Rare coding variants in PLCG2, ABI3 and TREM2 implicate microglial-mediated innate immunity in Alzheimer’s disease

Introduction

We identified rare coding variants associated with Alzheimer’s disease (AD) in a 3-stage case-control study of 85,133 subjects. In stage 1, 34,174 samples were genotyped using a whole-exome microarray. In stage 2, we tested associated variants (P<1×10^-4) in 35,962 independent samples using de novo genotyping and imputed genotypes. In stage 3, an additional 14,997 samples were used to test the most significant stage 2 associations (P<5×10^-8) using imputed genotypes. We observed 3 novel genome-wide significant (GWS) AD associated non-synonymous variants; a protective variant in PLCG2 (rs72824905/p.P522R, P=5.38×10^-10, OR=0.68, MAF_cases=0.0059, MAF_controls=0.0093), a risk variant in ABI3 (rs616338/p.S209F, P=4.56×10^-10, OR=1.43, MAF_cases=0.011, MAF_controls=0.008), and a novel GWS variant in TREM2 (rs143332484/p.R62H, P=1.55×10^-14, OR=1.67, MAF_cases=0.0143, MAF_controls=0.0089), a known AD susceptibility gene. These protein-coding changes are in genes highly expressed in microglia and highlight an immune-related protein-protein interaction network enriched for previously identified AD risk genes. These genetic findings provide additional evidence that the microglia-mediated innate immune response contributes directly to AD development.

Late-onset AD (LOAD) has a significant genetic component (h²=58-79%¹). Nearly 30 LOAD susceptibility loci^2-12 are known, and risk is significantly polygenic¹³. However, these loci explain only a proportion of disease heritability. Rare variants also contribute to disease risk^14-17. Recent sequencing studies identified a number of genes that have rare variants associated with AD^9-11,18-24. Our approach to rare-variant discovery is to genotype a large sample with micro-arrays targeting known exome variants with follow-up using genotyping and imputed genotypes in a large independent sample. This is a cost-effective alternative to de novo sequencing^25-29.

We applied a 3-stage design (Supplementary Figure 1) using subjects from the International Genomics of Alzheimer’s Project (IGAP)(Table 1, Supplementary Tables 1 & 2). In stage 1, 16,097 LOAD cases and 18,077 cognitively normal elderly controls were genotyped using the Illumina HumanExome microarray. Data from multiple consortia were combined in a single variant meta-analysis (Online Methods) assuming an additive model. In total, 241,551 variants passed quality-control (Supplementary Table 3). Of these 203,902 were polymorphic, 26,947 were common (minor allele frequency (MAF)≥5%), and 176,955 were low frequency or rare (MAF<5%). We analyzed common variants using a logistic regression model in each sample cohort and combined data using METAL³⁰. Rare and low frequency variants were analyzed using the score test and data combined with SeqMeta³¹ (Supplementary Figure 2).

Table 1.

Summary of the consortium data sets used for stages 1, 2 and stage 3. Data are from the Genetic and Environmental Risk for Alzheimer’s Disease (GERAD)/Defining Genetic, Polygenic and Environmental Risk for Alzheimer’s Disease (PERADES) Consortium, the Alzheimer’s Disease Genetic Consortium (ADGC), the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) and the European Alzheimer’s disease Initiative (EADI)(Supplement 1).

	Consortium	N Controls	N Cases	N Total
Stage 1	GERAD/PERADES	2974	6000	8974
	ADGC	7002	8706	15708
	CHARGE	8101	1391	9492
Total		18077	16097	34174
Stage 2	GERAD/PERADES genotype	5049	4049	9098
	CHARGE-genotype	1839	1434	3273
	CHARGE-in silico	3246	722	3968
	EADI-genotype	11787	7836	19623
Total		21921	14041	35962
Stage 3	ADGC-in silico	8345	6652	14997
Stage 1 + 2 + 3
Total		48402	37022	85133

We reviewed cluster plots for variants showing association (P<1×10^-4) and identified 43 candidate variants (Supplementary Table 4) exclusive of known risk loci (Supplementary Table 5). Stage 2 tested these for association in 14,041 LOAD cases and 21,921 controls, using de novo and imputation derived genotypes (Online Methods). We carried forward single nucleotide variants (SNVs) with GWS associations and consistent directions of effect to stage 3 where genotypes for 6,652 independent cases and 8,345 controls were imputed using the Haplotype Reference Consortium resource^32,33 (Online Methods, Supplementary Table 6).

We identified four rare coding variants with GWS association signals with LOAD (P<5×10^-8)(Table 2, Supplementary Tables 7 & 8). The first is a missense variant p.P522R (P=5.38×10^-10, OR=0.68) in Phospholipase C Gamma 2 (PLCG2)(Table 2, Figure 1a, Supplementary Table 9, Supplementary Figure 3). This variant is associated with decreased risk of LOAD, showing a MAF of 0.0059 in cases and 0.0093 in controls. The reference allele (p.P522) is conserved across several species (Supplementary Figure 4). Gene-wide analysis showed nominal evidence for association at P=1.52×10^-4 (Supplementary Tables 10 & 11) and we found no other independent association at this gene (Supplementary Figure 5).

Table 2.

Summary of stage 1, 2, 3 and combined meta-analysis results for SNVs at P<5×10^-8. Data includes p-values, odds ratios (OR), minor allele frequency (MAF) in cases and controls and number of subjects included in each analytical stage. For OR 95% confidence intervals see Supplementary Table 7.

SNV	rs75932628	rs143332484	rs72824905	rs616338
Chr	6	6	16	17
Position	41129252	41129207	81942028	47297297
Protein Variation	R47H	R62H	P522R	S209F
Gene	TREM2	TREM2	PLCG2	ABI3
Effect Allele	T	T	G	T
Stage 1
P	3.02E-12	3.48E-09	1.19E-05	2.16E-05
OR	2.46	1.58	0.65	1.42
MAF Cases	0.003	0.015	0.006	0.013
MAF Controls	0.001	0.010	0.011	0.010
N	30018	33786	33786	33786
Stage 2
P	4.38E-08	3.66E-07	1.35E-04	8.37E-05
OR	2.37	3.97	0.70	1.41
MAF Cases	0.004	0.014	0.006	0.010
MAF Controls	0.002	0.006	0.008	0.008
N	35831	3968	35831	35831
Stage 3
P	1.23E-06	2.45E-03	2.48E-02	1.75E-02
OR	2.58	1.55	0.69	1.58
MAF Cases	0.006	0.012	0.006	0.010
MAF Controls	0.003	0.008	0.007	0.008
N	14884	15288	15288	14876
Stage1, 2 and 3 Meta-Analysis
P	5.38E-24	1.55E-14	5.38E-10	4.56E-10
OR	2.46	1.67	0.68	1.43
MAF Cases	0.004	0.014	0.006	0.011
MAF Controls	0.002	0.009	0.009	0.008
N	80733	53042	84905	84493

Note: Concordance for alternate allele carrier genotypes between imputed versus called SNPs in Stage 3 was 75.2% for rs75932628, 91.1% for rs143332484, 95.7% for rs72824905, and 81.9% for rs616338 (Online Methods and Supplementary Table 6).

Figure 1.

Association plots of PLCG2, ABI3, and TREM2. (a) Regional plot of identified association at the PLCG2 locus. Top hit rs72824905 indicated in purple. Data presented for rs72824905 includes stage 1, stage 2 and stage 3 (N=84,905). (b) Regional plot of identified association at the ABI3 locus. Top hit rs616338 indicated in purple. Data presented for rs616338 includes stage 1, stage 2 and stage 3 (N=84,493). (c) Regional plot of identified association at the TREM2 locus. Top hit rs75932628 indicated in purple. Data presented for rs75932628 and rs143332484 includes stage 1, stage 2 and stage 3 (N=80,733 and 53,042, respectively). SNVs with missing LD information are shown in grey.

The second novel association is a missense change p.S209F (P=4.56×10^-10, OR=1.43) in B3 domain-containing transcription factor ABI3 (ABI3). The p.F209 variant shows consistent evidence for increasing LOAD risk across all stages, with a MAF of 0.011 in cases and 0.008 in controls (Table 2, Figure 1b, Supplementary Table 12, Supplementary Figure 6). The reference allele is conserved across multiple species (Supplementary Figure 7). Gene-wide analysis showed nominal evidence of association (P=5.22×10^-5)(Supplementary Tables 10 & 11). The B4GALNT2 gene, adjacent to ABI3, contained an independent suggestive association (Supplementary Figure 8), but this failed to replicate in subsequent stages (P_combined=1.68×10^-4)(Supplementary Table 7).

Following reports of suggestive association with LOAD^34,35, we report the first evidence for GWS association at TREM2 coding variant p.R62H (P=1.55×10^-14, OR=1.67), with a MAF of 0.0143 in cases and 0.0089 in controls (Table 2, Figure 1c, Supplementary Table 13, Supplementary Figures 9 & 10). We also observed evidence for the previously reported^9,11 TREM2 rare variant p.R47H (Table 2). These variants are not in linkage disequilibrium (Supplementary Table 14) and conditional analyses confirmed that p.R62H and p.R47H are independent risk variants (Supplementary Figure 11). Gene-wide analysis of TREM2 showed a GWS association (P_SKAT=1.42×10^-15)(Supplementary Tables 10 & 11). Removal of p.R47H and p.R62H variants from the analysis diminished the gene-wide association but the signal remains interesting (P_SKAT-O=6.3×10^-3, P_Burden=4.1×10^-3). No single SNV was responsible for the remaining gene-wide association (Supplementary Table 13, Supplementary Figure 11) suggesting that there are additional TREM2 risk variants in TREM2. We previously reported a common variant LOAD association near TREM2, in a GWAS of cerebrospinal fluid tau and P-tau³⁶. We also observed a different suggestive common variant signal in another LOAD case-control study (P=6.3×10^-7)².

We previously identified 8 gene pathway clusters significantly enriched in AD-associated common variants³⁶. To test whether biological enrichments observed in common variants are also present in rare variants we used the rare-variant data (MAF<1%) to reanalyze these eight AD-associated pathway clusters (Online Methods, Supplementary Table 15). We used Fisher’s method to combine gene-wide p-values for all genes in each cluster. After correction for multiple testing, we observed enrichment for immune response (P=8.64×10^-3), cholesterol transport (P=3.84×10^-5), hemostasis (P=2.10×10^-3), Clathrin/AP2 adaptor complex (P=9.20×10^-4) and protein folding (P=0.02). We also performed pathway analyses on the rare variant data presented here using all 9,816 pathways used previously. The top pathways are related to lipoprotein particles, cholesterol efflux, B-cell differentiation and immune response, areas of biology also enriched when common variants are analyzed³⁷(Supplementary Table 16).

Previous analysis of normal brain co-expression networks identified 4 gene modules that are enriched for common variants associated with LOAD risk^2,37,11. These 4 modules are enriched for immune response genes. We identified 151 genes present in 2 or more of these 4 modules and these showed a strong enrichment for LOAD-associated common variants (P=4.0×10^-6)³⁶ and for rare variants described here (MAF<1%)(Supplementary Table 15P=1.17×10^-6). We then used a set of high-quality protein-protein interactions³⁷ to construct, from these 151 genes, an interaction network containing 56 genes, including PLCG2, ABI3 and TREM2 (Figure 2)(Online Methods). This subset is strongly enriched for association signals from both the previous common variant analysis (P=5.0×10^-6, Supplementary Table 17) and this rare variant gene-set analysis (P=1.08×10^-7, Supplementary Table 15). The remaining 95 genes only have nominally-significant enrichment for either common or rare variants (Supplementary Tables 15 & 17), suggesting that the 56-gene (Supplementary Table 18) network is driving the enrichment.

Figure 2.

Protein-protein interaction network (using high-confidence human interactions from the STRING database) of 56 genes enriched for both common and rare variants associated with AD risk. Colours of edges refer to the type of evidence linking the corresponding proteins: red=gene fusion, dark blue = co-occurrence, black = co-expression, magenta = experiments, cyan=databases, light green = text mining, mauve = homology. TREM2, PLCG2 and ABI3 highlighted by red circles, SYK, CSF1R and TYROBP highlighted by blue circles, and INPP5D, SPI1 and CD33 identified as common variant risk loci^2,5-7, highlighted by black circles.

TREM2, ABI3 and PLCG2 have a common expression pattern in human brain cortex, with high expression in microglia cells and limited expression in neurons, oligodendrocytes, astrocytes and endothelial cells (Supplementary Figure 12)³⁸. Other known LOAD loci with the same expression pattern include SORL1, the MS4A gene cluster, and HLA-DRB1. PLCG2, ABI3, and TREM2 are up-regulated in LOAD human cortex and in two APP mouse models. However, when corrected for levels of other microglia genes, these changes in expression appear to be related to microgliosis (Supplementary Tables 19 & 20).

PLCG2 (Supplementary Figure 13) encodes a transmembrane signaling enzyme (PLCγ2) that hydrolyses the membrane phospholipid PIP2 (1-phosphatidyl-1D-myo-inositol 4,5-bisphosphate) to secondary messengers IP3 (myo-inositol 1,4,5-trisphosphate) and DAG (diacylglycerol). IP3 is released into the cytosol and acts at the endoplasmic reticulum where it binds to ligand-gated ion channels to increase cytoplasmic Ca²⁺. DAG remains bound to the plasma membrane where it activates two major signaling molecules, protein kinase C (PKC) and Ras guanyl nucleotide-releasing proteins (RasGRPs), which initiate the NF-κB and mitogen-activated protein kinase (MAPK) pathways. While the IP3/DAG/Ca+2 signaling pathway is active in many cells and tissues, in brain, PLCG2 is primarily expressed in microglial cells. PLCG2 variants also cause Antibody Deficiency and Immune Dysregulation (PLAID) and Autoinflammation and PLAID (APLAID)³⁹. Genomic deletions (PLAID) and missense mutations (APLAID) affect the cSH2 autoinhibitory regulatory region. The result is a complex mix of loss and gain of function in cellular signalling³⁹.

Functional annotation (Supplementary Table 21) suggests ABI3 (Supplementary Figure 14) plays a role in the innate immune response via interferon-mediated signaling⁴⁰. ABI3 is co-expressed with INPP5D (P=2.2×10^-10), a gene previously implicated in LOAD risk². ABI3 plays a significant role in actin cytoskeleton organization through participation in the WAVE2 complex⁴¹, a complex that regulates multiple pathways leading to T-cell activation⁴².

TREM2 encodes a transmembrane receptor present in the plasma membrane of brain microglia (Supplementary Figure 15). TREM2 protein forms an immune-receptor-signaling complex with DAP12. Receptor activation results in activation of Syk and ZAP70 signaling which in turn activates PI3K activity and influences PLCγ2 activity⁴³. In microglia, TREM2-DAP12 induces an M2-like activation⁴⁴ and participates in recognition of membrane debris and amyloid deposits resulting in microglial activation and proliferation^45-47. When TREM2 knockout (KO) or TREM2 heterozygous KO mice are crossed with APP-transgenics that develop plaques, the size and number of microglia associated with plaques are markedly reduced^46,47. TREM2 risk variants are located within exon 2, which is predicted to encode the conserved ligand binding extracellular region of the protein. Any disruption in this region may attenuate or abolish TREM2 signaling, resulting in the loss or decrease in TREM2 function⁴⁷.

The 56-gene interaction network identified here is enriched in immune response genes and includes TREM2, PLCG2, ABI3, SPI1, INPP5D, CSF1R, SYK and TYROBP (Figure 2). SPI1 is a central transcription factor in microglial activation state that has a significant gene-wide association with AD⁵ and is in the proximity of GWS signals identified by IGAP². Loss-of function mutations in CSF1R cause hereditary diffuse leukoencephalopathy with spheroids, a white matter disease related to microglial dysfunction⁴⁸. Activated microglial cells surround plaques^49,50, a finding consistently observed in AD brain and AD transgenic mouse models⁵¹. In AD mouse model brain, synaptic pruning associates with activated microglial signalling⁵². Pharmacological targeting of CSF1R inhibits microglial proliferation and shifts the microglial inflammatory profile to an anti-inflammatory phenotype in murine models⁵³. SYK regulates Aβ production and tau hyperphosphorylation⁵⁴, is affected by the INPP5D/CD2AP complex⁵⁵ encoded by two LOAD associated genes², and mediates phosphorylation of PLCG2⁵⁶. Notably, the anti-hypertensive drug Nilvadipine, currently in a phase III AD clinical trial, targets SYK as well as TYROBP, a hub gene in an AD-related brain expression network³⁸, that encodes the TREM2 complex protein DAP12.

We identified three rare coding variants in PLCG2, ABI3 and TREM2 with GWS associations with LOAD that are part of a common innate immune response. This work provides additional evidence that the microglial response in LOAD is directly part of a causal pathway leading to disease and is not simply a downstream consequence of neurodegeneration^46,47,57,58. Our network analysis supports this conclusion. In addition, PLCγG2, as an enzyme, represents the first classically drug-able target to emerge from LOAD genetic studies. The variants described here account for a small portion of the ‘missing heritability of AD’. The remaining heritability may be due to a large number of common variants of small effect size. For rare variants, there may be additional exonic sites with lower MAF or effect size, and/or intronic and intergenic sites. Complete resolution of AD heritability will be facilitated by larger sample sizes and more comprehensive sequence data.

Data Availability

Summary statistics for the 43 genetic associations identified are provided in Supplementary Table 6.

Stage 1 data (individual level) for the GERAD exome chip cohort can be accessed by applying directly to Cardiff University. Stage 1 ADGC data is deposited in NIAGADS and NIA/NIH sanctioned qualified access data repository. Stage 1 CHARGE data is accessible by applying to dbGaP for all US cohorts, and to ERASMUS University for Rotterdam data. AGES primary data are not available due to Icelandic laws. Stage 2 and stage 3 primary data is available upon request.

A detailed description of the Mayo Clinic RNAseq data is available to all qualified investigators through the Accelerating Medicines Partnership in Alzheimer’s Disease (AMP-AD) knowledge portal that is hosted in the Synapse software platform from Sage Bionetworks (Synapse IDs: syn3157182 and syn3435792 (mouse data), and syn3163039 (human data)).

Online Methods

Genotyping and Quality Control

Stage 1

GERAD/PERADES

Genotyping was performed at Life and Brain, Bonn, Germany, with the Illumina HumanExome BeadChip v1.0 (N=247,870 variants) or v1.1 (N=242,901 variants). Illumina’s GenTrain version 2.0 clustering algorithm in GenomeStudio or zCall¹ was used for genotype calling. Quality control (QC) filters were implemented for sample call rate excluding samples with >1% missingness, excess autosomal heterozygosity excluding outliers based on <1% and >1% minor allele frequency (MAF) separately, gender discordance, relatedness excluding one of each pair related with IBD ≥ 0.125 (the level expected for first cousins), and population outliers (i.e. non European ancestry). Variants were filtered based on call rate excluding variants with >1% missingness, genotype cluster separation excluding variants with a separation score < 0.4 and Hardy-Weinberg equilibrium (HWE) excluding variants with P_HWE < 1×10^-4. Ten principal components (PCs) were extracted using EIGENSTRAT, including the first three PCs as covariates had the maximum impact on the genomic control inflation factor, λ². After QC 6,000 LOAD cases and 2,974 elderly controls (version 1.0; 4,093 LOAD cases and 1,599 controls, version 1.1; 1,907 LOAD cases and 1,375 controls) remained. The version 1.0 array had 244,412 variants available for analysis and 239,814 remained for the version 1.1 array.

CHARGE

All four CHARGE cohorts were genotyped for the Illumina HumanExome BeadChip v1.0. To increase the quality of the rare variant genotype calls, the genotypes for all four studies were jointly called with 62,266 samples from 11 studies at the University of Texas HSC at Houston³. Quality control (QC) procedures for the genotype data were performed both centrally at UT Houston and at each study. The central QC procedures have been described previously³. Minimum QC included: 1) Concordance checking with GWAS data and removal of problematic samples, 2) Removal of individuals with low genotype completion rate (<90%), 3) Removal of variants with low genotype call rate (<95%), 4) Removal of individuals with sex-mismatches, 5) Removal of one individual from duplicate pairs, 6) Removal of first-degree relatives based on genetically calculated relatedness (IBS > 0.45), with cases retained over controls, 7) Removal of variants not called in over 5% of the individuals and those that deviated significantly form the expected Hardy-Weinberg Equilibrium proportions (P<1×10^-6).

ADGC

Genotyping was performed in subsets at four centers: NorthShore, Miami, WashU, and CHOP (“CHOP” and “ADC7” datasets) on the Illumina HumanExome BeadChip v1.0. One variant rs75932628 (p.R47H) in TREM2 clustered poorly across all ADGC cohorts, and was therefore re-genotyped using a Taqman assay. Data on all samples underwent standard quality control procedures applied to genome-wide association studies (GWAS), including excluding variants with call rates <95%, and then filtering samples with call rate <95%. Variants with MAF>0.01 were evaluated for departure from HWE and any variants for P_HWE<10^-6 were excluded. Population substructure within each of the five subsets (NorthShore, Miami, WashU, CHOP, and ADC7) was examined using PC analysis in EIGENSTRAT⁴, and population outliers (>6 SD) were excluded from further analyses; the first three PCs were adjusted for as covariates in association testing. Prior to analysis we harmonized the alternate and reference alleles over all datasets. See Supplementary Table 3 for an overview of cohort genotype calling and quality control procedures. All sample genotyping and quality control was performed blind to participant’s disease status.

Stage 2

Twenty-two variants successfully designed for replication genotyping on the Agena Bioscience MassARRAY^® platform. Genotyping was performed at Life and Brain, Bonn, Germany, and the Centre National de Génotypage (CNG), Paris, France. Twenty-one variants were successfully genotyped, with one variant (rs147163004 in ASTN2) failing visual cluster plot inspection. An additional nine variants were successfully genotyped using the Agena Bioscience MassARRAY^® platform or Thermo FisherTaqMan^® assay at the CNG, Paris, France in a subset of the replication samples N=16,850 (7,755 cases, 9,095 controls).

GERAD/PERADES and ACE QC

Filters were implemented for sample call rate, excluding samples with >10% missingness, and excess autosomal heterozygosity via visual inspection. Variants were filtered based on call rate excluding variants with >10% missingness and HWE excluding variants with P_HWE<1×10^-5 in either cases or controls.

IGAP and EADI QC

Variants were genotyped in 3 different panels and QC was performed in each panel separately. Samples with more than 3 missing genotypes were excluded, as were males heterozygous for X-Chromosome variants present within the genotyped panels. Variants were excluded based on missingness >5%, HWE (in cases and controls separately) <1×10^-5, and differential missingness between cases and controls <1×10^-5, for each Country cohort. All variants passed quality control. PCs were determined using previously described methods¹⁹.

Stage 3

Replication was performed using genotypes from 23 ADGC datasets as described above. Genotyping arrays used have been described in detail before for most datasets, except for the CHAP, NBB, TARCC, and WHICAP datasets. CHAP and WHICAP datasets were genotyped on the Illumina OmniExpress-24 array, while NBB was genotyped on the Illumina 1M platform. TARCC first wave subjects were genotyped using the Affymetrix 6.0 microarray chip, while subjects in the second wave (172 cases and 74 controls) were genotyped using the Illumina HumanOmniExpress-24 beadchip. Second wave TARCC subjects (TARCC2) were genotyped together with 84 cases and 115 controls from second wave samples ascertained at the University of Miami and Vanderbilt University. All samples used in stage 3 were imputed to the HRC haplotype reference panel^5,6, which includes 64,976 haplotypes with 39,235,157 SNPs that allows imputation down to an unprecedented MAF=0.00008.

Prior to imputation, all genotype data underwent QC procedures that have been described extensively elsewhere^7,8. Imputation was performed on the Michigan Imputation Server (https://imputationserver.sph.umich.edu/) running MiniMac3^9,10. Genotypes from genome-wide, high-density SNP genotyping arrays for 16,175 AD cases and 17,176 cognitive-normal individuals were imputed. Across all samples 39,235,157 SNPs were imputed, with the actual number of SNPs imputed for each individual varying based on the regional density of array genotypes available. As a subset of these samples had also been genotyped as part of stage 1, we examined the imputation quality for critical variants by comparing imputed genotypes to those directly genotyped by the exome array; overall concordance was >99%, while concordance among alternate allele genotypes (heterozygotes and alternate allele homozygotes) was >88.5% on average (N=13,000 samples). Concordance between Stage 3 imputed genotypes and exome chip genotypes for replicated SNPs is reported in Supplementary Table 6.

Analysis

Stage 1

We tested association with LOAD using logistic regression modelling for common and low frequency variants (MAF>1%) and implementing maximum likelihood estimation using the score test and ‘seqMeta’ package for rare variation (MAF≤1%). Analyses were conducted globally in the GERAD/PERADES consortium, and for each contributing centre in the CHARGE and ADGC consortia under two models (1) an ‘unadjusted’ model, which included minimal adjustment for possible population stratification, using Country of origin and the first three principal components from PCA, and (2) an ‘adjusted’ model, which included covariates for age, and sex, as well as Country of origin and the first three principal components. Age was defined as the age at onset of clinical symptoms for cases, and the age at last interview for cognitively normal controls.

Meta-analysis for common and low frequency variants were undertaken in METAL using a fixed-effects inverse variance-weighted meta-analysis. Rare variants were meta-analysed in the SeqMeta R package. In the SeqMeta pipeline, cohort-level analyses generated score statistics through the function ‘prepScores()’ which were captured in *. Rdata objects. These *. Rdata objects contain the necessary information to meta-analyse SKAT analyses: the individual SNP scores, MAF, and a covariance matrix for each unit of aggregation. Using the ‘singlesnpMeta()’ and ‘skatOmeta()’ functions of SeqMeta, the *. Rdata objects for individual studies were meta-analysed. The seqMeta coefficients and standard errors can be interpreted as a ‘one-step‘ approximation to the maximum likelihood estimates. Monomorphic variants in individual studies were not excluded as they contribute to the minor allele frequency information. Three independent analysts confirmed the meta-analysis results.

In the GERAD/PERADES consortium 1,740 participants (888 LOAD cases and 852 controls) did not have age information available and were excluded from the adjusted analyses. Therefore, 16,160 cases and 17,967 controls were included in the unadjusted analyses and 15,272 cases and 17,115 controls were included in the adjusted analyses. The primary analysis utilized the unadjusted model given the larger sample size this provided. See Supplementary Figure 2 for QQ plots of unadjusted and adjusted analyses.

Stage 2

We tested association with LOAD using the score test and ‘seqMeta’ package. Analyses were conducted under the two models described above, in the analysis groups indicated in Supplementary Table 2. Analyses were undertaken globally in the GERAD/PERADES cohort and by Country in the IGAP cohorts, with the EADI1 cohort only including French participants and the ACE cohort including only Spanish participants. Following the format of the IGAP mega meta-analysis⁷, four PCs were included for the EADI1 dataset, and one in the Italian and Swedish IGAP clusters. Meta-analysis was undertaken in the SeqMeta R package.

Stage 3

Association analyses performed followed Stage 1 and Stage 2 analytical procedures described below, and only variants in ABI3, PLCG2 and TREM2 were examined. For gene-based testing, 10 variants in ABI3, 35 in PLCG2, and 13 in TREM2 were examined.

Pathway/Gene-set Enrichment Analysis

The eight biological pathway clusters previously identified as enriched for association in the IGAP dataset¹¹ were tested for enrichment in this rare variation study (Supplementary Table 15) in order to test whether the biological enrichments observed in common variants also apply to rare variants. Genes were defined without surrounding genomic sequence, as this yielded the most significant excess of enriched pathways in the common variation dataset¹¹. Gene-wide SKAT-O P-values for the variants of interest were combined using the Fisher’s combined probability test. Given the low degree of LD¹² between rare variants our primary analyses did not control for LD between pathway genes. However, as a secondary analysis, the APOE region was removed, and for each pair of pathway genes within 1Mb of each other, the gene with the more significant SKAT-O P-value was removed. This highly conservative procedure removes any potential bias in the enrichment test both from LD between the genes, and also from dropping less significant genes from the analysis.

We also performed pathway analyses on the rare variant data presented here using all 9,816 pathways used previously. The top pathways are related to lipoprotein particles, cholesterol efflux, B-cell differentiation and immune response, and closely parallel the common variant results (Supplementary Table 16).

Protein interaction Analysis

Previous analysis of normal brain co-expression networks identified 4 gene modules that were enriched for common variants associated with AD risk in the IGAP GWAS. Each of these 4 modules was also found to be enriched for immune-related genes. The 151 genes present in 2 or more of these 4 modules were particularly strongly enriched for IGAP GWAS association⁴¹. This set of 151 co-expressed genes thus contains genes of relevance to AD aetiology. To identify these genes, and clarify biological relationships between them for future study, protein interaction analysis was performed. First, a list of high-confidence (confidence score >0.7) human protein-protein interactions was downloaded from the latest version (v10) of the STRING database (http://string-db.org). Then, protein interaction networks were generated as follows:

1. Choose a gene to start the network (the “seed” gene)

2. For each remaining gene in the set of 151 genes, add it to the network if its corresponding protein shows a high-confidence protein interaction with a protein corresponding to any gene already in the network.

3. Repeat step 2 until no more genes can be added

4. Note the number of genes in the network

5. Repeat, choosing each of the 151 genes in turn as the seed gene.

The largest protein interaction network resulting from this procedure resulted in a network of 56 genes connected by high-confidence protein interactions. To test whether this network was larger than expected by chance, given the total number of protein-protein interactions for each gene, random sets of 151 genes were generated, with each gene chosen to have the same total number of protein-protein interactions as the corresponding gene in the actual data. Protein networks were generated for each gene as described above, and the size of the largest such network compared to the observed 56-gene network. 1000 random gene sets were generated, and none of them yielded a protein interaction network as large as 56 genes. Note that the procedure for generating the protein interaction network relies only on protein interaction data, and is agnostic to the strength of GWAS or rare-variant association for each gene. Thus the strength of genetic association in the set of 56 network genes can be tested relative to that in the original set of 151 genes without bias.

Gene-set enrichment analysis of the protein network

The set of 56 network genes was tested for association enrichment in the IGAP GWAS using ALIGATOR¹³, as was done in the original pathway analysis, using a range of p-value thresholds for defining significant SNPs (and thus the genes containing those SNPs). The same analysis was also performed on the 95 genes in the module overlap but not the protein interaction network (Supplementary Table 17). It can be seen that the 56 network genes account for most of the enrichment signal observed in the set of 151 module overlap genes.

The set of 56 network genes, the set of 151 module overlap genes, and the set of 95 genes in the module overlap but not the network were tested for enrichment of association signal in variants with MAF<1% using the gene set enrichment method described above in section 11. Both the set of 151 genes (P=1.17×10^-6) and the subset of 56 genes (P =1.08×10^-7) show highly significant enrichment for association in the rare variants with MAF<1%. It can be seen that the 56 network genes account for most of the enrichment signal observed in the set of 151 module overlap genes (Supplementary Table 17). Again, the subset of 56 genes accounts for most of the enrichment signal observed in the set of 151 genes, as the remaining 95 genes have only nominally-significant enrichment (P=0.043). Both the set of 151 genes (P=5.15×10^-5) and the subset of 56 genes (P=2.98×10^-7) show significant enrichment under a conservative analysis excluding the APOE region and correcting for possible LD between the genes (Supplementary Table 17). Thus, the rare variants show convincing replication of the biological signal observed in the common variant GWAS, and furthermore, the protein network analysis has refined this signal to a set of 56 interacting genes. Given that TREM2 has a highly significant gene-wide p-value (P=1.01×10^-13) among variants with MAF<1%, enrichment analyses were run omitting it. Both the set of 151 genes (P=2.78×10^-3) and the subset of 56 genes (P=0.010) (Supplementary Table 18) still showed significant enrichment of signal, suggesting that the contribution of rare variants to disease susceptibility in these networks is not restricted to TREM2. Biological follow-up of genetic results is labour-intensive and expensive. It is therefore important to concentrate such work on the genes that are most important to AD susceptibility. Thus, the rationale for reducing the gene set is that it defines a network of genes that are not only related through co-expression and protein interaction, but also show enrichment for genetic association signal. These genes are therefore strong candidates for future biological study.

Gene Expression

We examined mRNA expression of the novel genes PLCG2 and ABI3 in neuropathologically characterized brain post-mortem tissue (508 persons): they are expressed at low levels in the dorsolateral prefrontal cortex of subjects from two studies of aging with prospective autopsy (ranked 12,965th out of 13,484 expressed genes)¹⁴. However, ABI3 and PLCG2 were more highly expressed in purified microglia/macrophage from the cortex of 11 subjects from these cohorts (1740th and 2600th respectively out of the 11,500 expressed genes)(unpublished data). These findings are consistent with the high levels of expression of both PLCG2 and ABI3 in peripheral monocytes, spleen, and whole blood reported by the ROADmap project and in microglia as reported by Zhang et al¹⁵. From the same brain tissue, we examined methylation (n=714)¹⁶ and H3K9ac acetylation (n=676) data and found differential methylation at four CpG sites and lower acetylation at two H3K9ac sites adjacent to PLCG2 and ABI3 in relation to increased global neuritic plaque and tangle burden (FDR < 0.05). Similarly, high TREM2 expression has been shown to correlate with increasing neuritic plaque burden¹⁷.

AMP-AD Gene Expression Data

RNA sequencing was used to measure gene expression levels in the temporal cortex of 80 subjects with pathologically confirmed AD and 76 controls without any neurodegenerative pathologies obtained from the Mayo Clinic Brain Bank and the Banner Sun Health Institute. The human RNA sequencing data is deposited in the Accelerating Medicines Partnership-AD (AMP-AD) knowledge portal housed in Synapse (https://www.synapse.org/#!Synapse:syn2580853/wiki/66722). After QC, our postmortem human cohort has 80 subjects with pathologically confirmed AD and 76 controls without any neurodegenerative pathologies. Assuming two samples of 100 per group, two-sample t-test, same standard deviation, we will have 80% power to detect effect sizes of 0.40, 0.49 and 0.59 at p<0.05, 0.01 and 0.001, respectively, where effect size is the difference in means between two groups divided by the within-group standard deviation. The human RNA sequencing data overview, QC and analytic methods are available at the following Synapse pages, respectively: syn3163039, syn6126114, syn6090802. Multivariable linear regression was used to test for association of gene expression levels with AD diagnosis (Dx) using two different models: In the Simple model, we adjust for age at death, sex, RNA integrity number (RIN), tissue source, and RNAseq flowcell. In the Comprehensive model, we adjust for all these covariates, and brain cell type markers for five cell-specific genes (CD68 (microglia), CD34 (endothelial), OLIG2 (oligodendroglia), GFAP (astrocyte), ENO2 (neuron)) to account for cell number changes that occur with AD neuropathology. TREM2, PLCG2 and ABI3 are significantly higher in AD temporal cortex prior to correcting for cell types (Simple model), but this significance is abolished after adjusting for cell-specific gene counts (Comprehensive model). This suggests that these elevations are likely a consequence of changes in cell types that occur with AD, most likely microgliosis given that TREM2, PLCG2 and ABI3 are microglia-enriched genes¹⁵ (Supplementary Table 19, Supplementary Figure 12).

Main Authors by Consortium and Author Contributions

Superscript number refers to institutional affiliation. This can be found below the main author list at the beginning of this article.

GERAD/PERADES

Rebecca Sims¹, Nandini Badarinarayan¹, Rachel Raybould¹, Stefanie Heilmann-Heimbach^12,13, Maria Vronskaya¹, Per Hoffmann^12,13,21, Markus M. Nöthen^12,13, Wolfgang Maier^30,31, Stefan Herms^12,13,21, Andreas J. Forstner^12,13, Denise Harold³⁵, Rhodri Thomas¹, Taniesha Morgan¹, Nicola Denning¹, Elisa Majounie¹, Michelle K Lupton^43,44, Christopher Medway⁴⁷, Kristelle Brown⁴⁷, Bernadette McGuinness⁵⁰, Petra Proitsi⁴³, Pau Pastor⁵⁴, Ana Frank-García^56,57,58, Ina Giegling⁶³, Harald Hampel^65,66,67, Patrizia Mecocci⁵⁴, Virginia Boccardi⁵⁴, Martin Scherer⁷², Markus Leber⁷⁴, Steffi Riedel-Heller⁷⁶, Anne Braae⁷⁹, Carlo Masullo⁸³, Gianfranco Spalletta⁸⁶, Paola Bossù⁸⁶, Eleonora Sacchinelli⁸⁶, Pascual Sánchez-Juan³⁷, Frank Jessen^30,31,74, John Morris^97,98, Chris Corcoran¹⁰⁰, JoAnn Tschanz¹⁰⁰, Maria Norton¹⁰⁰, Ron Munger¹⁰⁰, María J Bullido^57,58,106, Eliecer Coto¹⁰⁷, Victoria Alvarez¹⁰⁷, Maura Gallo¹¹⁶, Amalia Cecilia Bruni¹¹⁶, Martin Dichgans^119,120, Daniela Galimberti¹²², Elio Scarpini¹²², Michelangelo Mancuso¹²⁷, Ubaldo Bonuccelli¹²⁷, Antonio Daniele¹²⁹, GERAD/PERADES, Oliver Peters¹³¹, Benedetta Nacmias^132,133, Matthias Riemenschneider¹³⁴, Reinhard Heun³¹, Carol Brayne¹³⁵, David C Rubinsztein¹²³, Jose Bras^136,137, Rita Guerreiro^136,137, John Hardy¹³⁶, Ammar Al-Chalabi¹³⁸, Christopher E Shaw¹³⁸, John Collinge¹³⁹, David Mann¹⁴⁰, Magda Tsolaki¹⁴¹, Jordi Clarimón^58,142, Rebecca Sussams¹⁴³, Simon Lovestone¹⁴⁴, Michael C O’Donovan¹, Michael J Owen¹, Simon Mead¹³⁹, Clive Holmes¹⁴³, John Powell⁴³, Kevin Morgan⁴⁷, Peter Passmore⁵⁰, Dan Rujescu⁶³, Sara Ortega-Cubero^58,154, John “Keoni” Kauwe¹⁵⁷, Lesley Jones¹, Valentina Escott-Price¹, Peter A Holmans¹, Alfredo Ramirez^12,31,74, Julie Williams¹.

Study design or conception

Rebecca Sims, Valentina Escott-Price, Michael C O’Donovan, Michael J Owen, Peter A. Holmans, Julie Williams

Sample contribution

Markus M. Nöthen, Wolfgang Maier, Stefan Herms, Andreas J. Forstner, Julie Williams, Alfredo Ramirez, Michelle K Lupton, Christopher Medway, Kristelle Brown, Bernadette McGuinness, Petra Proitsi, Pau Pastor, Ana Frank-García, Ina Giegling, Harald Hampel, Patrizia Mecocci, Virginia Boccardi, Martin Scherer, Markus Leber, Steffi Riedel-Heller, Anne Braae, Carlo Masullo, Gianfranco Spalletta, Paola Bossù, Eleonora Sacchinelli, Pascual Sánchez-Juan, Frank Jessen, John Morris, Chris Corcoran, JoAnn Tschanz, Maria Norton, Ron Munger, María J Bullido, Eliecer Coto, Victoria Alvarez, Maura Gallo, Amalia Cecilia Bruni, Martin Dichgans, Daniela Galimberti, Elio Scarpini, Michelangelo Mancuso, Ubaldo Bonuccelli, Antonio Daniele, Oliver Peters, Benedetta Nacmias, Matthias Riemenschneider, Reinhard Heun, Carol Brayne, David C Rubinsztein, Ammar Al-Chalabi, Christopher E Shaw, John Collinge, David Mann, Magda Tsolaki, Jordi Clarimón, Rebecca Sussams, Simon Lovestone, Simon Mead, Clive Holmes, John Powell, Kevin Morgan, Peter Passmore, Dan Rujescu, Sara Ortega-Cubero, John “Keoni” Kauwe,

Data generation

Rebecca Sims, Rachel Raybould, Stefanie Heilmann-Heimbach, Per Hoffmann, Rhodri Thomas, Taniesha Morgan, Nicola Denning, Alfredo Ramirez, Julie Williams, Jose Bras, Rita Guerreiro, John Hardy

Analysis

Rebecca Sims, Nandini Badarinarayan, Maria Vronskaya, Denise Harold, Elisa Majounie, Peter A. Holmans

Manuscript preparation

Rebecca Sims, Lesley Jones, Peter A. Holmans, Julie Williams

Study supervision/management

Rebecca Sims, Alfredo Ramirez, Julie Williams

ADGC

Adam C. Naj³, Brian W. Kunkle⁸, Eden R. Martin^8,11, Amanda B. Kuzma¹⁶, Robert R. Graham²⁰, Badri N. Vardarajan^23,24,25, Kara L. Hamilton-Nelson⁸, Gary W. Beecham^8,11, Cory C. Funk³⁴, Hongdong Li³⁴, Otto Valladares¹⁶, Liming Qu¹⁶, Yi Zhao¹⁶, John Malamon¹⁶, Beth Dombroski¹⁶, Patrice Whitehead⁸, Shubhabrata Mukherjee⁴⁸, Laura B. Cantwell¹⁶, Jeremy D. Burgess⁵¹, Mariet Allen⁵¹, Nathan D Price³⁴, Paramita Chakrabarty⁶⁰, Xue Wang⁵¹, Paul K. Crane⁴⁸, Robert C. Barber⁷¹, Perrie M. Adams⁷³, Marilyn S. Albert⁷⁵, Duane Beekly⁷⁸, Deborah Blacker^81,82, Rachelle S. Doody⁸⁵, Thomas J. Fairchild⁸⁹, Matthew P. Frosch⁹², Bernardino Ghetti⁹³, Ryan M. Huebinger⁹⁴, M. Ilyas Kamboh^95,96, Mindy J. Katz⁹⁹, C. Dirk Keene¹⁰¹, Walter A. Kukull⁸⁸, Eric B. Larson^48,104, Richard B. Lipton⁹⁹, Thomas J. Montine¹⁰¹, Ronald C. Petersen¹⁰⁹, Eric M. Reiman^{112,113,114,115}, Joan S. Reisch^59,117, Donald R. Royall¹¹⁸, Mary Sano¹²¹, Peter St George-Hyslop^123,124, Debby W. Tsuang^125,126, Ashley R. Winslow¹²⁸, Chuang-Kuo Wu¹³⁰, ADGC, Timothy W. Behrens²⁰, Alison M. Goate^98a, Carlos Cruchaga^97,98, Todd E. Gold^60,146, Nilufer Ertekin-Taner^51,109, Steven G Younkin^51,109, Dennis W. Dickson⁵¹, Hakon Hakonarson¹⁵⁶, Lindsay A. Farrer¹⁴, Johnathan Haines¹⁵⁹, Richard Mayeux^23,24,25, Margaret A. Pericak-Vance^8,11, Li-San Wang¹⁶, Gerard D. Schellenberg¹⁶.

Study design or conception

Lindsay A. Farrer, Johnathan Haines, Richard Mayeux

Margaret A. Pericak-Vance, Li-San Wang, Gerard D. Schellenberg

Sample contribution

Ashley R. Winslow, Shubhabrata Mukherjee, Paul K. Crane, Robert C. Barber, Perrie M. Adams, Marilyn S. Albert, Deborah Blacker, Rachelle S. Doody, Thomas J. Fairchild, Matthew P. Frosch, Bernardino Ghetti, Ryan M. Huebinger, M. Ilyas Kamboh, Mindy J. Katz, C. Dirk Keene, Eric B. Larson, Richard B. Lipton, Thomas J. Montine, Ronald C. Petersen, Eric M. Reiman, Joan S. Reisch, Donald R. Royall, Mary Sano, Peter St George-Hyslop, Debby W. Tsuang, Chuang-Kuo Wu, Alison M. Goate, Carlos Cruchaga, Steven G Younkin, Dennis W. Dickson, Walter A. Kukull, Nilufer Ertekin-Taner

Data generation

Otto Valladares, Liming Qu, Yi Zhao, John Malamon, Cory C. Funk, Hongdong Li, Jeremy D. Burgess, Mariet Allen, Nathan D Price, Paramita Chakrabarty, Xue Wang, Todd E. Golde, Hakon Hakonarson, Timothy W. Behrens, Beth Dombroski, Walter A. Kukull, Nilufer Ertekin-Taner

Analysis

Adam C. Naj, Brian W. Kunkle, Eden R. Martin, Amanda Partch, Robert R. Graham, Badri N. Vardarajan, Kara L. Hamilton-Nelson, Gary W. Beecham

Manuscript preparation

Adam C. Naj, Gerard D. Schellenberg

Study supervision/management

Gerard D. Schellenberg, Laura B. Cantwell, Duane Beekly, Patrice Whitehead

CHARGE

Sven J. van der Lee², Johanna Jakobsdottir⁷, Joshua C. Bis¹⁰, Vincent Chouraki^14,15, Agustin Ruiz¹⁹, Megan L. Grove²², Charles C. White²⁸, Seung-Hoan Choi^14,32, Albert V. Smith^7,33, Shahzad Ahmad², Claudia L. Satizabal^14,15, Jennifer A. Brody¹⁰, Frank J. Wolters², Myriam Fornage³⁹, Thomas H. Mosley⁴⁰, Josée Dupuis³², Xueqiu Jian³⁹, Honghuang Lin¹⁶², Sonia Moreno-Grau¹⁹, Hieab H. Adams², L. Adrienne Cupples^15,32, Daniel Levy^14,15,59, Alexa S. Beiser^15,32, Melissa E. Garcia⁶⁸, Gudny Eiriksdottir⁷, Isabel Henández¹⁹, Lluis Tarraga¹⁹, Yuning Chen³², Valur Emilsson^7,77, Reinhold Schmidt⁸⁰, Helena Schmidt⁸⁴, WT Longstreth Jr^87,88, Oscar L. Lopez^90,91, Qiong Yang³², Shuo Li³², Oscar Sotolongo-Grau¹⁹, Jayanadra J. Himali¹⁴, Annette L. Fitzpatrick^88,102, Thor Aspelund^7,103, Jerome I. Rotter¹⁰⁵, Albert Hofman², Eric Boerwinkle^22,108, Fernando Rivadeneira^2,110,111, Christopher J. O’Donnell¹⁵, CHARGE, Andre G. Uitterlinden^2,110,111, David A. Bennett¹⁴⁵, Philip L. De Jager¹⁴⁸, Bruce M. Psaty^{10,88,104,149}, Vilmundur Gudnason^7,33, Anita L. DeStefano^15,32, Merce Boada¹⁹, M. Arfan Ikram^2,158, Lenore J. Launer⁶⁸, Najaf Amin², Cornelia M. van Duijn², Sudha Seshadri^14,15.

Study design or conception

S. J. van der Lee, Joshua C. Bis, Philip L. De Jager, Vilmundur Gudnason, Anita L. DeStefano, Lenore J. Launer, Najaf Amin, Cornelia M. van Duijn, Sudha Seshadri.

Sample contribution

Joshua C. Bis, Agustin Ruiz, Megan L. Grove, Claudia L. Satizabal, Frank J. Wolters, Thomas H. Mosley, Alexa S. Beiser, Melissa E. Garcia, Gudny Eiriksdottir, Reinhold Schmidt, Helena Schmidt, WT Longstreth, Jr, Oscar L. Lopez, Jayanadra J. Himali, Annette L. Fitzpatrick, Albert Hofman, David A. Bennett, Philip L. De Jager, Bruce M. Psaty, Vilmundur Gudnason, Merce Boada, M. Arfan Ikram, Lenore J. Launer.

Data generation

Sven J. van der Lee, Agustin Ruiz, Fernando Rivadeneira, Andre G., Uitterlinden, Joshua C. Bis, Megan L. Grove, Helena Schmidt, Johanna Jakobsdottir, Albert V. Smith, Jennifer A. Brody, Myriam Fornage, Xueqiu Jian, Honghuang Lin, L. Adrienne Cupples, Daniel Levy, Qiong Yang, Thor Aspelund, Eric Boerwinkle, Christopher J. O’Donnell, Merce Boada, Shahzad Ahmad, Sonia Moreno-Grau, Hieab H. Adams, Isabel Henández, Lluis Tarraga, Oscar Sotolongo-Grau, Najaf Amin

Analysis

Sven J. van der Lee, Agustin Ruiz, Joshua C. Bis, Megan L. Grove, Helena Schmidt, Johanna Jakobsdottir, Albert V. Smith, Sonia Moreno-Grau, Najaf Amin, Vincent Chouraki, Charles C. White, Seung-Hoan Choi, Josée Dupuis, Yuning Chen, Shuo Li, Anita L. DeStefano

Manuscript preparation

Sven J. van der Lee, Agustin Ruiz, Joshua C. Bis, Johanna Jakobsdottir, Vincent Chouraki, Charles C. White, Cornelia M. van Duijn, Sudha Seshadri

Study supervision/management

Cornelia M. van Duijn, Sudha Seshadri, M. Arfan Ikram

EADI

Céline Bellenguez^4,5,6, Anne Boland⁹, Benjamin Grenier-Boley^4,5,6, Kristel Sleegers^17,18, Robert Olaso⁹, Mikko Hiltunen^26,27, Jacques Epelbaum²⁹, Cécile Dulary⁹, Céline Derbois⁹, Delphine Bacq⁹, Alessio Squassina³⁶, Pascual Sanchez-Juan³⁷, Florentino Sanchez Garcia³⁸, Maria Candida Deniz Naranjo³⁸, David Wallon^41,42, Laura Fratiglioni^45,46, Lina Keller⁴⁶, Francesco Panza⁴⁹, Vincenzo Solfrizzi⁵², Davide Seripa⁵³, Didier Hannequin^42,55, Paolo Caffarra^61,62, Vimantas Giedraitis⁶⁴, Lars Lannfelt⁶⁴, Florence Pasquier^69,70, EADI, Martin Ingelsson⁶⁴, Caroline Graff^45,147, Onofre Cambarros³⁷, Claudine Berr^150,151,152, Jean-Francois Dartigues¹⁵³, Dominique Campion^41,42, Christine Van Broeckhoven^17,18, Christophe Tzourio^160,161, Jean-François Deleuze⁹, Philippe Amouyel^4,5,6,69, Jean-Charles Lambert^4,5,6.

Study design or conception

Phillippe Amouyel, Jean-Charles Lambert

Sample contribution

Jacques Epelbaum, David Wallon, Didier Hannequin, Florence Pasquier, Claudine Berr, Jean-Francois Dartigues, Dominique campion, Christophe Tzourio, Phillippe Amouyel, Jean-Charles Lambert, Vincent Dermecourt, Nathalie Fievet, Olivier Hanon, Carole Dufouil, Alexis Brice, Karen Ritchie, Bruno Dubois, Kristel Sleegers, Mikko Hiltunen, Maria Del Zompo, Ignacio Mateo, Florentino Sanchez Garcia, Maria Candida Deniz Naranjo, Laura Fratiglioni, Lina Keller, Francesco Panza, Paolo Caffarra, Lars Lannfelt, Martin Ingelsson, Caroline Graff, Onofre Cambarros, Christine Van Broeckhoven, Sebastien Engelborghs, Rik Vandenberghe, Peter P. De Deyn, Alession Squassina, Pascual Sanchez-Juan, Carmen, Munoz Fernadez, Yoland Aladro Benito, Hakan Thonberg, Charlotte Forsell, Lena Lilius, Anne Kinhult-stählbom, Vilmantas Giedraitis, Lena Kilander, RoseMarie Brundin, Letizia Concari, Seppo Helisalmi, Anne Maria Koivisto, Annakaisa Haapasalo, Vincenzo Solfrizzi, Vincenza Frisardi

Data generation

Anne Boland, Robert Aloso, Cécile Dulary, Céline Derbois, Delphine Bacq, Jean-François Deleuze, Fabienne Garzia, Feroze Golamaully, Gislain Septier

Analysis

Céline Bellenguez, Benjamin Grenier-Boley, Phillippe Amouyel, Jean-Charles Lambert

Manuscript preparation

Céline Bellenguez, Jean-Charles Lambert Study

supervision/management

Phillippe Amouyel, Jean-Charles Lambert