The 759 participants were enrolled from the Korea-Registries to Overcome and Accelerate Dementia research project (K-ROAD). The K-ROAD aims to develop a genotype–phenotype cohort to accelerate the development of novel diagnostic and therapeutic techniques for Alzheimer’s and concomitant cerebrovascular diseases. Nationwide, 25 university-affiliated hospitals in South Korea are participating in the K-ROAD. All participants underwent neuropsychological tests, brain magnetic resonance imaging (MRI), Aβ PET (¹⁸F-flutemetamol (FMM)), APOE genotyping, and microarray genotyping data. The 759 participants consist of Alzheimer’s disease (AD; n = 246), amnestic mild cognitive impairment (aMCI; n = 255), and cognitively unimpaired (CU) individuals (n = 258). All participants with CU met the following criteria: (1) no medical history that was likely to affect cognitive function based on Christensen’s health screening criteria; (2) no objective cognitive impairment in any cognitive domain on a comprehensive neuropsychological test battery (at least −1.0 SD above age-adjusted norms on any cognitive test); and (3) independence in daily living activities. All participants with MCI met the criteria for MCI with the following modifications (Albert et al., 2011); (1) subjective cognitive complaints by the participants or caregivers; (2) objective memory impairment below −1·0 SD on verbal or visual memory tests; (3) no significant impairment in daily living activities; and (4) non-demented status. Participants with dementia met the core clinical criteria of probable AD dementia proposed by the National Institute on Aging-Alzheimer’s Association (NIA-AA; McKhann et al., 2011).

Participants with significant white matter hyperintensities (cap or band >10 mm and longest diameter of deep white matter lesion >25 mm), structural lesions, including cerebral infarction, intracranial haemorrhage, brain tumours, and hydrocephalus on MRI, and abnormal laboratory results on complete blood count, electrolyte, vitamin B12, and folate levels, syphilis serology, and liver, kidney, or thyroid function tests were excluded from the study.

The institutional review board of the Samsung Medical Center approved this study. Written informed consent was obtained from all participants.

Genotyping was performed using the Illumina Asian Screening Array BeadChip (Illumina, CA, United States). The quality control (QC) procedures were conducted using PLINK software, and samples that did not satisfy the following criteria were excluded: call rate < 95%, sex-mismatch, excess heterozygosity rate (5 standard deviation from the mean), and identify-by-descent ≥0.125. Additionally, individual markers with the following criteria were excluded: call rate < 98%, minor allele frequency (MAF) < 1%, Hardy–Weinberg equilibrium (HWE) p < 10⁻⁶. Following QC, un-genotyped markers were imputed using Minimac4 and reference haplotypes from HRC-r1.1 on the University of Michigan Imputation Server (Das et al., 2016). Furthermore, post-imputation QC was performed using the following criteria: poor imputation, r² ≤ 0.8 and MAF < 1%. Finally, bi-allelic 4,906,407 SNPs in autosomal chromosomes (sex chromosome, mitochondrial, and pseudo autosomal SNPs were excluded) were used for the following analyses.

All participants underwent Aβ PET (¹⁸F-flutemetamol PET) scans using a Discovery STe PET/CT scanner (GE Medical Systems, Milwaukee, WI, United States). For ¹⁸F-flutemetamol PET, a 20-min emission PET scan in dynamic mode (consisting of 4 × 5 min frames) was performed 90 min after an injection of a mean dose of 311.5 MBq ¹⁸F-florbetaben or 197.7 MBq ¹⁸F-flutemetamol. Three-dimensional PET images were reconstructed in a 128 × 128 × 48 matrix with 2 mm × 2 mm × 3·27 mm voxel size using the ordered-subsets expectation maximisation algorithm (iteration = 4 and subset = 20).

For image processing, we used SPM8¹ running on MATLAB.² T1-weighted images were corrected for nonuniformity (Sled et al., 1998) and normalised to standard space using a linear transformation. PET images were co-registered, and a Gaussian kernel with an 8 mm full width at half maximum was used for spatial smoothing. Using the whole cerebellum as a reference region, the standardised uptake value ratio (SUVR) images were calculated. The mask of the reference region was obtained from the Global Alzheimer’s Association Information Network (GAAIN) websites³ (Klunk et al., 2015). The mean SUVR of cerebral cortical areas was determined using the masks of the standard cortical target region (Klunk et al., 2015) downloaded from GAAIN.

We acquired standardised three-dimensional T1 Turbo Field Echo and three-dimensional fluid-attenuated inversion recovery (FLAIR) images using a 3.0 T MRI scanner (Philips 3.0 T Achieva; Philips Healthcare, Andover, MA, United States), as previously described (Kang et al., 2021).

Images were processed using the CIVET anatomical pipeline (version 2.1.0). The native MRI images were registered to the MNI-152 template by a linear transformation and corrected for intensity nonuniformities using the N3 algorithm (Sled et al., 1998). We used an automated hippocampus segmentation method described in an earlier study that combined a graph cut algorithm with atlas-based segmentation and morphological opening to measure hippocampal volume (Kwak et al., 2013).

To assess the effect of genes on Aβ SUVR, a gene-based association analysis was performed by combining the SNP-based p-values from a GWAS, which was performed using the PLINK software (Purcell et al., 2007). The statistical analysis was performed on 759 participants from three diagnostic groups (CU, aMCI, and AD) and allelic effects for the Aβ SUVR were measured using an additive model with age, sex, and APOE ε4 counts as covariates. Gene-based association analysis was conducted using the extended Simes procedure (GATES; Li et al., 2011). SNPs that mapped within 20 kb of the 3′ and 5′ untranslated regions were considered “within” the genes. The linkage disequilibrium (LD) was calculated using data for East Asian ancestry from the 1,000 Genomes Project, and SNPs with r² < 0.005 were ignored in gene annotation.

We performed an eQTL analysis to determine the functional effect of genetic variants on gene expression. Genotype data and exon-specific expression in the hippocampus, frontal, and temporal cortex were downloaded from the Braineac database⁴ (Ramasamy et al., 2014). The eQTL analysis was performed with the most significant SNPs mapped to identified genes from gene-based association analysis. Multiple comparison correction for the number of tissues was performed, and eQTLs with q < 0.05 were considered significant.

We hypothesised that the identified genetic variants were related to biomarkers that indicate the severity of disease symptoms. The effects of the most significant SNPs mapped to identified genes on cognitive scores [Mini-Mental State Exam (MMSE) and Clinical Dementia Rating Sum of Boxes (CDR-SB)] and hippocampal volume were investigated. After applying inverse normal transformations to cognitive scores for data normality, associations were tested, and age, sex, education, and ICV volume were used as covariates, if appropriate. In addition, the mediation package in R was used for mediation analysis in order to identify the SUVR-mediated pathway from SNP to AD biomarkers (Imai et al., 2010). The genotypes were defined as independent variables, the SUVR as a mediator variable, and the clinical outcomes as continuous dependent variables. The mediate() function generated the direct and indirect effects as well as paths from the mediator variable to dependent variables by linear regression.

We performed pathway analysis using GSA-SNP (Nam et al., 2010) to identify functional gene sets associated with Aβ SUVR. The pathway annotations from the Gene Ontology (GO) resource⁵ were downloaded, and gene sets containing 5–200 genes were used for analysis. The summary statistics of GWAS were used, and SNPs that fell within 20 kb of the boundary of a gene were annotated to the gene on the human genome (hg19) coordinate. Gene sets were evaluated by Z-statistics for the identification of enriched gene sets, with gene sets having q < 0.05 considered significant.

We generated a protein–protein interaction network using the WEB-based gene set analysis toolkit (Liao et al., 2019) with the human PPI data from the Biological General Repository for Interaction Datasets. For the seed genes identified in gene-based association analysis, random walk analysis was performed to expand the network. The network was expanded by ranking genes based on their network proximity with the seed genes, and the resulting network consisted of the seed genes and the top 50 neighbouring genes. Functional module detection was performed using HumanBase (Greene et al., 2015) to find the functional gene cluster in human brain tissue. Genes were clustered into the functional modules with tissue-specific networks based on the shared k-nearest nearest-neighbours (SKNN) and the Louvain community-finding algorithm. Functional enrichment analysis was performed for the resulting functional modules using Go terms. Statistical significance was tested by a one-sided Fisher’s exact test and Go terms with q < 0.05 were considered significant.

The demographics and genotypic characteristics of the final samples are listed in Table 1. The age (mean [±standard deviation]) of the participants was 70.6 (±8.5) years; the proportions of female and Aβ + participants were 58.6 and 49.4%, respectively. The proportions of apolipoprotein ε4 carriers were 38.1%.

The considerable association of APOE4 with Aβ SUVR led us to conduct GWAS with APOE ε4 counts as a covariate, where no SNP passed the genome-wide significance threshold of 5 × 10⁻⁸ (Supplementary Figure S1). Ten SNPs on chromosomes 13, 16, 17, and 21 showed genome-wide suggestive significance (p < 1 × 10⁻⁶) and there is no genomic inflation (λ = 1.0).

In gene-based association analysis, 3,936,415 SNPs mapped to 26,443 genes on the human genome. One gene on chromosome 16, LCMT1 (Leucine Carboxyl Methyltransferase 1), and five genes on chromosome 17, SCRN2 (Secernin 2), LRRC46 (Leucine Rich Repeat Containing 46), MRPL10 (Mitochondrial Ribosomal Protein L10), SP6 (Sp6 transcription factor), and OSBPL7 (Oxysterol Binding Protein Like 7), were significantly associated with the Aβ SUVR (Figure 2). SCRN2 on chromosome 17 showed the strongest associations with Aβ SUVR (q = 2.33 × 10⁻²). Four additional genes, LRRC46, MRPL10, SP6, and OSBPL7, close to SCRN2, were found to be related to Aβ SUVR (p = 2.46 × 10⁻⁶, p = 2.64 × 10⁻⁶, p = 5.39 × 10⁻⁶, p = 6.71 × 10⁻⁶, respectively). Furthermore, the LCMT1 gene on chromosome 16 was observed to be significantly related to Aβ SUVR (p = 5.90 × 10⁻⁶).

The genetic effects of the most significant SNPs mapped to the identified six genes are shown in Figures 3A–C. The rs4787307 on the LCMT1 gene has a risk effect on Aβ SUVR (β = 0.258, p = 3.91 × 10⁻⁷), which means that participants with minor alleles have a higher Aβ SUVR than those without minor alleles. The rs9903904 mapped to SCRN2, LRRC46, and SP6 genes, and the rs11079797 mapped to MRPL10 and OSBPL7 also have risk effects (β = 0.655, p = 9.73 × 10⁻⁸ and β = 0.623, p = 1.14 × 10⁻⁷, respectively).

To identify the functional effect of genetic variants on gene expression, cis-eQTL analysis was performed. We performed eQTL analysis with three SNPs (rs4787307, rs9903904, and rs11079797) as well as the expression levels of the mapped genes in the frontal and temporal cortex and hippocampus (Figures 3D–G). Our findings indicated that rs4787307 on chromosome 16 significantly regulates the expression of LCMT1 in the temporal cortex (q = 4.20 × 10⁻²), while rs11079797 on chromosome 17 significantly regulates the expression of MRPL10 (q = 6.90 × 10⁻³) and OSBPL7 (q = 4.80 × 10⁻²) in the frontal cortex and hippocampus. In addition, rs9903904 on chromosome 17 influences the expression of SCRN2 in the hippocampus (q = 4.80 × 10⁻²).

We tested the association between the most significant SNPs on six genes and cognitive scores and hippocampal volume (Table 2). The association tests are shown in Table 2. The rs4787307 in LCMT1 was significantly associated with CDR-SB (β = 0.129, p = 2.25 × 10⁻²), left (β = −93.19, p = 6.39 × 10⁻³) and right hippocampal volume decline (β = −106.51, p = 1.00 × 10⁻²). The rs11079797 in MRPL10 and OSBPL7 was significantly associated with MMSE (β = −0.372, p = 2.61 × 10⁻³), left (β = −181.2, p = 1.90 × 10⁻²) and right hippocampal volume decline (β = −164.56, p = 3.83 × 10⁻²). In addition, rs9903904 in SCRN2, LRRC46, and SP6 was significantly associated with MMSE (β = −0.453, p = 4.74 × 10⁻⁴), CDR-SB (β = 0.321, p = 1.48 × 10⁻²), left (β = −213.2, p = 8.72 × 10⁻³), and right hippocampal volume (β = −203.45, p = 2.34 × 10⁻²).

To identify the Aβ uptake-mediated pathways from SNP to clinical outcomes, we performed mediation analysis with SUVR as the mediator variable and clinical outcomes as dependent variables (Figure 4; Supplementary Table S1). Three SNPs had significant indirect effects when CDR-SB was set as the outcome (rs4787307, effect = 0.08, p < 0.001; rs11079797, effect = 0.21, p < 0.001; and rs9903904, effect = 0.22, p < 0.001; respectively), while only rs9903904 had a significant direct effect (effect = 0.1, p < 3.65 × 10⁻²). In addition, when left hippocampal volume was set as the outcome, we found significant indirect effects of three SNPs (rs4787307, effect = −60.54, p < 0.001; rs11079797, effect = −124.94, p < 0.001; rs9903904, effect = −130.63, p < 0.001; respectively), but the direct effects were not significant. The mediation results for MMSE and right hippocampal volume are presented in Supplementary Table S1. These findings indicate that the SNP could affect hippocampal volume and cognitive scores, primarily through the mediation of Aβ SUVR.

To identify functional gene sets associated with Aβ SUVR, we performed pathway analysis using SNP p-values from GWAS. The 6,432 gene sets have been examined, and 19 gene sets were enriched for Aβ SUVR (Figure 5). Gene sets related to axon parts, cytokine, and chemokine activity were enriched. More detailed results and a list of genes contained in the gene set are presented in Supplementary Table S2.

We constructed a network consisting of six seed genes identified in the gene-based association analysis and 50 top-ranking neighbour genes (Supplementary Figure S2). The 56 genes in the constructed PPI network were clustered into five functional modules in the cerebral cortex. Modules 1, 2, 3, 4, and 5 included 6, 10, 6, 8, and 10 genes, respectively (Supplementary Figure S3; Supplementary Table S3). In enrichment analysis, eight genes clustered into M1 were enriched for DNA repair, recombination-related processes, and dephosphorylation (Table 3). M2 and M4 genes were enriched for cell cycle-related processes, and M5 was enriched for the proteolysis-related process.