al., 2016; Stocker et al., 2018), and some recent studies have focused on predicting the conversion from mild cognitive impairment to Alzheimer’s disease via models based on genetic factors (Lacour et al., 2017; Chaudhury et al., 2019). Escott-Price and colleagues demonstrated a prediction accuracy of 75–84% for Alzheimer’s disease risk with certain predictors (APOE, polygenic risk score calculated from more than 20 000 SNPs, sex, and age) (Escott-Price et al., 2015). In addition, Sultan and colleagues reported that the prediction accuracy of their models, with APOE SNPs (rs7412 and rs429358), 165 non-APOE SNPs, sex, and age as predictors was 82.5% for predicting the conversion from mild cognitive impairment to Alzheimer’s disease (Chaudhury et al., 2019). However, most of these models included too many SNPs, preventing their use in clinics. Therefore, we established predictive models to determine the risks of the possibility of developing Alzheimer’s disease. In our A models, models A1 and A2 may be easier to use in clinics, as the number of SNPs in model A1 or model A2 was smaller than that in model A3, and the AUC of model A1 or model A2 was similar to that of model A3. However, these models are no longer suitable for populations negative for APOE ε4 comprising a more important group. Thus, we constructed models B1–B7. These models incorporate fewer SNPs and show significant AUC values. In the B models, the AUC of model B4 was similar to that of model B2, but the number of SNPs in model B4 was approximately half of that of model B2. Therefore, models B4–B7 are recommended as more of the population can be covered by these models in clinical practice. Overall, the 11 predictive models appear to be useful for identifying the indications of Alzheimer’s disease risk in the sectional datasets. To confirm the capacity of the models to predict Alzheimer’s disease, we performed survival curve analysis on a longitudinal cohort. The results showed that individuals carrying risk variants included in either model A or model B had a shorter latency and higher incidence of Alzheimer’s disease, suggesting that our models can predict Alzheimer’s disease onset in a population with genetic susceptibility. The mechanism of this genetic susceptibility requires further analysis.

The current study had some limitations. First, the novel variants we identified are involved in the pathogenesis of Alzheimer’s disease based on bioinformatic analysis and literature mining, but we did not conduct functional research on these novel variants. However, validation of the predictive models indicated the contributions of these variants to sporadic Alzheimer’s disease development. Second, the prediction accuracies of our models were relatively low compared to those of other predictive models but the models were verified to be effective and accessible for predicting Alzheimer’s disease onset based on a 10-year longitudinal cohort and sectional datasets. Third, although the longitudinal study (COAST) we used was prospective, the use of current diagnoses of individuals was compared with the initial condition at baseline and was used to validate our predictive model and to confirm the accuracy and effectiveness of our models to predict Alzheimer’s disease. This is rational because the DNA of individuals at baseline may reflect the true genetic conditions. Fourth, the significant SNPs in the discovery stage, which were not replicated in the second stage, did not pass the heterogeneity test. This may be because of differences in the genetic background of the Chinese population (Supplementary Table 18), which is supported by other studies (Chen et al., 2009; Tan et al., 2013; Tao et al., 2017; Wang et al., 2018). Finally, despite the power calculation indicating that our sample size was sufficient to detect associations, larger sample sizes based on Chinese populations are required in future studies, as well as analyses of populations of other ethnicities, to generate more reliable results.

In conclusion, we identified four novel susceptibility variants for Alzheimer’s disease, improving the understanding of the genetic predisposition to Alzheimer’s disease. Annotated genes of these variants were related to APP metabolism, antioxidation, and neurogenesis. The contributions of these variants to sporadic Alzheimer’s disease development were confirmed to be efficient based on validation of the predictive models in a longitudinal study. This is the first study to validate GWAS-based predictive models for evaluating the risk of Alzheimer’s disease onset in a Chinese population. The clinical application of these models is of potential use for individuals harbouring these risk variants but must be validated in a larger population.

Glossary

AUC =

area under the curve

eQTL =

expression quantitative trait loci

GWAS =

genome-wide association study

SNP =

single nucleotide polymorphism