Hong et al. Translational Psychiatry (2020)10:403 https://doi.org/10.1038/s41398-020-01074-z Translational Psychiatry ARTICLE Open Access Genome-wide association study of Alzheimer’s disease CSF biomarkers in the EMIF-AD Multimodal Biomarker Discovery dataset Shengjun Hong1, Dmitry Prokopenko2, Valerija Dobricic1, Fabian Kilpert1, Isabelle Bos3,4, Stephanie J. B. Vos3, Betty M. Tijms 5, Ulf Andreasson6,7, Kaj Blennow 6,7, Rik Vandenberghe8,9, Isabelle Cleynen10, Silvy Gabel8, Jolien Schaeverbeke8, Philip Scheltens 5, Charlotte E. Teunissen11, Ellis Niemantsverdriet12, Sebastiaan Engelborghs 12,13, Giovanni Frisoni14,15, Olivier Blin16, Jill C. Richardson17, Regis Bordet18, José Luis Molinuevo19, Lorena Rami19, Alzheimer’s Disease Neuroimaging Initiative (ADNI), Petronella Kettunen 6,20, Anders Wallin 6, Alberto Lleó21, Isabel Sala21, Julius Popp22,23, Gwendoline Peyratout23, Pablo Martinez-Lage24, Mikel Tainta24, Richard J. B. Dobson 25,26,27,28,29, Cristina Legido-Quigley30,31, Kristel Sleegers 32,33, Christine Van Broeckhoven32,33, Mara ten Kate34,35, Frederik Barkhof36, Henrik Zetterberg 6,7,37,38, Simon Lovestone39, Johannes Streffer40,41, Michael Wittig42, Andre Franke 42, Rudolph E. Tanzi 2, Pieter Jelle Visser 5 and Lars Bertram 1,43 Abstract Alzheimer’s disease (AD) is the most prevalent neurodegenerative disorder and the most common form of dementia in the elderly. Susceptibility to AD is considerably determined by genetic factors which hitherto were primarily identified using case–control designs. Elucidating the genetic architecture of additional AD-related phenotypic traits, ideally those linked to 1234567890():,; 1234567890():,; 1234567890():,; 1234567890():,; the underlying disease process, holds great promise in gaining deeper insights into the genetic basis of AD and in developing better clinical prediction models. To this end, we generated genome-wide single-nucleotide polymorphism (SNP) genotyping data in 931 participants of the European Medical Information Framework Alzheimer’s Disease Multimodal Biomarker Discovery (EMIF-AD MBD) sample to search for novel genetic determinants of AD biomarker variability. Specifically, we performed genome-wide association study (GWAS) analyses on 16 traits, including 14 measures derived from quantifications of five separate amyloid-beta (Aβ) and tau-protein species in the cerebrospinal fluid (CSF). In addition to confirming the well-established effects of apolipoprotein E (APOE) on diagnostic outcome and phenotypes related to Aβ42, we detected novel potential signals in the zinc finger homeobox 3 (ZFHX3) for CSF-Aβ38 and CSF-Aβ40 levels, and confirmed the previously described sex-specific association between SNPs in geminin coiled-coil domain containing (GMNC) and CSF- tau. Utilizing the results from independent case–control AD GWAS to construct polygenic risk scores (PRS) revealed that AD risk variants only explain a small fraction of CSF biomarker variability. In conclusion, our study represents a detailed first account of GWAS analyses on CSF-Aβ and -tau-related traits in the EMIF-AD MBD dataset. In subsequent work, we will utilize the genomics data generated here in GWAS of other AD-relevant clinical outcomes ascertained in this unique dataset. Correspondence: Lars Bertram (lars.bertram@uni-luebeck.de) 1 Lübeck Interdisciplinary Platform for Genome Analytics (LIGA), Institutes of Neurogenetics and Cardiogenetics, University of Lübeck, Lübeck, Germany Introduction 2 Genetics and Aging Unit and McCance Center for Brain Health, Department of Alzheimer’s disease (AD) is a progressive and devas- Neurology, Massachusetts General Hospital, Boston, MA, USA tating neurodegenerative disorder, which leads to cogni- Full list of author information is available at the end of the article tive decline, loss of autonomy, dementia, and eventually Alzheimer’s Disease Neuroimaging Initiative (ADNI): Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging death. Neuropathologically, AD is characterized by the Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within accumulation of extracellular amyloid β (Aβ) peptide the ADNI contributed to the design and implementation of ADNI and/or deposits (“plaques”) and intracellular hyperpho- provided data but did not participate in analysis or writing of this report. A complete listing of ADNI investigators can be found at: http://adni.loni.usc. sphorylated tau protein aggregates (“tangles”) in the edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf. © The Author(s) 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/. Hong et al. Translational Psychiatry (2020)10:403 Page 2 of 12 brain1,2. Using genetic linkage analysis followed by posi- In this report, we focus exclusively on the description of tional cloning led to the discovery of rare mutations in the results from genome-wide association analyses using three genes encoding the amyloid-beta precursor protein various Aβ and tau-relevant outcomes available in EMIF- (APP) and presenilins 1 and 2 (PSEN1, PSEN2) that cause AD MBD. Specifically, we performed GWAS and poly- fully penetrant monogenic forms of AD3. However, the genic risk score (PRS) assessments for more than a dozen vast majority of patients likely suffer from a polygenic binary and quantitative phenotypes derived from five (“sporadic”) form of AD, which is driven by numerous measures of cerebrospinal fluid (CSF) Aβ and tau proteins genomic variants4, the identification of which are the in addition to using simple diagnostic status (i.e., AD, main aim of genome-wide association studies (GWAS). MCI, and control). Whenever available, we compare our The most strongly and most consistently associated AD findings using equivalent GWAS results from the ADNI risk gene (even prior to GWAS5) is APOE, which encodes dataset. apolipoprotein E, a cholesterol transport protein that has been implicated in numerous amyloid-specific pathways, Materials and methods including amyloid trafficking, as well as plaque clear- Sample and phenotype description ance2,6. In addition to APOE, nearly three dozen inde- Overall, the EMIF-AD MBD dataset comprises 1221 pendent loci have now been reported to be associated elderly individuals (years of age: mean = 67.9, SD = 8.3) with disease risk by GWAS7–9. Pathophysiologically, the with different cognitive diagnoses at baseline (NC = nor- risk genes identified to date appear to predominantly act mal cognition; MCI = mild cognitive impairment; AD = through modulations of the immune system response, AD-type dementia). In addition, Aβ status, cognitive test endocytotic mechanisms, cholesterol homeostasis, and results and at least two of the following were available at APP catabolic processes7,8. baseline for analyses in all EMIF-AD MBD individuals: Despite these general advances in the field of AD plasma (n = 1189), DNA (n = 929), magnetic resonance genetics, many key questions still remain to be answered. imaging (MRI; n = 862), or CSF (n = 767 individuals). First, even when analyzed in combination, the currently Furthermore, clinical follow-up data were available for known AD risk factors explain only a fraction of the 759 individuals. The demographic information of the 16 phenotypic variance7, and, accordingly, only have limited outcome phenotypes (9 binary and 7 quantitative) of the applicability as early markers for disease onset and pro- EMIF-AD MBD dataset utilized in this paper is sum- gression7,10. Second, most of the currently reported AD marized in Table 1. Depending on the availability of the susceptibility genes were identified using classic clinical records, each phenotype has different effective case–control designs comparing clinically manifest sample sizes. We categorized the phenotypes analyzed in dementia-stage AD vs. control individuals, typically lack- this study into three main categories, i.e., “diagnosis”, ing data on early-stage impairments (e.g., mild cognitive “amyloid protein assessment” and “tau protein assess- impairment [MCI]) and clinical follow-up to ascertain ment” (NB: the diagnostic criteria used here for AD are progression and eventually conversion to AD. Finally, not incorporating biomarker status so that some AD cases while some studies have investigated the correlation were classified as “amyloid negative”). Details related to between genetics and non-genetic biomarkers, this was sample ascertainment and phenotype/biomarker collec- hitherto typically done as bivariate assessments owing to tion in EMIF-AD MBD have been described previously11 the lack of a broad spectrum of biomarkers and imaging and are summarized for the relevant traits of this study in data in the same individuals. To overcome at least some of Supplementary Table 1. Whenever available, we attemp- these shortcomings we generated genome-wide single- ted to validate EMIF-AD MBD findings in the indepen- nucleotide polymorphism (SNP) genotyping data in the dent ADNI dataset using identical or comparable European Medical Information Framework Alzheimer’s phenotypes (this was possible for the two diagnostic Disease Multimodal Biomarker Discovery (EMIF-AD groups, as well as for 6 amyloid- and 2 tau-related traits; MBD) sample11. This powerful and unique dataset allows Table 1). The local medical ethical committee in each to combine genomic data (and “-omics” data from other participant recruitment center approved the study. Sub- domains) with preclinical biomarker levels to eventually jects had provided written informed consent at the time of improve our ability for an early detection and prevention inclusion in the cohort for use of data, samples and of AD. While similar to the Alzheimer’s Disease Neuroi- scans11. maging Initiative (ADNI) study12 in various aspects, Replication data used in the preparation of this article EMIF-AD MBD extends ADNI and scope in several were obtained from the ADNI database (adni.loni.usc. important ways, e.g., in the breadth of the biomarker edu). The ADNI was launched in 2003 as a public-private assessments as well as the availability of “-omics” data partnership, led by Principal Investigator Michael W. from various different domains in the same individuals Weiner, MD. The primary goal of ADNI has been to test (for more details see Bos et al.11). whether serial MRI, positron emission tomography (PET), Table 1 Overview of binary and quantitative traits available for genome-wide association study (GWAS) and polygenic risk score (PRS) analyses in EMIF-AD MBD and ADNI datasets. EMIF-AD MBD (Discovery) ADNI (Replication) Category Variable type Variable Description #sample_n GWAS PGS Variable #sample_n GWAS Clinical diagnosis Binary AD vs. NC Alzheimer disease (AD) vs. normal 545 (212 AD vs. YES YES AD vs. NC 303 (46 AD vs. YES cognition (NC) 333 NC) 257 NC) MCI vs. NC Mild cognitive impairment (MCI) vs. 659 (326 MCI vs. YES YES MCI vs. NC 705(448 MCI vs. YES normal cognition (NC) 333 NC) 257 NC) Amyloid protein Binary AMYLOIDstatus_ALL Dichotomous amyloid classification 871 (455 abnormal YES YES AMYLOIDstatus 618 (361 abnormal YES assessment variable across all diagnostic groups vs. 416 normal) vs. 257 normal) AMYLOIDstatus_MCI Dichotomous amyloid classification 326 (189 abnormal YES YES AMYLOIDstatus_MCI 371 (232 abnormal YES Hong et al. Translational Psychiatry (2020)10:403 variable in MCI subjects vs. 137 normal) vs. 139 normal) AMYLOIDstatus_NC Dichotomous amyloid classification 333 (77 abnormal vs. YES YES AMYLOIDstatus_NC 202 (88 abnormal vs. YES variable in NC subjects 256 normal) 114 normal) Central_CSF_ratiodich Dichotomous variable based on ratio of 677 (418 abnormal YES YES NA NA NA central CSF amyloid-42/40 values vs. 259 normal) Local_AB42_Abnormal Dichotomous variable of local CSF 726 (392 abnormal YES YES AB42_abnormal 578 (340 abnormal YES amyloid-beta-42 values vs. 334 normal) vs. 238 normal) Quantitative AB_Zscore Z-score for amyloid pathology 890 YES YES AB_Zscore 578 YES log_Central_CSF_AB42 Log-transformed central CSF amyloid- 677 YES YES ABETA.Lumi.bl 578 YES beta-42 values Central_CSF_AB38 Central CSF Amyloid-beta-38 values 675 YES YES ABETA38.MSM.bl 548 YES Central_CSF_AB40 Central CSF Amyloid-beta-40 values 677 YES YES ABETA40.MSM.bl 548 YES log_Central_CSF_AB4240ratio Ratio of log-transformed central CSF 677 YES YES NA NA NA amyloid-beta-42 vs. 40 values Tau protein Binary Local_TTAU_Abnormal Dichotomous variable of local CSF total- 724 (378 abnormal YES YES NA NA NA assessment tau values vs. 346 normal) Local_PTAU_Abnormal Dichotomous variable of local CSF 726 (354 abnormal YES YES NA NA NA phospho-tau values vs. 372 normal) Quantitative Ttau_ASSAY_Zscore Z-score for CSF total-tau values 723 YES YES log_TTAU_Zscore 571 YES Ptau_ASSAY_Zscore Z-score for CSF phospho-tau values 726 YES YES log_PTAU_Zscore 576 YES Page 3 of 12 Hong et al. Translational Psychiatry (2020)10:403 Page 4 of 12 other biological markers, and clinical and neuropsycho- Classification of APOE genotypes logical assessment can be combined to measure the pro- For all but 80 samples APOE genotype (i.e., for SNPs gression of mild MCI and early AD. The ADNI rs7412 [a.k.a. as “ε2-allele”] and rs429358 [a.k.a. “ε4- participants utilized for our analyses originate from both allele”]) was determined locally at the sample collection ADNI1 and ADNIGo/2 and relate to those with available sites. To ensure that these prior genotypes correctly align whole-genome sequencing (WGS; see below) data. to those resulting from genome-wide genotyping, local Accordingly, we label the subset of ADNI participants APOE genotypes were compared to those either inferred utilized here as “ADNI-WGS”. directly (i.e., rs7412) or indirectly (i.e., by imputation: rs429358) from GSA genotyping. These comparisons DNA extraction resulted in a total of 5 mismatches (~0.6%). In these 5 and Our laboratory at University of Lübeck, Germany, had the 80 samples without prior APOE genotype information, access to 953 DNA samples from EMIF-AD MBD partici- genotyping was determined manually in our laboratory pants11 for genetic (this paper) and epigenetic (DNA using TaqMan assays (ThermoFisher Scientific, Foster methylation profiling, m.s. in preparation) experiments. All City, CA) on a QuantStudio-12K-Flex system in 384-well participants had provided written consent to these experi- format. TaqMan re-genotyping confirmed all five local ments and institutional review board (IRB) approvals for the genotype calls (which were used as genotypes in all sub- utilization of the DNA samples in the context of EMIF-AD sequent analyses). MBD were obtained by the sample collection sites. For 805 participants, DNA was extracted locally at the collection Biochemical analyses of CSF biomarkers sites. For 148 whole-blood samples DNA extraction was CSF sampling and storage conditions have been performed in our laboratory using the QIAamp® DNA described elsewhere11. CSF concentrations of Aβ38, Aβ40 Blood Mini Kit (QIAGEN GmbH, Hilden, Germany). and Aβ42 were measured using the V-PLEX Plus Aβ Overall, this resulted in a total number of 953 DNA samples Peptide Panel 1 (6E10) Kit from Meso Scale Discovery available for subsequent processing and analysis. Quality (MSD, Rockville, MD). The measurements were per- control (QC; by agarose gel electrophoresis, determination formed at the Clinical Neurochemistry Laboratory in of A260/280 and A260/230 ratios, and PicoGreen quanti- Gothenburg in one round of experiments, using one batch fication) resulted in 936 DNA samples of sufficient quality of kit reagents, by board-certified laboratory technicians, and quantity to attempt genome-wide SNP genotyping who were blinded to clinical data. For phosphorylated tau using the Infinium Global Screening Array (GSA) with (Ptau) and total tau (Ttau), available data from the local Shared Custom Content (Illumina Inc.). GSA genotyping cohorts were used. These were derived in clinical was performed at the Institute of Clinical and Medical laboratory practice using INNOTEST ELISAs (Fujirebio, Biology (UKSH, Campus-Kiel) on an iScan instrument Ghent, Belgium), as previously described17. In the absence (Illumina, Inc) following the manufacturer’s recommenda- of CSF for new analyses of CSF Aβ proteins, we used local tions. All 936 DNA samples passed post-experiment QC INNOTEST ELISA-derived CSF Aβ42 data to allow according to the manufacturer’s instructions. classifying as many subjects as possible as either Aβ-positive or -negative (see ref. 11 and below). Genotype imputation and quality control Data processing was performed from raw intensity data GWAS and post-GWAS analyses (idat format) in GenomeStudio software (v2.0.4; Illumina, SNP-based association tests were performed using Inc.). We then used PLINK software (v1.9)13 to perform logistic regression models in mach2dat18,19, for binary pre-imputation QC and bcftools (v1.9)14 to remove traits and linear regression models in mach2qtl18,19, for ambiguous SNPs, flipping and swapping alleles to align to quantitative traits. Association analyses utilized human genome assembly GRCh37/hg19 before imputa- imputation-derived allele dosages as independent vari- tion. The QC’ed data (i.e., 931 samples and 498 589 SNPs) ables and were adjusted for sex, age at examination, and were then phased using SHAPEIT2 (v2.r837)15 and principle components (PC) 1 to 5 (using PLINK –pca to imputed locally using Minimac316 based on a pre- compute eigenvalues for up to 20 PCs; the number of PCs compiled Haplotype Reference Consortium (HRC) refer- was then determined visual inspection of the scree plot). ence panel (EGAD00001002729 including 39,131,578 Diagnostic groups (coded as AD = 3, MCI = 2, controls = 1) SNPs from ~11 K individuals). Following post-imputation were included as additional covariates in all analyses QC, we retained a total of 7,778,465 autosomal SNPs with except for diagnostic outcome. QQ and Manhattan plots minor allele frequency (MAF) ≥ 0.01 in 898 individuals of were constructed in R version 3.3.3 (https://www.r- European ancestry for downstream association analysis. A project.org) using the “qqman” package20. The genomic full description of data processing and QC procedures is inflation factor was calculated in R using the “GenABEL” provided in the Supplementary Material. package21. Statistical significance for the SNP-based Hong et al. Translational Psychiatry (2020)10:403 Page 5 of 12 analyses was defined as α = 5E-08, a widely used threshold Weinberg equilibrium (P > 1E-05). We calculated princi- that accounts for the approximate number of independent pal components to account for population stratification variants (~1 M) in European populations22,23. Post- based on an LD-pruned subset of common variants GWAS, we used FUMA (http://fuma.ctglab.nl/)24 to (MAF > 0.1). Association statistics were calculated using perform functional mapping and annotation of the PLINK v2.0 using linear and logistic regression models (as genome-wide association results. This included calculat- appropriate), controlling for age, sex and four principal ing gene-based association statistics using MAGMA25 components as basic covariates in our models. For the using predefined sets of genes as implemented in FUMA. phenotypes not representing the diagnostic outcome, we Statistical significance for the gene-based analyses was included diagnosis as an additional covariate. For more defined as α = 0.05/18720 = 2.671E-06 based on the information on ADNI, please see Supplementary Material. number of genes (n = 18720) utilized for these analyses, as suggested by FUMA24. Results GWAS on diagnostic outcomes Polygenic risk score (PRS) analysis First, we performed GWAS for diagnostic outcomes, PRS were calculated for each individual from the sum- i.e., all cases diagnosed with AD (n = 212) and MCI (n = mary statistics of two partially overlapping AD 326), against normal control subjects (n = 333). Com- case–control GWAS, i.e., the paper by Jansen et al.7 paring AD vs. controls showed the expected strong signals including data from >380,000 individuals from the UK in the APOE region on chromosome 19q reaching biobank, and a 2013 GWAS meta-analysis from the genome-wide suggestive significance in the SNP-based International Genomics of Alzheimer’s Project (IGAP)8. analyses (best SNP rs429358: OR = 2.26, 95% CI = Note that the genome-wide screening data of the IGAP 1.66–3.07, P = 2.68E-07; Supplementary Fig. 1 and Sup- study (“stage 1”) was also included in the meta-analysis by plementary Table 2) and genome-wide significant asso- Jansen et al. After removal of ambiguous SNPs (A/T and ciation in the gene-based tests (APOC1 [P = 3.24E-07] C/G) and filtering SNPs by MAF > 0.01 and imputation and APOE [P = 3.39E-07]; Supplementary Fig. 1B and quality Rsq >0.8, PLINK 1.9 software13 was used for Supplementary Table 2). Interestingly, and in contrast to linkage disequilibrium (LD) pruning and scoring for a most other previous AD GWAS (e.g., Jansen et al.7 and variety of P-value thresholds (5E-08, 5E-06, 1E-04, 0.01, Lambert et al.8), the best-associated SNP (rs429358) in 0.05, 0.10, 0.20, 0.30, 0.40, 0.50, 1.00). The resulting PRSs this region is the variant defining the “ε4” allele in the were used as independent variable in the regression commonly used “ε/2/3/4” haplotype (the “ε2” allele is models adjusting for sex, age, and PC1 to PC5 as cov- defined by rs7412). APOE was also the top-associated ariates. For the phenotypes not representing the diag- region in the ADNI dataset (Supplementary Table 2), as nostic outcome, we also included diagnosis as additional previously described26. Interestingly, in analyses compar- covariate. For linear models, variance explained (R2) was ing MCI vs. controls, the APOE region did not emerge as derived from comparing results from the full model strongly associated (i.e., P-value for rs429358 = 0.17; (including PRS and covariates) vs. the null model (linear Supplementary Fig. 2 and Supplementary Table 3). model with covariates only). Using a similar partitioning Instead, the best-associated SNP was rs153308 (OR = approach, we estimated the percent trait variance 0.51, 95% CI = 0.39–0.67, P = 1.25E-06; Supplementary explained by our GWAS results for the five CSF traits. For Fig. 2A, Supplementary Table 3), located in an intergenic logistic models, we calculated Nagelkerke’s r2 using the R region on chromosome 5q13.3. However, neither this nor package fmsb. A full description of these and all other any of the other variants showing P-values <1E-05 showed statistical procedures is provided in the “Supplementary any evidence of association in the ADNI dataset, thus they Methods”. may—at least in part—not reflect genuine association signals. Validation analyses in ADNI Whenever possible, we used whole-genome sequencing GWAS on dichotomous “amyloid classification” data from the ADNI cohort to assess replicability of the In the remainder of our analyses we focus on AD- EMIF-AD MBD findings. The ADNI-WGS sample used relevant CSF biomarker data as phenotypic outcomes of here comprises 808 subjects with available whole-genome our GWAS analyses (see Table 1 for an overview of all sequencing data (for more details on the generation of traits analyzed). These analyses included a dichotomous these data, see http://adni.loni.usc.edu/study-design/). For “amyloid classification” (normal/abnormal) variable, the analyses performed here, we only used unrelated representing a combination of CSF Aβ42/40 ratio, local subjects of European origin (n = 751). Variant-based fil- CSF-Aβ42, and the standardized uptake value ratio tering was performed based on minor allele count (MAC (SUVR) on an amyloid PET scan (see methods section on > 3), missingness rate (not more than 5 %) and Hardy- “amyloid classification” in Bos et al.11 for more details). Hong et al. Translational Psychiatry (2020)10:403 Page 6 of 12 Fig. 1 GWAS results using amyloid status in the EMIF-AD MBD dataset. Manhattan plots of A SNP level and B gene-level genome-wide association results with the “amyloid classification” variable across all diagnostic groups (n = 871). All plots include gene assignments made with FUMA. Dotted red lines represent the threshold for genome-wide significance, i.e., α = 5.0E-08 for SNP-based (A) and α = 2.671E-06 for gene-based (B) analyses. Using this amyloid variable (available for n = 871 indivi- gene(s) associated with the “amyloid classification” vari- duals, of which n = 455 were classified as “abnormal” and able (P = 1.13E-18, Fig. 1B and Supplementary Table 4). n = 416 as “normal”) across all diagnostic groups, we In addition, a second locus emerged at genome-wide observed multiple genome-wide significant signals in the significance in the gene-based analyses, i.e., fermitin APOE region on chromosome 19 (Fig. 1). Similar to the family homolog 2 (FERMT2) (P = 1.03E-06) located on GWAS on diagnostic outcome, the strongest association chromosome 14q22.1. While this gene was originally was observed with the APOE “ε4” allele (i.e., SNP reported to represent an AD risk locus8, this finding was rs429358: OR = 5.66, 95% CI = 4.09−7.84, P = 1.95E-25; not replicated in the much larger GWAS by Jansen et al.7. Fig. 1A and Supplementary Table 4), which was also very Furthermore, gene-based tests in ADNI did not reveal strong in equivalent analyses in the ADNI dataset (P = evidence for association between markers in FERMT2 and 4.15E-22). Despite the consistency of the APOE findings, the “amyloid classification” phenotype (P = 0.23), sug- none of the other suggestive signals outside the APOE gesting that this gene is at best marginally involved in region replicated in ADNI (Supplementary Table 4). As determining variance of this trait. Additional analyses expected, gene-based tests using MAGMA highlighted using the “amyloid classification” variable limited to MCI APOE (and neighboring loci) as the most significant and control individuals revealed a similar picture as in the Hong et al. Translational Psychiatry (2020)10:403 Page 7 of 12 Fig. 2 GWAS results using CSF-Aβ38 in the EMIF-AD MBD dataset. Manhattan plots of A gene-level genome-wide association results with CSF- Aβ38 levels across all diagnostic groups (n = 675), B regional association results zoomed into a 450 kb region surrounding the ZFHX3 gene on chromosome 16q22. All plots include gene assignments and linkage disequilibrium estimates made with FUMA. Dotted red line represents the threshold for genome-wide significance (α = 2.671E-06) for the gene-based analyses. full dataset (see Supplementary Tables 5 and 6, respec- phenotypes available that were used as outcome variables tively), i.e., the strongest association signals were observed in the GWAS. With the exception of “CSF-Aβ38” and for APOE (all replicated in equivalent analyses in ADNI). “CSF-Aβ40” all showed strong and highly significant In addition, we identified a few suggestive non-APOE association with markers in the APOE region but no signals on other chromosomes, albeit none of these better than suggestive (P < 1E-05) signals in the remaining showed evidence for independent replication in ADNI. genome (see Supplementary Figs. 3–7 and Supplementary Tables 7–11). None of the non-APOE, suggestive signals GWAS on additional dichotomous and continuous CSF replicated in the ADNI dataset for phenotypes with amyloid variables available data. GWAS results on “CSF-Aβ38” (Fig. 2, Overall, there were a total of two binary (“Central_CS- Supplementary Fig. 8 and Supplementary Table 12) and F_ratiodich” and “Local_AB42_Abnormal”) and five “CSF-Aβ40” (Supplementary Fig. 9 and Supplementary quantitative (“AB_Zscore”, “Central_CSF_AB38”, “Cen- Table 13) showed highly similar GWAS results owing to tral_CSF_AB40”, “log_Central_CSF_AB42” and “log_- the—well-established27—high correlation between both Central_CSF_AB4240ratio”; see Table 1) CSF-Aβ markers, which we also observed in this dataset (Pearson’s Hong et al. Translational Psychiatry (2020)10:403 Page 8 of 12 r2 = 0.96, P-value < 2.2E-16 across all samples with avail- also observe in our data (Pearson r2 = 0.87, P-value < able data). The most noteworthy finding emerging from 2.2E-16). Owing to this phenotypic correlation, the these analyses was genome-wide significant association GWAS results for both variables also look quite similar, as between CSF-Aβ40 and markers in the gene encoding expected (Supplementary Fig. 13B and Supplementary zinc finger homeobox 3 (ZFHX3) in gene-based analyses Table 18), and replicate the sex-specific difference. (gene-based P-value = 7.48E-08; best SNP P-value = Genome-wide significant (P < 5E-08) and suggestive (P < 7.29E-08; Supplementary Fig. 9 and Supplementary Table 5E-06) SNPs explained about 28–34% of the phenotypic 13; similar results were obtained with CSF-Aβ38, Fig. 2). variance in our dataset (Supplementary Table 14). How- Despite the highly significant and consistent association ever, we note that this number likely overestimates the between CSF-Aβ40 and CSF-Aβ38 levels and ZFHX3, true variance explained as GWAS and genome parti- this finding was not replicated in ADNI (gene-based tioning were performed in the same dataset. P-value = 0.77; Supplementary Table 13). Genome-wide significant (P < 5E-08) and suggestive (P < 5E-06) SNPs Polygenic risk score (PRS) analyses on all outcome traits explained 29-32% of the phenotypic variance in our In addition to testing all above-mentioned traits for dataset (Supplementary Table 14). However, we note that SNPs and genes in the context of genome-wide analyses, this number likely overestimates the true variance we also computed association statistics with aggregated explained as GWAS and genome partitioning were per- variant data in the form of PRS using AD case–control formed in the same dataset. results from Lambert et al.8 and Jansen et al.7. The aim was to assess the degree at which established AD- GWAS on dichotomous and continuous CSF tau variables associated markers also show association with the “AD- Similar to the GWAS analyses for CSF Aβ measures, related” (endo-) phenotypes analyzed here. Although a there were several dichotomous and continuous measures considerable amount of sample overlap exists across the of CSF tau available in the EMIF-AD MBD dataset (see two AD risk GWAS (i.e., both use the “stage I” GWAS Table 1, Supplementary Figs. 10–13, Supplementary data from the IGAP sample in their discovery phase), both Tables 15–18). Using CSF-Ttau levels as z-scored con- studies use different analysis paradigms and different tinuous outcome we identified genome-wide significant replication cohorts. Given that the final effective sample association with markers in geminin coiled-coil domain size used in Jansen et al. is nearly ten-times larger than containing (GMNC) on chromosome 3q28 in gene-based that in Lambert, we hypothesized that the results and, analyses (P = 1.61E-06; Supplementary Fig. 11B, Supple- accordingly PRS, derived from the larger study are more mentary Table 16). In the SNP-based analyses, markers in “precise” and will, therefore, show stronger association this gene showed genome-wide suggestive P-values ran- and explain more of the trait variance analyzed here. PRS- ging between 4.3E-06 and 9.5E-06 (Supplementary Table based results are summarized in Table 2, while full results 16). These results showed marginal evidence for associa- can be found in Supplementary Table 19 (for PRS from tion in the ADNI cohort on the SNP level (for rs6444469 Jansen et al.7) and 20 (for PRS from IGAP8). Overall, we with P-value = 0.09), but not on the gene-level (P-value = observed significant PRS-based associations with many, 0.59; Supplementary Table 16). Interestingly, association but not all, traits analyzed in the EMIF-AD MBD dataset. between markers in GMNC and CSF-Ttau levels were For both PRS models, the best-associated trait was a previously described28. A follow-up study to the original diagnosis of AD and all measures involving CSF-Aβ42 report provided evidence for sex-specific differences at levels. In contrast, no noteworthy associations were this locus (rs1393060 proximal to GMNC and in strong observed with CSF-Aβ38 and CSF-Aβ40 levels and, per- LD with rs6444469 [r2 = 1]: P-value = 4.57E-10 in females haps more interesting, not with either of the two available compared to P-value = 0.03 in males), suggesting stronger CSF-tau measures (CSF-Ttau and CSF-Ptau). Comparing effects in females29. In EMIF-AD MBD, we also observed the “performance” of both PRS against each other a stronger association in females for GMNC at the gene revealed that—against our expectation—the IGAP-based and SNP level, respectively (gene: P-value = 4.09E-06 [in results tended to show the stronger statistical support (i.e., females] vs. P-value = 0.08 [in males]; SNP [for smaller P-values) and explained slightly more of the rs6444469]: P-value = 1.47E-04 [in females] vs. P-value = phenotypic variance (i.e., showed higher r2) than the PRS 8.91E-03 [in males]), hence, providing independent derived from more recent and larger GWAS by Jansen replication of the previous report. Similarly, the associa- et al. (Supplementary Table 19 vs. 20). The only exception tion results for rs6444469 and CSF-Ttau in ADNI were being the case–control analyses on AD status, where the more pronounced in females (P-value = 0.013) than males Jansen PRS outperformed that from IGAP (i.e., r2 = (P-value = 0.93). The other available CSF tau measure in 4.35%, P-value = 4.03E-06 vs. r2 = 2.98%, P-value = our GWAS was CSF-Ptau, which is known to strongly 1.07E-04, respectively). Interestingly, the association of correlate with CSF-Ttau levels30, a correlation, which we AD-based PRS with risk for MCI was minor in both Table 2 Summary of polygenic risk score (PRS) analyses using two P-value thresholds and two different GWAS datasets with and without markers in the APOE region. PRS constructed based on GWAS by Jansen et al.7 PRS constructed based on GWAS by IGAP, 2013 Including APOE Excluding APOE Including APOE Excluding APOE S1(P < 5E-8) S5( < 0.05) S1(P < 5E-8) S5( < 0.05) S1(P < 5E-8) S5( < 0.05) S1(P < 5E-8) S5( < 0.05) Phenotype r2 P-value r2 P-value r2 P-value r2 P-value r2 P-value r2 P-value r2 P-value r2 P-value AD 3.09% 8E-05 2.89% 0.0002 1.22% 0.0126 0.99% 0.0244 2.98% 0.0001 1.69% 0.0034 0.78% 0.0443 0.10% 0.4745 MCI 0.12% 0.3767 0.77% 0.0249 0.07% 0.4894 0.61% 0.0449 0.51% 0.0682 0.13% 0.3537 0.49% 0.0743 0.00% 0.8764 AMYLOIDstatus 1.38% 0.0004 1.01% 0.0025 0.38% 0.0643 0.04% 0.5599 3.82% 9E-09 2.95% 3E-07 1.09% 0.0018 0.34% 0.0766 Amyloid.MCI 4.59% 0.0008 5.82% 0.0002 1.33% 0.0684 0.87% 0.1394 12.54% 1E-07 7.45% 2E-05 4.63% 0.0008 0.82% 0.1513 Hong et al. Translational Psychiatry (2020)10:403 Amyloid.NC 0.83% 0.1637 0.23% 0.4657 0.20% 0.4928 0.01% 0.8786 2.45% 0.0174 5.10% 0.0008 0.24% 0.4592 1.28% 0.0866 AB_Zscore 1.20% 0.0003 0.80% 0.0031 0.27% 0.0871 0.01% 0.7683 2.61% 7E-08 3.08% 4E-09 0.39% 0.0382 0.38% 0.04 Central_CSF_ratiodich 3.86% 2E-06 1.94% 0.0007 1.49% 0.0029 0.10% 0.4414 5.87% 9E-09 3.52% 6E-06 1.58% 0.0022 0.22% 0.2439 Local_AB42_Abnormal 3.68% 4E-06 1.13% 0.0098 1.65% 0.0019 0.02% 0.7287 3.40% 1E-05 1.66% 0.0018 0.94% 0.0184 0.01% 0.7873 Central_CSF_AB38 0.21% 0.2273 0.00% 0.8587 0.32% 0.1305 0.01% 0.7838 0.00% 0.8816 0.28% 0.1614 0.03% 0.6416 0.33% 0.1279 Central_CSF_AB40 0.24% 0.1805 0.01% 0.7974 0.28% 0.1514 0.05% 0.5298 0.05% 0.5418 0.44% 0.0718 0.07% 0.4874 0.36% 0.106 log_Central_CSF_AB4240ratio 2.06% 4E-05 1.68% 0.0002 0.49% 0.0453 0.07% 0.4428 3.74% 3E-08 2.55% 5E-06 0.56% 0.0323 0.05% 0.5324 log_Central_CSF_AB42 1.52% 0.0005 0.52% 0.0445 0.56% 0.0373 0.01% 0.8297 2.31% 2E-05 2.17% 4E-05 0.53% 0.0425 0.25% 0.1605 Local_PTAU_Abnormal 0.31% 0.1602 0.03% 0.662 0.06% 0.5386 0.01% 0.8148 0.69% 0.0347 0.38% 0.1192 0.19% 0.2697 0.09% 0.4424 Local_TTAU_Abnormal 0.93% 0.0107 0.01% 0.7505 0.56% 0.0465 0.12% 0.3625 0.82% 0.0166 0.27% 0.1644 0.17% 0.267 0.04% 0.5771 Ptau_ASSAY_Zscore 0.34% 0.0909 0.00% 0.9833 0.09% 0.3922 0.05% 0.5191 0.37% 0.0806 0.08% 0.4258 0.01% 0.7755 0.01% 0.8237 Ttau_ASSAY_Zscore 0.51% 0.032 0.03% 0.6082 0.23% 0.1529 0.00% 0.9666 0.09% 0.3566 0.11% 0.3269 0.03% 0.5811 0.00% 0.8676 Italicized values = nominally significant association. r2 = variance explained; bold font = largest r2 (= most variance explained) for trait in question. A full listing of results from these PRS analyses can be found in Supplementary righthandcolumns of Table 2 and bottom part of SupplementaryTable 19 (for Jansen et al. GWAS) and Supplementary Table 20 (for IGAP GWAS). Page 9 of 12 Hong et al. Translational Psychiatry (2020)10:403 Page 10 of 12 models (r2 = 0.83%, P-value = 0.02, and r2 = 0.65%, P- paving the way for future genomic discoveries with value = 0.039, for Jansen and IGAP, respectively). additional phenotypes available in this unique cohort11. To investigate the contribution of markers in the APOE Despite these promising first results, our study is region, we repeated all analyses excluding variants within 1 potentially confined by a number of possible limitations: MB of APOE (chr19:45409039-45412650; right-hand col- First and foremost, despite the breadth of available phe- umns of Table 2 and affected were the analyses on AD notype data, the overall sample size of the EMIF-AD MBD (strongest reduction in bottom part of Supplementary dataset is comparatively small and, as a result, may not be Tables 19 and 20). As expected, removal of APOE region adequately powered to detect genetic variants exerting markers from the PRS decreased the variance explained for smaller effects. To a degree, this limitation is alleviated by most of the traits analyzed here, albeit to varying degrees. the fact that many available outcome phenotypes are of a Most affected were the analyses on AD (strongest reduction quantitative nature, which are more powerful than ana- in r2 = 73.8% in analyses excluding APOE effects vs. the full lyses of dichotomous traits (e.g., disease risk). Second, model including APOE in IGAP) and essentially all CSF- many of the phenotypes available in EMIF-AD MBD are Aβ42 related measures (strongest reduction in r2 = 88.9% not currently ascertained in other, independent datasets for trait “AB_Zscore” in IGAP) for both PRS models. Less (such as ADNI), making independent replication of any affected by the removal of APOE were the analyses of CSF- novel findings difficult. This situation can be expected to tau species (strongest reduction in r2 = 29.0% for CSF-Ttau improve somewhat once the phenotypic breadth in ADNI in non-APOE vs. APOE models). and other cohorts is extended. Still, until independent replication is available, novel GWAS findings from EMIF- Discussion AD MBD must be interpreted with caution. This includes This is the first GWAS utilizing part of the wide col- the putative association between CSF-Aβ38 and CSF- lection of AD-relevant phenotypes and biomarkers avail- Aβ40 and markers in ZFHX3, which were highly sig- able in the EMIF-AD MBD dataset. The phenotypes nificant and consistent in EMIF-AD MBD but not repli- analyzed here related either to clinical diagnosis (i.e., AD cated in ADNI. Until more independent data on these or MCI) or to levels of CSF biomarkers revolving around outcome traits are available, the ZFHX3 association various biochemical species of amyloid or tau proteins. should be considered “provisional”. In this context it is While GWAS results have already been reported for some comforting, however, that many well-established genetics of the biomarkers analyzed here (e.g., in ADNI), ours are findings (such as the association between APOE and the first to combine genomic and biomarker data in the measures of CSF-Aβ or GMNC and CSF-tau) were newly established EMIF-AD MBD dataset. The main reproduced in EMIF-AD MBD. Third, while the genome- findings of our study can be summarized in the following wide SNP genotype data was generated in one run of five points: (1) the most prominent genetic signals in consecutive experiments in one laboratory, the same is analyses of either diagnostic outcome or phenotypes not true for the phenotype measurements, which were related to CSF-Aβ42 were observed with markers in or performed locally in each of the 11 participating sites. In near APOE, which is in good agreement with equivalent some instances, the compiled phenotype data are not analyses in ADNI and other datasets; (2) our analyses based on the same biochemical assays across sites for identified one novel association in analyses of CSF-Aβ38 some variables, e.g., measurements of tau protein. While and CSF-Aβ40 levels and DNA sequence variants in the EMIF-AD MBD phenotype team went to great lengths ZFHX3 (a.k.a. ATBF1 [AT-motif binding factor 1]), to alleviate this potential problem by normalizing vari- although these signals were not replicated in the ADNI ables for each center (see Bos et al.11 for more details), the dataset; (3) using CSF-tau species (i.e., Ttau and Ptau), we possibility of artefactual findings owing to phenotypic confirmed the previously described association with SNPs heterogeneity remains. Finally, as described in the overall in GMNC, including the recently reported effect mod- cohort description manuscript, the EMIF-AD MBD ification by sex at this locus; (4) PRS analyses revealed that dataset is not designed to be “representative” of the gen- AD risk SNPs are mostly associated with phenotypes eral population but was assembled with the aim to achieve related to CSF-Aβ42 but not CSF-Aβ38, CSF-Aβ40, and approximately equal proportions of amyloid+ vs. amy- most notably CSF-tau; (5) exclusion of APOE from the loid- individuals in all three diagnostic subgroups. While PRS analyses suggest that non-APOE AD GWAS SNPs this ascertainment strategy does not invalidate our GWAS explain at most 2.5% of the phenotypic variance under- results per se, they may not be generalizable to the lying a diagnosis of AD in this dataset. Collectively, these population as a whole. However, this limitation may affect results implicate that the genetic architecture underlying any study with clinically ascertained participants and, many traits relevant for AD research in EMIF-AD MBD thus, applies to most previously published GWAS in the compare well to other datasets of European descent field, including those performed in ADNI. Hong et al. Translational Psychiatry (2020)10:403 Page 11 of 12 In conclusion, our first-wave of GWAS analyses in the Gothenburg, Gothenburg, Sweden. 7Clinical Neurochemistry Laboratory, EMIF-AD MBD dataset provides a first important step in Sahlgrenska University Hospital, Mölndal, Sweden. 8Laboratory for Cognitive Neurology, Department of Neurosciences, KU Leuven, Leuven, Belgium. a series of additional genome-wide and epigenome-wide 9 Neurology Service, University Hospital Leuven, Leuven, Belgium. 10Laboratory (using DNA methylation profiles) association analyses in for Complex Genetics, Department of Human Genetics, KU Leuven, Leuven, this valuable and unique cohort. Belgium. 11Neurochemistry Laboratory, Department of Clinical Chemistry, Amsterdam Neuroscience, Amsterdam University Medical Centers, Vrije Universiteit, Amsterdam, The Netherlands. 12Department of Biomedical Acknowledgements Sciences, Institute Born-Bunge, University of Antwerp, Antwerp, Belgium. The present study was conducted as part of the EMIF-AD MBD project, which 13 Department of Neurology and Center for Neurosciences, UZ Brussel and Vrije has received support from the Innovative Medicines Initiative Joint Undertaking Universiteit Brussel (VUB), Brussels, Belgium. 14University of Geneva, Geneva, under EMIF grant agreement n° 115372, the resources of which are composed Switzerland. 15IRCCS Istituto Centro San Giovanni di Dio Fatebenefratelli, of financial contribution from the European Union’s Seventh Framework Brescia, Italy. 16AIX Marseille University, INS, Ap-hm, Marseille, France. Program (FP7/2007–2013) and EFPIA companies’ in kind contribution. Parts of 17 Neurosciences Therapeutic Area, GlaxoSmithKline R&D, Stevenage, UK. this study were made possible through support from the German Research 18 University of Lille, Inserm, CHU Lille, Lille, France. 19Alzheimer’s disease and Foundation (DFG grant FOR2488: Main support by subproject “INF-GDAC” other cognitive disorders unit, Hospital Clinic I Universitari, Barcelona, Spain. BE2287/7-1 to LB). R.V. acknowledges the support by the Stichting Alzheimer 20 Department of Neuropathology, Nuffield Department of Clinical Onderzoek (#13007, #11020, #2017-032) and the Flemish Government (VIND IWT Neurosciences, University of Oxford, Oxford OX3 9DU, UK. 21Memory Unit, 135043). H.Z. is a Wallenberg Academy Fellow supported by grants from the Neurology Department, Hospital de Sant Pau, Barcelona and Centro de Swedish Research Council (#2018-02532), the European Research Council Investigación Biomédica en Red en enfermedades Neurodegenerativas (#681712) and Swedish State Support for Clinical Research (#ALFGBG-720931). S. (CIBERNED), Madrid, Spain. 22Geriatric Psychiatry, Department of Mental Health J.B.V. received funding from the Innovative Medicines Initiative 2 Joint and Psychiatry, Geneva University Hospitals, Geneva, Switzerland. Undertaking under ROADMAP grant agreement No. 116020 and from ZonMw 23 Department of Psychiatry, University Hospital of Lausanne, Lausanne, during the conduct of this study. No conflict of interest exists. Research at VIB- Switzerland. 24Department of Neurology, Center for Research and Advanced UAntwerp was in part supported by the University of Antwerp Research Fund. Therapies, CITA-Alzheimer Foundation, San Sebastian, Spain. 25Department of The research was supported by ALF clinical grants from Region Västra Götaland Biostatistics and Health Informatics, Institute of Psychiatry, Psychology and to A.W. and to P.K. We acknowledge the assistance of Ellen De Roeck, Naomi De Neuroscience, King’s College London, London, UK. 26NIHR BioResource Centre Roeck, and Hanne Struyfs (UAntwerp) with data collection. The Lausanne study Maudsley, NIHR Maudsley Biomedical Research Centre (BRC) at South London was funded by a grant from the Swiss National Research Foundation (SNF and Maudsley NHS Foundation Trust (SLaM) & Institute of Psychiatry, 320030_141179) to J.P. We thank Mrs. Tanja Wesse and Mrs. Sanaz Sedghpour Psychology and Neuroscience (IoPPN), King’s College London, London, UK. Sabet at the Institute of Clinical Molecular Biology, Christian-Albrechts-University 27 Health Data Research UK London, University College London, 222 Euston of Kiel, Kiel, Germany for technical assistance with the GSA genotyping. The LIGA Road, London, UK. 28Institute of Health Informatics, University College London, team acknowledges computational support from the OMICS compute cluster at 222 Euston Road, London, UK. 29The National Institute for Health Research the University of Lübeck. The computations in the ADNI cohort were run on the University College London Hospitals Biomedical Research Centre, University Odyssey cluster supported by the FAS Division of Science, Research Computing College London, 222 Euston Road, London, UK. 30Steno Diabetes Center, Group at Harvard University. For ADNI: Data was used for this project of which Copenhagen, Denmark. 31Institute of Pharmaceutical Sciences, King’s College collection and sharing was funded by the Alzheimer’s Disease Neuroimaging London, London, UK. 32Neurodegenerative Brain Diseases Group, Center for Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD Molecular Neurology, VIB, Antwerp, Belgium. 33Department of Biomedical ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is Sciences, University of Antwerp, Antwerp, Belgium. 34Alzheimer Center and funded by the National Institute on Aging, the National Institute of Biomedical Department of Neurology, Amsterdam University Medical Centers, Amsterdam Imaging and Bioengineering, and through generous contributions from the Neuroscience, Amsterdam, The Netherlands. 35Department of Radiology and following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Nuclear Medicine, Amsterdam University Medical Centers, Amsterdam Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Neuroscience, Amsterdam, The Netherlands. 36Institutes of Neurology and Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly Healthcare Engineering, University College London, London, UK. 37Department and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company of Neurodegenerative Disease, UCL Queen Square Institute of Neurology, Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Queen Square, London, UK. 38UK Dementia Research Institute at UCL, London, Immunotherapy Research & Development, LLC.; Johnson & Johnson UK. 39Department of Psychiatry, University of Oxford, Oxford, UK. 40Reference Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Center for Biological Markers of Dementia (BIODEM), Institute Born-Bunge, Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; University of Antwerp, Antwerp, Belgium. 41Translational Medicine Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Neuroscience, UCB Biopharma SPRL, Braine l’Alleud, Belgium. 42Institute of Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Clinical Molecular Biology, Christian-Albrechts-University of Kiel, Kiel, Germany. Institutes of Health Research is providing funds to support ADNI clinical sites in 43 Department of Psychology, University of Oslo, Oslo, Norway Canada. Private sector contributions are facilitated by the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the Northern California Institute for Research and Education, and the study is Author contributions coordinated by the Alzheimer’s Therapeutic Research Institute at the University S.H. performed all the analyses and interpretation on data and wrote the of Southern California. ADNI data are disseminated by the Laboratory for manuscript. D.P. and R.T. contributed to replication in ADNI. V.D. was NeuroImaging at the University of Southern California. responsible for EMIF-AD MBD DNA sample preparation and extraction. F.K. was responsible for EMIF-AD MBD genotypic data QC and imputation. I.B., S.V., B.T., Author details and P.J. coordinated the collection and harmonization of phenotypes and 1 Lübeck Interdisciplinary Platform for Genome Analytics (LIGA), Institutes of biosamples in EMIF-AD MBD and helped identifying equivalent phenotypes Neurogenetics and Cardiogenetics, University of Lübeck, Lübeck, Germany. from the ADNI catalog. A.F. and M.W. supervised the genotyping experiments. 2 Genetics and Aging Unit and McCance Center for Brain Health, Department of U.A., K.B., and H.Z. performed CSF biomarker measurements and took part in Neurology, Massachusetts General Hospital, Boston, MA, USA. 3Department of cut-point determinations. K.S. and C.V.B. contributed to genetic Psychiatry and Neuropsychology, School for Mental Health and Neuroscience, characterization of samples and design of the genomics studies in EMIF-AD Alzheimer Centrum Limburg, Maastricht University, Maastricht, The MBD, and critically revised the manuscript drafts. R.V., S.G., J.S., and I.C. Netherlands. 4Alzheimer Center Amsterdam, Department of Neurology, contributed to sample and data collection. E.N. and I.S. were responsible for Amsterdam Neuroscience, Vrije Universiteit Amsterdam, Amsterdam UMC, The data collection of the Antwerp cohort. A.W. and P.K. contributed to data and Netherlands. 5Alzheimer Center Amsterdam, Department of Neurology, sample collection/handling of the Gothenburg MCI Study, Gothenburg, Amsterdam Neuroscience, Vrije Universiteit Amsterdam, Amsterdam UMC, The Sweden. J.S., P.J.V., and S.L. are leads for the EMIF-AD MBD; designed and Netherlands. 6Institute of Neuroscience and Physiology, Department of managed the platform. L.B. designed and supervised the genomics portion of Psychiatry and Neurochemistry, The Sahlgrenska Academy, University of the EMIF-AD MBD project and co-wrote all drafts of the manuscript. All authors Hong et al. Translational Psychiatry (2020)10:403 Page 12 of 12 critically revised all manuscripts drafts, read, and approved the final 9. Bertram, L. et al. Genome-wide association analysis reveals putative Alzhei- manuscript. mer’s disease susceptibility loci in addition to APOE. Am. J. Hum. Genet. 83, 623–632 (2008). Data availability 10. Stocker, H., Möllers, T., Perna, L. & Brenner, H. The genetic risk of Alzheimer’s GWAS summary statistics for the top (P-value < 1E-05) results are listed in the disease beyond APOE ε4: systematic review of Alzheimer’s genetic risk scores. Supplementary Tables. Full GWAS summary statistics are available from the Transl. Psychiatry 8, 166 (2018). authors upon request. Clinical data and genome-wide genotyping data are 11. Bos, I. et al. The EMIF-AD multimodal biomarker discovery study: design, stored on an online data platform using the “tranSMART” data warehouse methods and cohort characteristics. Alzheimers Res. Ther. 10, 64 (2018). framework. Access to the genome-wide genotyping data can be requested 12. Mueller, S. G. et al. The Alzheimer’s disease neuroimaging initiative. Neuroi- from the corresponding author of this study who will forward each request to maging Clin. N. Am. 15, 869–877 (2005). the EMIF-AD data access team. 13. Purcell, S. et al. PLINK: a tool set for whole-genome association and population-based linkage analyses. Am. J. Hum. Genet. 81, 559–575 (2007). Code availability 14. Danecek, P., Schiffels, S. & Durbin, R. Multiallelic Calling Model In bcftools (-m). All scripts used to generate the primary GWAS and PRS analyses are available https://samtools.github.io/bcftools/. from the authors upon request. 15. Delaneau, O., Marchini, J. & Zagury, J.-F. A linear complexity phasing method for thousands of genomes. Nat. Methods 9, 179–181 (2012). 16. Das, S. et al. Next-generation genotype imputation service and methods. Nat. Conflict of interest Genet. 48, 1284–1287 (2016). The authors declare that they have no conflict of interest. 17. Bos, I. et al. Cerebrospinal fluid biomarkers of neurodegeneration, synaptic integrity, and astroglial activation across the clinical Alzheimer’s disease spectrum. Alzheimer’s Dement. 15, 644–654 (2019). Publisher’s note 18. Li, Y., Willer, C., Sanna, S. & Abecasis, G. Genotype imputation. Annu. Rev. Springer Nature remains neutral with regard to jurisdictional claims in Genomics Hum. Genet. 10, 387–406 (2009). published maps and institutional affiliations. 19. Li, Y., Willer, C. J., Ding, J., Scheet, P. & Abecasis, G. R. MaCH: using sequence and genotype data to estimate haplotypes and unobserved genotypes. Genet. Supplementary Information accompanies this paper at (https://doi.org/ Epidemiol. 34, 816–834 (2010). 10.1038/s41398-020-01074-z). 20. Turner, S. D. qqman: an R package for visualizing GWAS results using Q-Q and manhattan plots. J. Open Source Soft., 3, 731, https://doi.org/10.21105/ Received: 6 March 2020 Revised: 23 April 2020 Accepted: 18 May 2020 joss.00731 (2018). 21. Aulchenko, Y. S., Ripke, S., Isaacs, A. & van Duijn, C. M. GenABEL: an R library for genome-wide association analysis. Bioinformatics 23, 1294–1296 (2007). 22. T. I. H. International HapMap Consortium. A haplotype map of the human genome. Nature 437, 1299–1320 (2005). References 23. Pe’er, I., Yelensky, R., Altshuler, D. & Daly, M. J. Estimation of the multiple testing 1. Sperling, R., Mormino, E. & Johnson, K. The evolution of preclinical Alzheimer’s burden for genomewide association studies of nearly all common variants. disease: implications for prevention trials. Neuron 84, 608–622 (2014). Genet. Epidemiol. 32, 381–385 (2008). 2. Mattsson, N. et al. Revolutionizing Alzheimer’s disease and clinical trials 24. Watanabe, K., Taskesen, E., van Bochoven, A. & Posthuma, D. Functional through biomarkers. Alzheimer’s Dement. Diagnosis, Assess. Dis. Monit. 1, mapping and annotation of genetic associations with FUMA. Nat. Commun. 8, 412–419 (2015). 1826 (2017). 3. Tanzi, R. E. & Bertram, L. Twenty years of the alzheimer’s disease amyloid 25. de Leeuw, C. A., Mooij, J. M., Heskes, T. & Posthuma, D. MAGMA: generalized hypothesis: a genetic perspective. Cell 120, 545–555 (2005). gene-set analysis of GWAS data. PLOS Comput. Biol. 11, e1004219 (2015). 4. Gatz, M. et al. Role of genes and environments for explaining Alzheimer 26. Apostolova, L. G. et al. Associations of the top 20 Alzheimer disease risk disease. Arch. Gen. Psychiatry 63, 168 (2006). variants with brain amyloidosis supplemental content. JAMA Neurol. 75, 5. Bertram, L., McQueen, M. B., Mullin, K., Blacker, D. & Tanzi, R. E. Systematic meta- 328–341 (2018). analyses of Alzheimer disease genetic association studies: the AlzGene data- 27. Schoonenboom, N. S. et al. Amyloid β 38, 40, and 42 species in cerebrospinal base. Nat. Genet. 39, 17–23 (2007). fluid: More of the same?. Ann. Neurol. 58, 139–142 (2005). 6. Kim, J., Basak, J. M. & Holtzman, D. M. The role of apolipoprotein E in Alz- 28. Cruchaga, C. et al. GWAS of cerebrospinal fluid tau levels identifies risk variants heimer’s disease. Neuron 63, 287–303 (2009). for Alzheimer’s disease. Neuron 78, 256–268 (2013). 7. Jansen, I. E. et al. Genome-wide meta-analysis identifies new loci and func- 29. Deming, Y. et al. Sex-specific genetic predictors of Alzheimer’s disease bio- tional pathways influencing Alzheimer’s disease risk. Nat. Genet. 51, 404–413 markers. Acta Neuropathol. 136, 857–872 (2018). (2019). 30. Mulder, C. et al. Amyloid-beta(1-42), total tau, and phosphorylated tau as 8. Lambert, J.-C. et al. Meta-analysis of 74,046 individuals identifies 11 new sus- cerebrospinal fluid biomarkers for the diagnosis of Alzheimer disease. Clin. ceptibility loci for Alzheimer’s disease. Nat. Genet. 45, 1452–1458 (2013). Chem. 56, 248–253 (2010).