Study sample

To identify AD-associated rare-variant effects, we evaluated three publicly available large-scale WGS datasets collected for LOAD patients, downloaded as multi-sample VCF files. The Alzheimer’s disease sequencing project (ADSP) Umbrella is a collection of sequencing data from the ADSP and other AD and Related Dementia studies. Under this Umbrella, the ADSP group sequenced a large number of well-characterized Alzheimer’s Disease (AD) patients at three National Human Genome Research Institute Genome Centers (NHGRI) (Baylor College of Medicine Human Genome Sequencing Center, the Broad Institute, and the McDonnell Genome Institute at Washington University). The ascertainment methods and inclusion criteria are described in detail on the National Institute on Aging Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS).[51,52] The sequencing results were mapped to the human reference genome (GRCh38) and processed using the VCPA 1.0 pipeline, which follows GATK best-practices pipeline.[53] Details of the variant calling pipeline can also be found on the NIAGADS. The ADSP discovery extension phase sequenced whole genomes of 1,466 cases and 1,534 controls from five cohorts provided by the Alzheimer’s Disease Genetics Consortium (ADGC) and included samples with diverse ancestry backgrounds (Non-Hispanic White, Caribbean Hispanic, and African American). Another WGS project shared under the ADSP Umbrella is the Alzheimer’s Disease Neuroimaging Initiative (ADNI), which is a longitudinal multi-center (63 sites across North America) study designed for early detection and tracking of AD. The ADNI WGS data contains 808 participants with 238 AD cases, 322 mild cognitive control (MCI) subjects, and 248 controls. A full list of the ascertainment methods and inclusion criteria can be found in detailed descriptions in the online ADNI protocol.[54] As of 2018, the ADNI was recalled under the same VCPA 1.0 pipeline as the ADSP discovery extension WGS data and mapped to the same human reference genome (GRCh38), which were then released together. This combined ADSP case-control dataset contained WGS data from a total of 3,896 individuals (accessed by us on Nov 20, 2018), which then underwent a sequence of quality control steps discussed later before including in our stage 1 analysis. Detailed demographic information of this dataset can be found in Table 1 and the distribution of age among AD cases and controls in S12 Fig. To note, we removed samples in the MCI category to ensure a strict bipartite definition of disease status from all our analyses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of clinical, demographic, and technical information of individuals from three large WGS datasets.

https://doi.org/10.1371/journal.pgen.1009772.t001

Our stage 2 replication included 1,894 WGS samples from the Accelerating Medicines Partnership-Alzheimer’s Disease (AMP-AD) Target Discovery and Preclinical Validation Project (accessed by us on Dec 13, 2018). The samples were separately sequenced at three centers: the Religious Orders Study and Memory and Aging Project (ROSMAP) (1,200 samples), the Mount Sinai Brain Bank (MSBB) study (354 samples), and the Mayo Clinic Brain Bank (Mayo) (350 samples). Previously reports have the detailed data collection scheme and sample inclusion and exclusion criteria.[55–58] This sequence data were mapped to the human reference genome (GRCh37) and were processed using the GATK best-practices workflow v3.4.0.[58] Another stage 2 replication was performed on the ADSP discovery extension phase family samples, which were released together with the ADSP case-control data. Therefore, this family WGS data were also mapped to the human reference genome (GRCh38) and processed using the VCPA 1.0 pipeline. The ADSP WGS family dataset contains 888 samples from 161 multiplex families. The inclusion criteria prioritized families loaded with LOAD with minimal APOE ε4 alleles. A detailed description of the study design and sample ascertainment methods can be found in previous reports.[59,60] For AMP-AD case-control study, this resulted in 642 AD patients and 969 controls after removing low-quality samples. For the ADSP family study, we obtained 545 AD patients and 285 cognitively normal older individuals (Table 1).

RNA-Seq data used for functional analysis were also obtained from the ROSMAP study of the AMP-AD Consortium [58]. The bulk RNA-Seq data were generated for 636 samples (254 AD cases, 368 controls, 12 other dementia, and two without annotation) from the dorsolateral prefrontal cortex (DLPFC) tissues by the Broad Institue’s Genomics Platform and processed in an automatic and parallelized pipeline.[55] The ROSMAP group also selected 48 post-mortem samples (24 with severe AD pathology and 24 with low-to-no pathology) and conducted droplet-based single-nucleus RNA sequencing of the prefrontal cortex region.[61] Metadata of the RNA-Seq data were then used to map samples to cases and controls following the same rule as in stage 2 replication, as well as to merge with genotyping data.

Data processing and quality control of WGS data

Identification and annotation of rare variants

To identify rare variants, we used both external and internal sources of allele frequency to avoid potential inflation of the allele frequency introduced by the study design. For the external sources, we looked at the Europeans (EUR) in 1KG[63] and Non-Finnish Europeans (NFE) in the gnomAD v2 database [65], which matched the ancestral backgrounds of our datasets. We used two different MAF thresholds (0.1% and 1%) to define rare variants, as there is no one consensus definition of rarity and we will correct for testing multiple MAF thresholds in future analysis. In practice, when a variant was present in either of 1KG EUR or gnomAD v2 NFE samples and below the aforementioned threshold, we would keep it for further analysis. For the internal sources, we retained only samples with European ancestry based on the previous PCA, as different ancestral groups would have different allele frequency distributions. Then when a variant was absent in both external databases, we would look at the MAF estimated from the European samples within our dataset and selected rare variants based on 0.1% and 1% MAF thresholds separately. We then annotated rare variants using Ensembl Variant Effect Predictor (VEP)[66]. We defined a variant to be ‘deleterious’ if it is within one of the following categories: stop-gain, stop-loss, frameshift, splice-donor, splice-acceptor, and missense variants. Particularly, for missense variants, we additionally consulted PolyPhen-2[67] and retained only confident missense variants predicted to be ‘damaging.’ This definition of deleteriousness focused on coding regions, primarily due to the fact that the effect of non-coding variants was challenging to predict.[68,69] A distribution of variant types and singletons among the selected set of rare deleterious variants could be found in S9 Fig and S11 Table. In an additional validation of the deleteriousness, we further introduced the CADD score [70] as a third deleterious criterion in phase 1 analysis combined with VEP and PolyPhen-2. The distribution of CADD scores among the set of rare deleterious variants could be found in S8 Fig. As suggested by the CADD documentation, variants with scaled CADD > 15 were retained as pathogenic variants and the set of rare deleterious/pathogenic variants passing all three annotation tools were used in this validation test.

Identification of genes in endocytic pathways

We identified genes involved in endocytic pathways using AmiGO 2 [71,72] gene ontology database to select all genes participating in this pathway. We identified three specific GO terms related to the endocytic system in the Homo Sapiens category, which corresponded to three specific compartments in the endocytic system (endosome, lysosome, and trans-Golgi network). The endosome compartment is a membrane-bound vacuole in eukaryotic, participating in the endocytic trafficking from the trans-Golgi network to the plasma membrane and vice versa.[73] The trans-Golgi network serves as an interconnected tubular network and the final cisternal structure involved in packaging and transporting of cargos to the lysosome, endosome, and cell surface.[74] The lysosome, a small membrane-bound lytic vacuole, is one of the end-point in the endocytic transporting pathway, which contains hydrolytic enzymes to break down various biomolecules.[75] The combination of these three compartments formed the essential backbone of the endocytic system, which we named as “endo-system” and used this term throughout the paper. After removing duplicates, we obtained 1,435 genes in total in the endo-system, while the three compartmental gene-sets contained 899 (endosome), 678 (lysosome), and 236 (trans-Golgi network) genes, respectively (S16 Table). We confirmed their biological functions with a functional enrichment analysis using the Database for Annotation, Visualization, and Integrated Discovery (DAVID)[76], where the top enriched GO terms were indeed lysosome, endosome, and trans-Golgi network. (S13 Fig) A comparison of the endo-system gene-set to the findings in the recent AD GWASes [11,12] has been provided by checking the number of endocytic genes implicated in Jansen et al. and Kunkle et al. (S7 Fig). To note, some genes were related to multiple compartmental gene-sets and thus only one of the duplicated genes was included in the endo-system gene-set (S11 Fig).

Analysis of association between the burden of rare deleterious SNVs and AD status

To identify whether rare variants in the endocytic pathway are associated with AD, we compared the burden of rare deleterious SNVs between AD patients and controls. The burden was defined as the fraction of the alternative minor alleles that each individual carried for all rare deleterious SNVs, using the—score function in PLINK [62]. We additionally performed this procedure on the three compartmental gene-sets and obtained a burden score for each individual within each gene-set. To correct for potential confounding factors, for each gene-set, we first regressed the burden against the total number of rare SNVs and the top ten principal components (PCs). Due to randomness, the distribution of the number of rare SNVs might be naturally variable from sample to sample, in which case the distribution of rare deleterious SNVs would also be greatly affected. Similarly, the PCs helped to correct for potential population stratification within European ancestries. Both aspects could influence the burden score in ways unrelated to AD and thus need to be controlled. Once we had removed the confounding covariates, we performed three logistic regression models as proposed by Zhang et al.[77] using the residuals and AD status for all case-control studies. The three models differed in the covariates they corrected for. The minimal adjustment Model 0 (M0) controlled for the ten PCs and sequencing centers. This model has been previously reported to improve power for detecting variants whose effects are confounded with age and sex.[60] This phenomenon could be introduced by study design where the mean age between cases and controls are substantially disproportionate, as in the case of ADSP studies. Model 1 (M1) was built upon M0 by additionally including age and sex. Model 2 (M2) was further built upon M1 and included the count of APOE ε2 and ε4 alleles. For the ADSP family dataset, we generated kinship matrices and used a generalized linear mixed model (GLMM) to take kin relationships into consideration when calculating association p-values. In particular, we used the glmmkin function in the R package, GMMAT.[78] We computed odds ratio (OR) and p-values of association between the burden of rare deleterious SNVs and AD status in each model for European samples in each dataset (ADSP case-control study, AMP-AD case-control study, and ADSP family study). Our stage 1 analysis involved only the ADSP case-control dataset as the discovery set, while the AMP-AD case-control and the ADSP family study served as replication sets in our stage 2 analysis. We chose this analysis scheme because the ADSP case-control study encompassed the largest sample size, including non-European samples, even though we identified fewer samples with European ancestry compared to the AMP-AD case-control study. To note, the AMP-AD case-control study provided only the age-of-death for each individual, while the ADSP case-control and family studies provided only the age-at-onset. As a result, we used different definitions of age in analyzing different datasets. To validate our gene-set AD association analysis, we tested two additional methods provided by MAGMA [79] using the same set of rare deleterious variants. The first was the SNP-wise method applicable to both common and rare variants and the second was the burden method that MAGMA suggested to use for rare-variant-only analysis and was similar to the aforementioned gene-set AD association analysis using PLINK. We applied both methods to the set of rare deleterious variants previously defined and computed two types of p-values: a competitive p-value that tests whether the association within the gene-set is greater than in other genes and a self-contained p-value that tests whether there is an association within the gene-set of interest at all. The latter concept is the same as what our main analysis method aimed for. Due to our study design with multiple gene-sets and MAF thresholds, a Bonferroni correction was applied in accordance with the number of tests we performed in each analysis to define the study-wide significance threshold in each stage and each dataset. Although our analysis started with the whole endocytic pathway and then moved onto individual compartments, we, nonetheless, utilized a stringent multiple-testing correction threshold. Specifically, as we tested for four gene-sets (endo-system gene-set and three sub-compartmental gene-sets) and two MAF thresholds (1% and 0.1%), we set our significant threshold at α = 0.05/8 = 0.00625 for both stage 1 discovery phase and stage 2 replication phase analyses. Accordingly, we set our nominal significance threshold at α = 0.05.

To combine results from two stages (three studies) for each of the four gene-sets we tested previously, we performed meta-analyses on p-values using estimates from our best model, namely the model producing the smallest p-values among the three models tested. We used two meta-analysis methods to combine the results. The first was a fixed-effects inverse variance weighted method in METAL [80], which took ORs, standard deviations (SDs), and p-values for separate tests and combined them into one ‘Gene-set level’ p-value with an estimate of the unified effect. The second was Fisher’s method which only required p-values and has been shown to be more robust to some situations where a small portion of p-values are very small.[81,82] In particular, we used the sumlog function from the R package, ‘metap,’[83] which took into account the direction of effects in each study and the corresponding p-values. It then computed a ‘Gene-set level’ p-value similar to METAL indicating the significance of rare variants’ effect shared across studies but without an estimated effect size.

Analysis of association between the burden of rare deleterious SNVs and AD-related phenotypes

To test for association between the burden of rare deleterious SNVs and NFTs, we leveraged the Braak stages and followed a similar workflow as in testing AD status. As the sample size of patients with Braak staging information was limited in the ADSP family study, we tested for replication only in AMP-AD case-control study after analyzing the ADSP case-control study in stage 1. We obtained 626 and 1,399 individuals with Braak staging information in ADSP and AMP-AD case-control datasets, respectively. To note, even though the ADSP case-control study had fewer samples with Braak staging information, we, nonetheless, followed the same analysis scheme as in the previous AD analysis. In practice, after removing confounding effects from the burden score, we applied three ordinal logistic regression (OLR) models (M0, M1, M2) to account for multiple ordered categories present in the Braak staging (stage 0 to VI). The regular logistic regression only allows binary dependent variables, which is not feasible for Braak stages. In particular, we used the polr function from the R package, MASS [84], which fits a logistic regression model to an ordered factor response. Similar to the previous burden analysis, our M0 accounted for sequencing centers and the top 10 PCs; our M1 additionally controlled for sex and age; finally, our M2 further included the count of APOE ε2 and ε4 alleles. For analyses in all datasets, our significance threshold after the multiple-testing correction was still at α = 0.00625 because we tested for two MAF thresholds and four gene-sets. Finally, the nominal significance threshold was also at α = 0.05. To increase statistical strength and precision in estimating effects [85], we again performed meta-analyses and combined these two independent tests similar to what we did for AD association analyses.

We additionally tested the age-specific risk of rare deleterious SNVs in the endocytic pathway. As aforementioned, the AAO and AOD information was provided by the ADSP studies and the AMP-AD study, respectively, which allowed us to test for two different age-specific risks within each gene-set. Different from AD risk, age-specific risk leveraged the information of age and estimated the association between the age-to-event (survival time) of patients and the rare-variant burden score. Therefore, we adopted a genetic epidemiological framework proposed by Desiken et al.[86], in which a Cox Proportional Hazard Regression (CPHR) was performed to account for age-to-event information. Specifically, we first used the Surv function from the R package, “survival”[87], and computed a survival time for each sample in each dataset. Then, we conducted CPHR using the coxph function from the R package, ‘survminer’[88], to estimate the hazard ratio, or the ratio of risk-to-event (onset or death), depending on the input age we used. We performed three CPHR models (M0, M1, and M2) similar to the previous burden analysis on AD status and Braak staging, except that age was not a covariate in either of the three models. Therefore, since we tested for two different MAF thresholds and four gene-sets (though in a stepwise fashion), we set a stringent significant threshold at α = 0.05/8 = 0.00625 and our nominally significant threshold at α = 0.05 for analyses in all three datasets. Finally, we combined the results of AAO in the same way as we did for AD and NFT association tests. The resulting p-value then indicated the shared rare-variant effect on AAO-specific risk across the ADSP case-control and family studies.

Single-gene analysis

To identify specific genes within the endocytic pathway associated with AD, we extracted rare deleterious SNVs as defined previously for each gene in the endo-system gene-set that were present in European samples for the ADSP case-control, the AMP-AD case-control, and the ADSP family study. Association test was performed for AD status by first building a null model using the SKAT_Null_Model function in the R package, SKAT, [89] followed by running the SKATBinary function using the SKAT-O feature to obtain association p-values for binary traits. We used a full model that included age, sex, sequencing center, the number of APOE ε2 and ε4 alleles, and top 10 PCs. To note, we also applied SKAT_Null_Model to the ADSP family dataset without incorporating kinship structure. This procedure could only be valid in the case where the family structure was relatively simple and did not contribute to a large effect in our analysis. By re-running the previous AD burden analysis with and without kinship information, we indeed observed only small deviations between these two tests. Specifically, for the full model of the endo-system gene-set, we observed an OR of 1.34 with kinship structure provided (p = 0.035) while we observed a similar OR of 1.36 assuming an independent setup (p = 0.02), which indicated that the family structure within the ADSP family study did not influence our analyses to a large extent.

To test for association with Braak stages, we first extracted only European samples with Braak staging information available for each dataset, before extracting rare deleterious SNVs for each gene within the endo-system gene-set. We leveraged the fact that it is a semi-quantitative trait and performed the association test with the SKAT function for continuous traits with the ‘optimal’ option after building null models as described for testing AD status. In the attempt to remove confounding factors and unbalanced sample distribution for Braak staging association test, we additionally included AD status in null models. Finally, we meta-analyzed variants across datasets and computed ‘Gene-level’ p-values for AD status as well as Braak staging. We combined genotyping matrices across three datasets for each gene using the R package, MetaSKAT.[90] Specifically, we first transformed our genotyping matrices into an SSD format for a single population and then analyzed all three populations at once using the function MetaSKAT_MSSD_ALL. This procedure increased the power to analyze the effects of rare variants that are shared across different studies. To correct for testing multiple genes within the endo-system gene-set, we obtained the number of genes we tested in each separate dataset and computed their corresponding Bonferroni corrected significance thresholds. Specifically, for the AD single-gene analysis, we tested 1,195, 1,228, and 683 genes in ADSP case-control, AMP-AD case-control, and ADSP family datasets, respectively, which corresponded to Bonferroni corrected significance thresholds of α = 4.18*10⁻⁵; 4.07*10⁻⁵; 7.32*10⁻⁵, respectively. In meta-analyses, we identified 642 genes in common and computed a Bonferroni corrected significance threshold of α = 7.79*10⁻⁵. For the Braak staging single-gene analysis, we retained only rare deleterious SNVs present in samples with Braak staging information available and tested for 1,035 and 1,176 genes for the ADSP and AMP-AD case-control studies, respectively. The corresponding Bonferroni corrected significance thresholds were then computed as α = 4.83*10⁻⁵ for the ADSP case-control dataset and 4.25*10⁻⁵ for the AMP-AD case-control dataset. When performing meta-analyses, we examined 967 genes in common between these two datasets, which led to a Bonferroni corrected significance threshold of α = 5.17*10⁻⁵.

Functional analysis on AD

One approach to understanding how the effect of rare variants would influence the risk of AD status is to investigate how they regulate gene expression. A gene with a variation that is associated with its gene expression is called an eGene. Here, we obtained the bulk RNA-Seq data of DLPFC tissues of 636 individuals from the ROSMAP [55] study and performed an association test between the expression of a gene and rare variants in cis with the corresponding gene. In particular, for each gene within the endo-system gene-set, we included all variants within gene boundary and additionally all rare variants within 20kb up- and down-stream of the transcription start sites (TSS), which might potentially regulate the expression of a gene through cis-regulation, such as the effect of enhancer region. To overcome the problem of low power to detect the effect of single rare variants, we aggregated the effects of all rare variants within as well as near the TSS of each gene. We analyzed this aggregated effect on gene expression using the SKAT function to compute ‘Gene level’ p-values, while taking into account confounding covariates, including age, sex, sequencing locations, APOE ε2 and ε4 alleles, and top 10 PCs. To correct for testing multiple genes, we calculated false discovery rate for all tested genes and used FDR of 0.05 as the q-value threshold, following the suggestions of previous studies.[91,92] Follow-up validation was performed using genes previous identified from the burden and functional analyses, by directly comparing their expression levels between AD cases and controls using student t-test and computing the Pearson correlation between their expression levels and Braak stages. The multiple-testing issue was then addressed using the Bonferroni correction method.

The resolution of bulk RNA-Seq data may limit our capability of observing cell-type specific effects on AD.[55,61,93,94] To elucidate the underlying complexity of variation across cell types, we further obtained single-cell RNA-Seq (scRNA-Seq) of 48 samples (24 AD patients and 24 cognitively normal controls) from the ROSMAP study and investigated the pattern of expression for each of the six major cell types defined on a priori cell-type-specific gene-sets: excitatory neuron (Ex), inhibitory neuron (In), astrocyte (Ast), oligodendrocyte (Oli), oligodendrocyte-precursor-cell (Opc), and microglia (Mic)[61]. The six major cell types were further divided into sub-clustered cells based on the heterogeneity of gene expression within each cell type: 13 Exs, 12 Ins, 4 Asts, 5 Olis, 3 Opcs, and 4 Mics [61]. The whole dataset in 10X format was first processed using the R package, Seurat.[95] We followed the preprocessing steps as proposed by the Seurat developer by first filtering out cells with reads quantified for less than 200 or more than 2,500 genes, followed by filtering out cells with the percentage of mitochondrial gene counts over 5 percent. We then employed a global-scaling normalization method provided by the LogNormalize function, which normalized the feature expression measurements for each cell by the total expression, followed by a log-transformation. The major and sub-cell types were identified a priori for this scRNA-Seq data. Therefore, we extracted all significant genes identified in the previous single-gene and functional analyses for each specific cell type and conducted differential gene expression analysis using the student t-test method between cases and controls for each major cell type, as well as for each subcellular population within each major cell type.

The burden of rare deleterious SNVs in endo-system gene-set for ADSP case-control study

To investigate whether rare deleterious SNVs in the endocytic pathway were associated with AD, we leveraged a gene-set method of burden analysis that collapsed individual effects of multiple variants into one ‘gene-set level’ effect, hence increasing the power of detecting rare variants’ effect. We defined rare SNVs using both an external source of allele frequency and allele frequency observed in 1,291 European samples (664 AD cases and 627 controls) from the ADSP case-control study (see Methods). We focused on deleterious SNVs as defined in Methods, in which most were protein-altering variants. We identified rare deleterious SNVs in 1,133 of the 1,435 genes in our gene-set (see Methods). For each individual, we computed the burden of these rare deleterious SNVs. We then compared the genetic burden between AD cases and cognitively normal controls, while taking into account confounding covariates that can potentially influence the amount of burden. Such covariates include ancestral principal components, age, sex, the sequencing location, the number of APOE e2 and e4 alleles, and the total number of rare SNVs of each individual. The last procedure is necessary to account for individual differences in the total amount of variation; an individual is likely to carry more rare deleterious SNVs if she/he carries more rare SNVs overall. To note, we found that the total number of rare SNVs on the genome-wide scale has no statistically significant difference between cases and controls (p = 0.67, student t-test). As described in Methods, we applied three logistic regression models to find associations between AD status and the burden scores while the three models were built on top of each other and tested for two MAF thresholds (1% and 0.1%). Looking at our best model in terms of the strongest association, we observed that the risk of AD, as indicated by the odds ratio (OR), increased by 1.24 for every one unit increase in residual burden score (p = 0.00018 using GLM), which was a significant association after stringent multiple testing correction (α = 0.00625) for all gene-sets (including sub-gene-sets we analyzed in next steps) (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Rare deleterious variants are enriched in AD patients across the endocytic and corresponding compartmental gene-sets in stages 1 and 2.

We compared the burden of rare deleterious variants between AD patients and controls across the endo-system (endo-sys) gene-set and three compartmental sub-gene-sets (endosome, lysosome, and trans-Golgi network) in stage 1 ADSP case-control dataset (leftmost), which were then tested for replication in stage 2 AMP-AD case-control (middle) and ADSP family (rightmost) datasets. Enrichment (ORs) and p-value were computed using a linear regression model controlling for covariates, including the total count of rare variants (see Methods). P-values of enrichment in each gene-set are indicated above horizontal bars which represent 95% confidence intervals.

https://doi.org/10.1371/journal.pgen.1009772.g001

Additionally, we identified three major cellular compartments participating in the endocytic pathway and their corresponding genes, which constituted subsets of the endo-system gene-set. The first two compartmental gene-sets were endosome (n = 811) and lysosome (n = 620) gene-sets, which served as the major sorting station in the endocytic pathway and the final destination of proteolytic destinations [96], respectively. The third important compartment was the trans-Golgi network gene-set (n = 208) which represented a pathway sorting station for retrograde trafficking. In summary, we identified 689 endosomal genes, 544 lysosomal genes, and 181 trans-Golgi network genes, respectively. We found the burden scores of rare deleterious SNVs were higher in cases than in controls for all three sub-gene-set. In our best model, the OR, representing the risk of AD, increased by 1.18 per unit for the endosome gene-set (p = 0.0056 using GLM), 1.08 per unit for the lysosome gene-set (p = 0.09 using GLM; Fig 1), and 1.14 per unit for the trans-Golgi network gene-set (p = 0.019 using GLM). After the multiple-testing correction, we observed the endosome gene-set showed a gene-set-wide significant association with AD while the trans-Golgi network displayed a nominally significant association signal. In addition to exploring sub-gene-sets, we also checked the specificity of the association in the endocytic pathway by obtaining gene-sets unrelated to AD. Specifically, we explored two non-disease complex traits, BMI and height, and obtained related genes (212 and 78, respectively) from GeneRIF, a publically available database for functional annotations.[97] Indeed, we did not observe an enriched rare-variant burden in AD cases compared to controls in these gene-sets and the directions of effects were different across datasets, suggesting the observed rare-variant effect was specific to the endocytic pathway (S13 Table).

Stage 2 replication of the burden analysis in two independent WGS datasets

The gene-set burden analysis in the ADSP case-control study demonstrated statistically significant enrichment of rare deleterious SNVs in cases in the endocytic pathway, indicating an increase of risk conferring AD. We further examined the endo-system gene-set in 1,611 European samples (642 AD cases and 969 controls) from the AMP-AD study. We obtained 1,198 endo-system-related genes and observed an elevated risk of AD in terms of OR of 1.19 (p = 0.0038 using GLM; Fig 1), replicating the observation of a significantly higher burden of rare deleterious SNVs in the stage 1 analysis, using a multiple testing threshold of α = 0.00625.

We performed additional gene-set burden analysis on the sub-gene-sets of the functional compartments in the AMP-AD study. We identified 735 endosomal genes, 576 lysosomal genes, and 187 trans-Golgi network genes, respectively. We again observed an increase in AD risk among cases for all three sub-gene-sets. A nearly significant signal was observed in the lysosome gene-set with an OR of 1.17 (p = 0.0063 using GLM; Fig 1). For the other two gene-sets, we observed an OR of 1.08 (p = 0.16 using GLM; endosome gene-set) and 1.10 (p = 0.083 using GLM; trans-Golgi network). None of these gene-sets showed gene-set-wide significant association after multiple testing correction at α = 0.00625, although the lysosome gene-set nearly reached the gene-set-wide significance threshold.

As described above, the AMP-AD study consisted of three sub-cohorts and the largest one, ROSMAP, contained around 71.5% of the total sample size. To avoid potential batch effect diluting the association signal, we re-performed the analysis on only the ROSMAP data. In fact, we observed slightly more significant results in nearly all gene-sets, where the endo-system and the lysosome gene-sets both reached gene-set-wide significance threshold. (S2 Table) Overall, the associations were similar between the AMP-AD and ROSMAP data, indicating a relatively low level of batch effect among the three sub-cohorts.

Given the observed risk in stage 1 ADSP case-control study and the stage 2 AMP-AD replication, we further examined the genetic burden in the ADSP Family study. We filtered and annotated rare deleterious SNVs based on the same workflow using 353 European samples (144 AD cases and 209 controls) of the ADSP family study. Due to the smaller sample size compared to the previous two case-control studies, we obtained 683 endo-system-related genes. To examine the AD risk, we performed GLMM using the burden of each individual. Due to family structure, we utilized the generalized linear mixed model to account for the relatedness between samples. We observed an OR of 1.42 (p = 0.013 using GLMM), conferring an elevated AD risk among cases compared to controls. (Fig 1) This observation was not gene-set-wide significant using the Bonferroni correction threshold at α = 0.00625. However, it displayed a nominally significant association with the same direction of effect as in the ADSP and AMP-AD case-control studies.

Nonetheless, we looked into the sub-gene-sets of the three functional compartments in the ADSP family dataset. We identified 402 endosomal genes, 342 lysosomal genes, and 106 trans-Golgi network genes, respectively. We observed a significant elevation of AD risk among cases for endosome gene-set with an OR of 1.48 (p = 0.0045 using GLM). Similar increases were also observed in the lysosome and trans-Golgi network gene-sets, with OR of 1.18 (p = 0.22 using GLMM) and 1.04 (p = 0.77 using GLMM), respectively (Fig 1). Only the endosome gene-set remained gene-set-wide significant after multiple-testing correction, which was in concordance with our observation in the stage 1 ADSP case-control study.

A meta-analysis of stage 1 and 2 burden analysis

The stage 1 burden analysis using the ADSP case-control study demonstrated a significant increase in AD risk in the endo-system gene-set, which was replicated in one independent dataset, the AMP-AD case-control dataset, and displayed a nominal significance in the ADSP family study. We meta-analyzed the results using two different methods (see Methods) and computed a ‘Gene-set level’ p-value of 2.17*10⁻⁷ (by METAL; Fisher’s method produced similar results; Table 2) for the endo-system gene-set, which was improved compared to stages 1 and 2. The same was also observed for sub-gene-sets where we computed a meta-analysis p-value of 9.78*10⁻⁵ for the endosome gene-set, 9.83*10⁻⁴ for the lysosome gene-set, and 1.19*10⁻² for the trans-Golgi network gene-set. Except for the trans-Golgi network gene-set that has the smallest number of genes, all other gene-sets remained gene-set-wide significant after multiple-testing correction (α = 0.00625), which strongly demonstrated a shared effect of rare deleterious variants within the endocytic pathway across multiple independent studies. To note, although we meta-analyzed the results from the best models, as proposed by Zhang et al. to improve the power of detection, the same pattern of rare-variant association could be observed using the same models for each gene-set across the three datasets. (S1 Table) For all models in the endo-system, endosome, and lysosome gene-sets except M2 of lysosome, we observed gene-set-wide significant p-values, regardless of the meta-analysis methods used, demonstrating a high consistency with the observations made using the best models.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Meta-analysis of stages 1 and 2 gene-set burden analyses using AD, NFT, and AAO.

https://doi.org/10.1371/journal.pgen.1009772.t002

A similar pattern of meta-analysis results was also observed in the additional validation tests from two aspects. Firstly, we wanted to check our results using different annotation tools. Given the set of deleterious variants used in previous phase 1 analyses, we additionally filtered by CADD scores (see Methods) and re-ran the gene-set AD association analyses with the resulting set of pathogenic/deleterious variants. In the meta-analysis, we observed that the endocytic, endosome, and lysosome gene-sets reached gene-set-wide significance threshold (see S5 Table), consistent with the rare-variant effect we observed in the endocytic pathway using the original set of rare deleterious variants.

The second aimed to validate our gene-set burden analysis using MAGMA with two different aggregation methods (see Methods). In the meta-analysis, both the SNP-wise and burden methods provided gene-set-wide significant self-contained p-values for nearly all gene-sets (S3 and S4 Tables; for endo-system, SNP-wise: 9.28*10⁻⁷; burden: 5.16*10⁻⁸), similar to the results shown above in Table 2. Compared to the MAGMA burden method, the SNP-wise method was not designed for rare-variant-only analysis and indeed showed weaker association signals. Especially for the competitive p-values, we observed gene-set-wide significant results for nearly all gene-sets using the MAGMA burden method, but not the SNP-wise method (for endo-system, SNP-wise: 2.41*10⁻²; burden: 1.90*10⁻³). We also attempted to compute a weighted burden score using pLI scores by PLINK and observed gene-set-wide significant associations in the endo-system gene-set in the meta-analysis. (S10 Fig, S12 Table) Compared to our main method above, the MAGMA methods and the weighted method displayed some fluctuations in individual datasets and models but consistent results in meta-analysis, indicating a robust rare-variant effect in the endocytic pathway under different statistical methods. Besides, as APOE was a major risk determinant in AD, in this validation, we also checked whether our observed rare-variant enrichment was mainly contributed from this gene, rather than the whole endo-system gene-set, by re-run the analysis with APOE excluded. Indeed, we observed nearly the same p-values in the meta-analysis, indicating a rare-variant enrichment in AD cases even without APOE.

The burden of rare deleterious SNVs on NFTs

NFT, measured in Braak staging, was one of the most important histopathological indicators of AD [48–50]. It is designed as an ordinal scale from 0 to VI of NFT pathology where AD patients with high Braak stages (V or VI) are diagnosed with high confidence.[98] Therefore, Braak stages may serve as a finer spectrum or proxy of AD severity and provide higher power in assessing the effect of rare variants in AD progression. Based on our previous AD analysis, we hypothesized that the burden of rare deleterious variants in the endocytic pathway would be higher in patients with later Braak stages. To test our hypothesis, we applied an ordinal logistic regression (OLR) method to Braak stages (see Methods). This method has been previously shown to be effective in studies of Braak staging as well as of other ordered phenotypes, such as oral cancers.[99,100] We obtained 626 individuals (475 AD cases and 151 cognitively normal controls) from the stage 1 ADSP case-control dataset and 1,399 individuals (533 AD patients and 866 controls) from AMP-AD case-control dataset with Braak staging information, which were used to fitted OLR models. In stage 1, We observed an OR of 1.16 (p = 0.039 using OLR; S1 Fig) in the endocytic pathway, implicating a nominally significant association of rare-variant enrichment to later Braak stages. However, this result did not replicate in stage 2 with sufficient significance (OR = 1.08, p = 0.13 using OLR; S1 Fig). Comparing the stages 1 and 2 samples, we observed a distinct distribution of Braak stages. In particular, the stage 2 samples were concentrated in Braak stage III (23.1%), IV (28.1%), and V (23.3%), whereas most stage 1 samples were clustered in stage V (26.4%) and VI (34.8). (S3 Fig) We did not test for replication in the ADSP Family study due to limited samples with Braak staging information (n = 38 individuals where only one sample had AD).

Our analyses of two independent datasets suggested a trend of increased risk of bearing later Braak stages with elevated rare-variant burden in the endocytic pathway. To improve power, we meta-analyzed the results from the ADSP and AMP-AD case-control studies, producing a ‘Gene-set level’ p-value between 0.0099 and 0.012, which did not pass our multiple-testing correction threshold of α = 0.00625. (Table 2) Further Braak staging burden analysis using compartmental sub-gene-sets, however, revealed a gene-set-wide significant signal in the meta-analysis for lysosome gene-set (p = 0.0066, Fisher’s method). A full list of results for NFT burden analysis can be found in S6 Table.

Hazard analysis on population risk of AD age of onset and death

Previous gene-set burden analyses have demonstrated a significant correlation between the burden of rare deleterious variants within the endocytic gene-set and AD risk. One important aspect of AD development is its age-specific phenotypes, such as AAO. Previous studies on AD have shown a large genetic component in the heritability of AAO [101,102], with multiple risk loci associated with it. [103–107] It is thus of interest to also examine the genetic risk identified within the endocytic gene-set in this context. One approach is to evaluate whether AD patients with earlier AAO are associated with greater rare-variant burden within the endocytic gene-set. Previous studies have proposed a genetic epidemiological framework, where age-specific phenotypes were analyzed using a Cox Proportional Hazard Regression (CPHR) that considered a time-to-event probability, as opposed to the simple event probability estimated in logistic regression.[86,108] Therefore, we leveraged our previously computed burden score for each individual in the ADSP case-control study and constructed a cox proportional hazard (CPHR) model to estimate the instantaneous risk of developing AD, in consideration of genotype and AAO. A positive estimate of hazard in this model would indicate a higher risk of developing AD in early ages. We built three models as in the burden analysis and observed in our best model that an AAO-specific genetic risk increased by 1.14 per unit increase in the residual burden score (p = 0.00083 using CPHR; Fig 2), which reached gene-set-wide significance after multiple testing correction (α = 0.00625). We further examined the AAO-specific genetic risk within the functional sub-gene-sets. In our best model, we observed a gene-set-wide significant hazard ratio of 1.14 (p = 0.00097 using CPHR) for lysosome gene-set and a nominal significant hazard ratio of 1.10 (p = 0.011 using CPHR) for trans-Golgi network gene-set.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. The enrichment of rare deleterious variants is associated with AD AAO across the endocytic and corresponding compartmental gene-sets in stages 1 and 2.

We computed a hazard ratio of obtaining AD in earlier ages using the burden of rare deleterious variants across the endo-system gene-set and three compartmental sub-gene-sets (endosome, lysosome, and trans-Golgi network) in stage 1 ADSP case-control dataset (left), which were then tested for replication in stage 2 ADSP family datasets (right). Enrichment (ORs) and p-value were computed using CPHR. P-values of enrichment in each gene-set are indicated above horizontal bars which represent 95% confidence intervals.

https://doi.org/10.1371/journal.pgen.1009772.g002

To test for replication, we examined the ADSP family study under the same statistical framework. Applying the CPHR models, we observed a gene-set-wide significant hazard ratio of 1.31 (p = 0.00091 using CPHR; Fig 2) in the endo-system gene-set. Carefully examining the sub-gene-sets also revealed gene-set-wide significant AAO-specific risk within the endosome gene-set (HR = 1.35, p = 3.83*10⁻⁵ using CPHR). We did not observe significant associations using the other two compartmental gene-sets (S7 Table).

To increase power, we performed meta-analyses to identify rare-variant effects shared across multiple studies. We combined the best results from ADSP case-control and family studies and observed a gene-set-wide significant p-value of 2.47*10⁻⁶ (by METAL; Fisher’s method produced similar results; Table 2) for the endo-system gene-set, which was greatly improved compared to results in either stage. Similarly, the endosome gene-set also demonstrated an improved gene-set-wide significant p-value of 3.33*10⁻⁵. However, the lysosome and the trans-Golgi network gene-sets showed only nominally significant p-values in our meta-analysis, potentially due to the absence of signal in the ADSP family study. These findings strongly demonstrated that this AAO-specific rare-variant effect in the endocytic pathway was shared in European samples across different studies.

Another age-specific phenotype is the age of death (AOD), which has been shown to be affected by genetic groups implicated in AD AAO as well as in other dementia.[109,110] We thus followed the same analysis framework using the CPHR model and assessed whether AD-affected patients with earlier AOD were associated with a higher rare-variant burden in the endocytic pathway. We looked at European samples in the AMP-AD case-control study, where the AOD information was available. We observed a hazard ratio of 1.10 (p = 0.024 using CPHR; S2 Fig), indicating an increase of risk of death in AD patients as well as a worse prognosis along with an elevation in genetic burden. Further analysis using the lysosome sub-gene-set displayed a hazard ratio of 1.09 (p = 0.036 using CPHR). Both endo-system and lysosome gene-sets demonstrated nominally significant associations with AOD but did not reach gene-set-wide significance after multiple-testing correction. Analysis using other sub-gene-sets did not provide significant hazard ratios.

Single-gene analysis on AD risk using AD and NFT status

From the previous analysis, the endo-system gene-set conferred a large rare-variant effect on AD and related phenotypes. Thus, we decided to examine the effect of rare variants in single endocytic genes, attempting to identify those associated with AD with large effect sizes. To increase power, we aggregated previously defined rare deleterious SNVs in each gene and tested for association with AD. We did not observe a single gene passing the Bonferroni corrected significance threshold in all three datasets, as well as in meta-analysis (See Methods; α = 4.18*10⁻⁵; 4.07*10⁻⁵; 7.32*10⁻⁵; 7.79*19⁻⁵, for ADSP case-control, AMP-AD case-control, ADSP family studies, and meta-analysis respectively; S15 Table).

As mentioned previously, NFT status may provide more detailed information of the pathological progression of AD and thus a greater power to detect signals of rare-variant effect. We, therefore, performed single-gene analysis using NFT status, as a proxy for AD status. For all datasets, we retained only rare deleterious SNVs that were present in samples with Braak staging information. We controlled for the same set of covariates as in previous analyses, except that we also included the AD phenotype (AD affected/unaffected) for each individual as one additional covariate (see Methods). The latter is necessary because the Braak staging and the AD phenotype are correlated, and the numbers of individuals with and without AD were vastly disproportionate among the samples with Braak staging information. For the ADSP case-control study, we observed six genes that reached Bonferroni corrected significance threshold (α = 4.83*10⁻⁵). None of the genes passed the Bonferroni corrected significance threshold (α = 4.25*10⁻⁵) in the AMP-AD study. Results of the top ten most significant genes can be found in S8 Table. We conducted meta-analyses for these two independent studies using MetaSKAT as before. In the combined results, we observed one gene, ANKRD13D, reached Bonferroni corrected significance threshold (p = 3.56e-05; α = 5.17*10⁻⁵). This gene has been previously implicated in AD through RNA expression analysis [111] and protein interactome mapping [112].

The identification of functional effects of rare variants within the endocytic pathway

The hypothesis that the endo-system gene-set contains rare variants that are influential to AD development is endorsed by the previous gene-set burden analyses and single-gene analyses. One approach to investigating how the effect of rare variants takes place is to analyze how these rare variants are associated with gene expression. Such gene containing variations affecting its expression is often called an eGene.[91] To identify eGenes, we obtained bulk RNA-Seq data of DLPFC brain tissues of 636 individuals from the ROSMAP study as part of the AMP-AD study and tested for association of all variants in cis with a gene with its gene expression. Specifically, we grouped all variants within one gene, as well as those near the corresponding TSS, and assessed whether the aggregated rare-variant effect in an endocytic gene is associated with its expression level using SKAT (see Methods). Intersecting the bulk RNA-Seq and WGS data revealed 547 individuals with 224 AD patients and 323 controls. By taking an FDR of 5%, we discovered two genes, HLA-A and SLC26A7, whose rare variants were significantly associated with expressional changes. To note, previous studies have demonstrated that proteins from the same families of these two genes are associated with AD status. Specifically, two proteins from the HLA families and one from the SLC families have been implicated in AD through meta-analyses of large GWAS and brain DNA-methylation association analysis.[9,113] We first examined their single-gene analysis results and observed that none of them was significant using the AMP-AD dataset (p = 4.74e-01; 2.14e-01, for HLA-A and SLC26A7, respectively). To validate our results and determine the direction of effects, we compared the expression of these two genes between cases and controls. Indeed, their expression levels were both significantly decreased in cases compared to controls (p = 0.00073 HLA-A; Fig 3A; p = 0.0054 SLC26A7; Fig 3B; student t-test; α = 0.017;). We further examined the distribution of their expression levels across multiple Braak stages. Similarly, both expressions were strongly negatively correlated with greater Braak stages (r = -0.129, p = 0.0024 HLA-A; r = -0.127, p = 0.0029 SLC26A7; Pearson correlation; α = 0.017).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Comparison of the gene expression of HLA-A, SLC26A7, and ANKRD13D between AD cases and controls from the ROSMAP study.

Violin plots were used to represent the distribution of gene expression within each AD status, where a symmetric deviation from the middle line on both sides indicated a higher abundance of samples at the corresponding gene expression level. Comparisons between AD cases and controls were assessed using boxplots. P-values were computed using the student t-test. All three genes, HLA-A, SLC26A7, and ANKRD13D, are down-regulated in AD cases compared to controls.

https://doi.org/10.1371/journal.pgen.1009772.g003

We also investigated ANKRD13D, which we previously identified to be associated with Braak stages, in the context of gene expression. Although not an eGene, ANKRD13D exhibited a significant expressional decrease in cases compared to controls (p = 0.0026; student t-test; Fig 3C). The analysis on Braak staging also revealed a strong negative correlation (r = -0.122, p = 0.0042; Pearson correlation).

scRNA expression analysis

Recent advancement in analyzing gene expression in single-cell resolution has provided opportunities to uncover complex alterations across cell types and identify cell-type specific effects on AD.[55,61,93,94] For example, previous studies have pointed the imbalance of excitatory and inhibitory neurons could lead to overexcitability and early dysregulation in the development of AD [114]. Many other studies also demonstrated abnormalities in innate immune cells, primarily microglia, in the pathogenesis of AD.[115] Therefore, to investigate the potential cell-type specific effects of rare variants within the endocytic pathway, we obtained the single-cell RNA-seq data of 48 samples (24 AD patients and 24 cognitively normal controls) from the ROSMAP study. We focused on three genes we identified through the previous analysis, which demonstrated significant associations to AD progression. The scRNA-Seq data were labeled with six major cell types using a priori marker genes (Ex, In, Ast, Oli, Mic, and Opc), and sub-clustering within each cell type revealed cellular subpopulation (see Methods). We examined the expression of the three target genes in all major cell types and observed that ANKRD13D was up-regulated in Ex (p = 1.92*10–18; student t-test), Ast (p = 0.011; student t-test), and In (p = 0.028; student t-test) (S9 Table). However, it exhibited a down-regulation in Oli (p = 0.0018; student t-test). SLC26A7 was observed to be up-regulated in Ex (p = 0.0049; student t-test), while HLA-A displayed a pattern of down-regulation in both In and Mic (p = 9.72*10-6and p = 0.0031 respectively; student t-test). Four AD pathology-associated cellular subpopulations (Ex4, In0, Ast1, and Oli0) have been previous demonstrated for this scRNA-Seq data [61,91]. Our differential expression analysis within these four subpopulations showed a pattern of up-regulation of ANKRD13D in Ex4 and In0 (p = 5.76*10–8 and p = 0.036, respectively; student t-test; S10 Table). The other two genes, however, were not significantly differentially expressed in these four cell subpopulations.

Acknowledgments to ADSP

The Alzheimer’s Disease Sequencing Project (ADSP) is comprised of two Alzheimer’s Disease (AD) genetics consortia and three National Human Genome Research Institute (NHGRI) funded Large Scale Sequencing and Analysis Centers (LSAC). The two AD genetics consortia are the Alzheimer’s Disease Genetics Consortium (ADGC) funded by NIA (U01 AG032984), and the Cohorts for Heart and Aging Research in Genomic Epidemiology (CHARGE) funded by NIA (R01 AG033193), the National Heart, Lung, and Blood Institute (NHLBI), other National Institute of Health (NIH) institutes and other foreign governmental and non-governmental organizations. The Discovery Phase analysis of sequence data is supported through UF1AG047133 (to Drs. Schellenberg, Farrer, Pericak-Vance, Mayeux, and Haines); U01AG049505 to Dr. Seshadri; U01AG049506 to Dr. Boerwinkle; U01AG049507 to Dr. Wijsman; and U01AG049508 to Dr. Goate and the Discovery Extension Phase analysis is supported through U01AG052411 to Dr. Goate, U01AG052410 to Dr. Pericak-Vance and U01 AG052409 to Drs. Seshadri and Fornage. Data generation and harmonization in the Follow-up Phases is supported by U54AG052427 (to Drs. Schellenberg and Wang).

The ADGC cohorts include: Adult Changes in Thought (ACT), the Alzheimer’s Disease Centers (ADC), the Chicago Health and Aging Project (CHAP), the Memory and Aging Project (MAP), Mayo Clinic (MAYO), Mayo Parkinson’s Disease controls, University of Miami, the Multi-Institutional Research in Alzheimer’s Genetic Epidemiology Study (MIRAGE), the National Cell Repository for Alzheimer’s Disease (NCRAD), the National Institute on Aging Late Onset Alzheimer’s Disease Family Study (NIA-LOAD), the Religious Orders Study (ROS), the Texas Alzheimer’s Research and Care Consortium (TARC), Vanderbilt University/Case Western Reserve University (VAN/CWRU), the Washington Heights-Inwood Columbia Aging Project (WHICAP) and the Washington University Sequencing Project (WUSP), the Columbia University Hispanic- Estudio Familiar de Influencia Genetica de Alzheimer (EFIGA), the University of Toronto (UT), and Genetic Differences (GD).

The CHARGE cohorts are supported in part by National Heart, Lung, and Blood Institute (NHLBI) infrastructure grant HL105756 (Psaty), RC2HL102419 (Boerwinkle) and the neurology working group is supported by the National Institute on Aging (NIA) R01 grant AG033193. The CHARGE cohorts participating in the ADSP include the following: Austrian Stroke Prevention Study (ASPS), ASPS-Family study, and the Prospective Dementia Registry-Austria (ASPS/PRODEM-Aus), the Atherosclerosis Risk in Communities (ARIC) Study, the Cardiovascular Health Study (CHS), the Erasmus Rucphen Family Study (ERF), the Framingham Heart Study (FHS), and the Rotterdam Study (RS). ASPS is funded by the Austrian Science Fond (FWF) grant number P20545-P05 and P13180 and the Medical University of Graz. The ASPS-Fam is funded by the Austrian Science Fund (FWF) project I904), the EU Joint Programme—Neurodegenerative Disease Research (JPND) in frame of the BRIDGET project (Austria, Ministry of Science) and the Medical University of Graz and the Steiermärkische Krankenanstalten Gesellschaft. PRODEM-Austria is supported by the Austrian Research Promotion agency (FFG) (Project No. 827462) and by the Austrian National Bank (Anniversary Fund, project 15435. ARIC research is carried out as a collaborative study supported by NHLBI contracts (HHSN268201100005C, HHSN268201100006C, HHSN268201100007C, HHSN268201100008C, HHSN268201100009C, HHSN268201100010C, HHSN268201100011C, and HHSN268201100012C). Neurocognitive data in ARIC is collected by U01 2U01HL096812, 2U01HL096814, 2U01HL096899, 2U01HL096902, 2U01HL096917 from the NIH (NHLBI, NINDS, NIA and NIDCD), and with previous brain MRI examinations funded by R01-HL70825 from the NHLBI. CHS research was supported by contracts HHSN268201200036C, HHSN268200800007C, N01HC55222, N01HC85079, N01HC85080, N01HC85081, N01HC85082, N01HC85083, N01HC85086, and grants U01HL080295 and U01HL130114 from the NHLBI with additional contribution from the National Institute of Neurological Disorders and Stroke (NINDS). Additional support was provided by R01AG023629, R01AG15928, and R01AG20098 from the NIA. FHS research is supported by NHLBI contracts N01-HC-25195 and HHSN268201500001I. This study was also supported by additional grants from the NIA (R01s AG054076, AG049607 and AG033040 and NINDS (R01 NS017950). The ERF study as a part of EUROSPAN (European Special Populations Research Network) was supported by European Commission FP6 STRP grant number 018947 (LSHG-CT-2006-01947) and also received funding from the European Community’s Seventh Framework Programme (FP7/2007-2013)/grant agreement HEALTH-F4-2007-201413 by the European Commission under the programme "Quality of Life and Management of the Living Resources" of 5th Framework Programme (no. QLG2-CT-2002-01254). High-throughput analysis of the ERF data was supported by a joint grant from the Netherlands Organization for Scientific Research and the Russian Foundation for Basic Research (NWO-RFBR 047.017.043). The Rotterdam Study is funded by Erasmus Medical Center and Erasmus University, Rotterdam, the Netherlands Organization for Health Research and Development (ZonMw), the Research Institute for Diseases in the Elderly (RIDE), the Ministry of Education, Culture and Science, the Ministry for Health, Welfare and Sports, the European Commission (DG XII), and the municipality of Rotterdam. Genetic data sets are also supported by the Netherlands Organization of Scientific Research NWO Investments (175.010.2005.011, 911-03-012), the Genetic Laboratory of the Department of Internal Medicine, Erasmus MC, the Research Institute for Diseases in the Elderly (014-93-015; RIDE2), and the Netherlands Genomics Initiative (NGI)/Netherlands Organization for Scientific Research (NWO) Netherlands Consortium for Healthy Aging (NCHA), project 050-060-810. All studies are grateful to their participants, faculty and staff. The content of these manuscripts is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health or the U.S. Department of Health and Human Services.

The four LSACs are: the Human Genome Sequencing Center at the Baylor College of Medicine (U54 HG003273), the Broad Institute Genome Center (U54HG003067), The American Genome Center at the Uniformed Services University of the Health Sciences (U01AG057659), and the Washington University Genome Institute (U54HG003079).

Biological samples and associated phenotypic data used in primary data analyses were stored at Study Investigators institutions, and at the National Cell Repository for Alzheimer’s Disease (NCRAD, U24AG021886) at Indiana University funded by NIA. Associated Phenotypic Data used in primary and secondary data analyses were provided by Study Investigators, the NIA funded Alzheimer’s Disease Centers (ADCs), and the National Alzheimer’s Coordinating Center (NACC, U01AG016976) and the National Institute on Aging Genetics of Alzheimer’s Disease Data Storage Site (NIAGADS, U24AG041689) at the University of Pennsylvania, funded by NIA, and at the Database for Genotypes and Phenotypes (dbGaP) funded by NIH. This research was supported in part by the Intramural Research Program of the National Institutes of health, National Library of Medicine. Contributors to the Genetic Analysis Data included Study Investigators on projects that were individually funded by NIA, and other NIH institutes, and by private U.S. organizations, or foreign governmental or nongovernmental organizations.

Acknowledgments to ADNI

Data collection and sharing for this project was funded by the Alzheimer’s Disease Neuroimaging Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech; BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.; Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is providing funds to support ADNI clinical sites in 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 coordinated by the Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are disseminated by the Laboratory for Neuro Imaging at the University of Southern California.

Acknowledgments to AMP-AD

Mayo RNAseq Study- Study data were provided by the following sources: The Mayo Clinic Alzheimer’s Disease Genetic Studies, led by Dr. Nilufer Ertekin-Taner and Dr. Steven G. Younkin, Mayo Clinic, Jacksonville, FL using samples from the Mayo Clinic Study of Aging, the Mayo Clinic Alzheimer’s Disease Research Center, and the Mayo Clinic Brain Bank. Data collection was supported through funding by NIA grants P50 AG016574, R01 AG032990, U01 AG046139, R01 AG018023, U01 AG006576, U01 AG006786, R01 AG025711, R01 AG017216, R01 AG003949, NINDS grant R01 NS080820, CurePSP Foundation, and support from Mayo Foundation. Study data includes samples collected through the Sun Health Research Institute Brain and Body Donation Program of Sun City, Arizona. The Brain and Body Donation Program is supported by the National Institute of Neurological Disorders and Stroke (U24 NS072026 National Brain and Tissue Resource for Parkinson’s Disease and Related Disorders), the National Institute on Aging (P30 AG19610 Arizona Alzheimer’s Disease Core Center), the Arizona Department of Health Services (contract 211002, Arizona Alzheimer’s Research Center), the Arizona Biomedical Research Commission (contracts 4001, 0011, 05–901 and 1001 to the Arizona Parkinson’s Disease Consortium) and the Michael J. Fox Foundation for Parkinson’s Research.

ROSMAP

We are grateful to the participants in the Religious Order Study, the Memory and Aging Project. This work is supported by the US National Institutes of Health [U01 AG046152, R01 AG043617, R01 AG042210, R01 AG036042, R01 AG036836, R01 AG032990, R01 AG18023, RC2 AG036547, P50 AG016574, U01 ES017155, KL2 RR024151, K25 AG041906-01, R01 AG30146, P30 AG10161, R01 AG17917, R01 AG15819, K08 AG034290, P30 AG10161 and R01 AG11101.

ROSMAP (gene expression data)

The results published here are in whole or in part based on data obtained from the AD Knowledge Portal (https://adknowledgeportal.org). Study data were provided by the Rush Alzheimer’s Disease Center, Rush University Medical Center, Chicago. Data collection was supported through funding by NIA grants P30AG10161 (ROS), R01AG15819 (ROSMAP; genomics and RNAseq), R01AG17917 (MAP), R01AG30146, R01AG36042 (5hC methylation, ATACseq), RC2AG036547 (H3K9Ac), R01AG36836 (RNAseq), R01AG48015 (monocyte RNAseq) RF1AG57473 (single nucleus RNAseq), U01AG32984 (genomic and whole exome sequencing), U01AG46152 (ROSMAP AMP-AD, targeted proteomics), U01AG46161(TMT proteomics), U01AG61356 (whole genome sequencing, targeted proteomics, ROSMAP AMP-AD), the Illinois Department of Public Health (ROSMAP), and the Translational Genomics Research Institute (genomic). Additional phenotypic data can be requested at www.radc.rush.edu.

Mount Sinai Brain Bank (MSBB)

This work was supported by the grants R01AG046170, RF1AG054014, RF1AG057440 and R01AG057907 from the NIH/National Institute on Aging (NIA). R01AG046170 is a component of the AMP-AD Target Discovery and Preclinical Validation Project. Brain tissue collection and characterization was supported by NIH HHSN271201300031C.