Novel early-onset Alzheimer-associated genes influence risk through dysregulation of glutamate, immune activation, and intracell signaling pathways

Alzheimer Disease (AD) is a highly polygenic disease that presents with relatively earlier onset (≤70yo; EOAD) in about 5% of cases. Around 90% of these EOAD cases remain unexplained by pathogenic mutations. Using data from EOAD cases and controls, we performed a genome-wide association study (GWAS) and trans-ancestry meta-analysis on non-Hispanic Whites (NHW, NCase=6,282, NControl=13,386), African Americans (AA NCase=782, NControl=3,663) and East Asians (NCase=375, NControl=838 CO). We identified eight novel significant loci: six in the ancestry-specific analyses and two in the trans-ancestry analysis. By integrating gene-based analysis, eQTL, pQTL and functional annotations, we nominate four novel genes that are involved in microglia activation, glutamate production, and signaling pathways. These results indicate that EOAD, although sharing many genes with LOAD, harbors unique genes and pathways that could be used to create better prediction models or target identification for this type of AD


Introduction
Alzheimer Disease (AD) is a highly polygenic disease that often affects individuals over the age of 65-70 years but can affect individuals as young as 30 years old (yo) 1 .Many of the known genomic loci in uencing AD have been discovered through large-scale genome-wide association studies (GWAS) and family studies 2 mainly studying late-onset AD (LOAD).While LOAD dementia 3 is more common, the earlier onset form of AD (EOAD) makes up 5%-10% of all AD cases 4 .Additionally, it has a higher heritability 5 , has a more aggressive progression and presentation 6,7 , and often has a more severe impact on life and family.This is especially the case for high-risk populations such as racial and ethnic minorities 8 .Despite this, EOAD remains largely understudied 9 .
EOAD is characterized by AD onset before the age of 65-70 yo 5,10 , but this age cutoff is arbitrary and does not appropriately re ect the underlying biology 9 .In practice, age thresholds vary (often 60, 65, or 70) 11 depending on the scienti c question being studied.Previous studies have identi ed autosomal dominant mutations in well-established AD genes, APP, PSEN1, and PSEN2 7,[12][13][14][15][16] , which cause AD, often presenting with earlier-onset; however, such mutations are rare and are only present in about 10% of EOAD cases 17 .Given the limited genomic studies on EOAD, the genetic etiology of EOAD remains unclear.It is not clear if EOAD is genetically distinct from LOAD; if they are genetically identical, but with a spectrum of onset; or if there some but incomplete overlap.
The AD knowledge gap is further burdened by a lack of ethnic and racial diversity in studies 9 .One of the largest AD-risk GWAS of nearly 800,000 non-Hispanic white (NHW) participants, published by Bellenguez et al., 18 (referred to from here on as "Bellenguez") uncovered 75 genomic loci in uencing AD risk, 42 of which were novel.However recent studies that include non-European ancestries but with signi cantly smaller sample sizes identi ed novel loci additional loci.One such example is a recent LOAD GWAS of nearly 8,000 (2,748 cases, 5,222 controls) African Americans (AA), performed by Kunkle et al., 19 (referred to from here on as "Kunkle").Kunkle identi ed one novel genome-wide signi cant locus and ten novel disease associated loci compared to Bellenguez 18 .A more recent study of 9,168 AA was performed by Ray et al., 20 (2,903 cases, 6,265 controls) as a follow-up of the analyses by Kunkle.Ray et al., identi ed 12 additional novel disease associated loci, including one at genome-wide signi cance.In the most recent large-scale GWAS for 8,036 (3,962   cases, 4,074 controls) Japanese participants 21 (referred to from here on as "Shigemizu") a novel ancestry-speci c susceptibility locus near FAM47E was identi ed.Additionally, in a recent trans-ancestry meta-analysis with the 2013 IGAP study 22 , Shigemizu discovered a novel trans-ancestry susceptibility locus near OR2B2 21 .Lastly, Sarnowski et al., recently carried out an ancestry-speci c GWAS of circulating total tau levels in datasets of 14,721 NHW and 953 AA participants 23 .They identi ed three tau-associated loci in the AA dataset that were not replicated in NHW datasets.Furthermore, across these four non-NHW studies, besides APOE, only ten loci overlap with the now 80 discovered in NHW datasets.In addition, in a recent multi-ethic meta-analysis of 56,241 ADGC participants by Rajabli et al., 24 trans-ancestral meta-analyses identi ed two novel loci, and Lake et al., 25 (n = 644,188) identi ed two novel disease associated loci by meta-analysis of summary statistics from ve recent LOAD GWASs of various ancestries and a de novo GWAS of Caribbean Hispanics.Taken together, these studies clearly demonstrate that there is missing biological information surrounding AD that cannot be addressed solely by increasing sample size.Studies of multiple ancestral ancestriers are, thus, critical to identify AD-associated variants not present in NHW.
We hypothesize that by performing multi-ancestry EOAD GWAS, we will identify novel loci that will not be uncovered by studying only Europeans.In addition, as we and others have demonstrated that EOAD is enriched in genetic risk factors, which should lead to more statistical power.This will lead to identify novel loci as well as to determine the overlap and unique genetic architecture of EOAD compared to LOAD.To this end, we utilized genetic data of over 70,000 participants in three ancestries -NHW, AA, and East Asian (Asian)-from the Alzheimer Disease Genetics Consortium (ADGC) and the Charles F. and Joanne Knight Alzheimer Disease Research Center (Knight-ADRC) at Washington University in St. Louis to perform single variant analysis, trans-ancestry metaanalysis, and ne-mapping for gene identi cation in EOAD.

Analysis Datasets
This project uses genetic and phenotypic data from participants from the ADGC and Knight-ADRC who are self-described as either NHW, AA, Asian, or Hispanic (HIS).Imputation was performed using the TOPMed imputation server (GRCh38; Materials and Methods 4.2).Principal component analysis (PCA) was used to group subjects based on genetic similarity to populations of known ancestry (Figure S1A).Datasets were restricted to participants with of age at onset (AAO) ≤ 70 yo for EOAD and age at last visit > 70 yo for controls, to maximize statistical power across datasets.Sensitivity analyses, restricted to AD cases with age at onset ≤ 65, were also performed.The nal analysis datasets consisted of 19,668 NHW (6,282 CA, 13,386 CO), 4,445 AA (782 CA, 3,663 CO), and 1,213 Asian (375 CA, 838 CO).No analyses were performed for the ADGC Hispanic dataset due to the limited sample size at time of analysis (n = 550).A summary of the analysis work ow is displayed in Fig. 1A.Subject demographics are summarized in Table S1 for all subjects of each ancestry and is broken down by cohort in Table S2.

Single variant and multi-ethic meta-analysis
Within-ancestry single variant analysis identi ed 13 genome-wide signi cant loci across AA, Asian, and NHW.All 13 loci were identi ed in the NHW analysis (Figure S2A), while for AA (Figure S2B) and Asian (Figure S2C) only the locus near APOE -driven by the ε4 variant-was identi ed.Seven of the 13 novel loci are located near CR1, BIN1 TREM2, MS4A, PICALM, APOE, and LILRA5and were previously identi ed in LOAD 18,19,21,22,26 ; the remaining six loci were novel (Figure S2 and Table 1).AS this study includes dataset from multiple cohorts and arrays, we performed sensitivity analyses by cohort to con rm the analyses and determine potential heterogeneity for each locus.Analyses for each of these loci separated by cohort (40 NHW cohorts, 21 AA cohorts) only found signi cant heterogeneity (logP<-1.30103)for the APOE locus (logP NHW =-6.662, logP AA =1.532).For all the other sings, many cohorts contributed to these associations (Figure S3-S6, Table S4), and there was no signi cant heterogeneity for any of the loci besides APOE, suggesting low underlying variation between studies.For subsequent trans-ancestry meta-analyses, we used the summary statistics from all three ancestries in a random effects model in Plink v1.9.Metaanalysis identi ed two additional novel loci near TRIM49B (P Meta =1.02×10 − 08 ) on chromosome 11 and SSBP4 (P Meta =4.44×10 − 09 ) on chromosome 19 (Fig. 1B, Figure S7, Table 1).Both AA and NHW contributed to the signal for SSBP4 (P AA =0.021, OR AA =1.197; P NHW =1.30×10 − 07 , OR NHW =1.392; Table 1), but only NHW contributed to the TRIM49B signal.To determine how robust the analyses were, we also performed three additional meta-analysis models to determine how robust the results were..The additional models included I) an inverse-variance weighted, xed effects model in Metal; II) a basic xed-effects meta-analysis in Plink; and III) meta-regression with MR-MEGA, which was developed to be robust against heterogeneous effects.There were no additional loci, novel or known, identi ed from these other models, but both of the identi ed novel loci near TRIM49B and SSBP4 were genome-wide signi cant regardless of the model used (Table S3).
There were no genome-wide or suggestive associations with genetic variants in known causal Mendelian dementia genes (APP, PSEN1, PSEN2, and GRN; Figure S8-S9), suggesting that common variants in these genes do not signi cantly contribute to sporadic EOAD, or that we do not have enough power.
To identify additional independent signals at each locus, we then performed step-wise conditional analyses.The top hit at each ancestry-speci c locus was included as a covariate in single-ancestry analyses restricted to the genomic region at one megabase (1MB) anks to the GWAS signal.Each subsequent genome-wide signi cant variant was included in the model at each locus until signi cance was attenuated (P ≥ 5×10 − 08 ).Independent signals were identi ed in four of the NHW GWAS loci (Figure S10-S13, Table S3).In known loci, we identi ed one additional independent signal in NHW for the TREM2 locus (Figure S10).The original GWAS locus was driven by the TREM2 p.R47H variant, while the second independent signal is driven by the p.R62H variant.Additionally, there were eight independent signals in the locus near APOE (Figure S12).These, notably, are also independent of APOE4 (rs429358) and APOE2 (rs7412) as none are in linkage disequilibrium with those variants.At the novel loci, we identi ed one independent signal at the locus near ANKRD30BL (Figure S10C-D) and two independent signals at the locus near AC109635.7 (Figure S11A-C) in NHW.No independent signals were detected in AA or Asian (Figure S11D-E), likely due to lack of power.
Given the differing age thresholds for EOAD seen in AD literature (60, 65, 70) 5,9,10 we performed a sensitivity analysis for NHW only including cases with an onset 65 yo or younger (CA65).Asian and AA were excluded due to limited sample size and APOE being the only GWAS result.The new age trehsoild leads to a decrease of the number of cases in a 49% decrease.We compared the effect sizes of the sentinel passing the suggestive P value threshold, P < 1×10 − 05 (73 variants).Effect sizes (R 2 = 0.958, P = 6.74×10 − 51 , Figure S14A) were well correlated between our main analysis and the CA65 analysis and all betas were in the same direction, suggesting the results including the cases with onset below 70 is similar to those from 65.
Finally, we run another sensitivity analyses by removing array as covariate, as this model can lead to overcorrection as in some instances there may be a collinearity between array and cohort, which is also included as co-variate.We found effect sizes were very strongly correlated (R 2 = 0.960, P = 1.74×10-51; Figure S14B, ).However, this model identi ed two additional novel loci near MSMB (chr10:47025293) and IGLV (chr22:22690934; Table S3).These results suggest the model without array can identify additional loci which may have been missed by our default model.However, in order to be conservative, and present the more robust results the post-GWAS analysis (below) were done with he model included array as covariate.

Overlap of EOAD and LOAD
Another goal of this project was to understand the degree of overlapping genetic architecture between EOAD and LOAD.As described in the previous section, we identi ed seven genomic loci which were previously identi ed in LOAD GWAS by Bellenguez 18 , Kunkle 19 , or Shigemizu 21 at genome-wide signi cance (Table S3).Speci cally, the lead variants near CR1, BIN1, TREM2, MS4A, and PICALM in the NHW EOAD analyses were also signi cant in Bellenguez 18 , that also showed strong colocalization (PP.H4 > 0.984; Table 1, Table S3 and Table S5).The lead variant in the AA and Asian APOE locus, the APOE ε4 variant, was signi cant in all LOAD and EOAD analyses.The summary statistics from Bellenguez et al., did not included many variants in the APOE locus and therefore no therefore colocalization analyses could not been performed.
Additionally, it should be noted that signal in AA at the BIN1 locus colocalize with that form Bellenguez 18 , although this locus did not pass the genome-wide threshold, but reached nominal signi cance in both AA (P = 1.07×10 − 05 , OR = 1.400;Table 1).This in combination with the main GWAS results suggests that BIN1 might be a risk locus for EOAD in AA as well, but we lack the power to detect it.We also noted that our limited sample sizes for non-white ancestries may limit our power to detect other loci which might otherwise reach signi cance.However, these results suggest shared loci between EOAD and LOAD are likely driven by the same variants.
For the novel loci, only variants near TRIM49B and SSBP4 loci were nominally signi cant in Bellenguez (P = 0.022 and 0.043, respectively; Table S3) 18 and the lead variant near AC109635.7 was nominally signi cant in Kunkle (P = 0.029; Table S3) 19 .We believe these ndings reinforce the idea that EOAD has at least a partially unique genetic architecture compared to LOAD.

Total genetic covariance between EOAD and LOAD
For a broader view of genetic overlap between EOAD and LOAD, we use two methods to examine the genetic correlation between the two clinical presentations.First, we used polygenic risk scores (PRS) generated with LOAD summary statistics as a predictive variable for EOAD status in logistic regression and determine signi cant correlations.We found that PRS generated with genome-wide signi cant SNPs (P ≤ 5×10 − 08 ) from Bellenguez 18 was strongly correlated with NHW EOAD status (R 2 = 0.170, P < 10 -300 ; Fig. 2B Figure S15A).Noting that APOE is likely driving most of this association, we also performed the same analysis after removing the APOE region (chr19: 44,000,009-47,999,435) from LOAD summary statistics.After removing the APOE region, there was still a strong and signi cant correlation (R 2 = 0.048, P = 1.20×10 -143 ; Fig. 2B, S15B).
Conversely PRS using genome-wide SNPs from Shigemizu 21 had a very weak correlation with EOAD status in Asians.We did not nd strong correlation using Shigemizu PRS: neither with APOE (R 2 = 0.008, P = 0.007) nor without (R 2 = 0.005, P = 0.036; Figure S15C-D respectively).This may be due to relatively low overlapping variants between Asian EOAD and Shigemizu LOAD analysis 21 as well as there being only three loci passing the 5×10 − 08 threshold in Shigemizu's analysis 21 .Using a relaxed threshold of 5×10 − 05 to calculate PRS, we see much stronger correlation as the R 2 increases to 0.085 with APOE and 0.084 without APOE (P = 1.13×10 -16 and 2.64×10 -16 respectively, Figure S15C-D).This suggests that there is, in fact an overlap between LOAD and EOAD in Asian populations but the current sample size is not large enough to determine the extent of the overlap.It should be noted; however, that Shigemizu's analysis focuses on Japanese datasets 21 , while our Asian dataset largely consists of Japanese participants, there are still other East Asian samples included which may impact the risk score correlation.Finally, PRS using GWS from Kunkle 19 correlated signi cantly with EOAD AA (R 2 = 0.063, P = 1.46×10 -37 ) with APOE included.However, APOE was the only GWS signal in Kunkle 19 , so strong correlation here is to be expected.Correlation was still robust using Kunkle 19 SNPs at the P value threshold 5×10 − 05 , both with APOE (R 2 = 0.020, P = 1.61×10 -13 , Figure S15E) and without APOE (R 2 = 0.061, P = 5.63×10 -37 , Figure S15F).Overall, the PRS results suggests that there is covariance between EOAD and LOAD in other loci besides APOE in NHW.There is also covariance between LOAD PRS and EOAD status in Asians and AA; however, we suspect the limited covariance at lower p-values is due to low sample size in those datasets.
Speci cally, the relatively low statistical power restricts the number of loci which contribute to PRS.This as well as that there are potentially overlapping subjects between our AA analysis and those used by Kunkle et al., 19 makes interpretation of those results di cult.

Gene prioritization in GWAS loci
To help determine the causal gene at each locus we prioritized functional genes by creating a 50-point weighted presence/absence matrix (Table S6) summarizing evidence from colocalization with molecular quantitative trait loci (QTL), including CSF protein QTLs (pQTL) and brain expression QTLs (eQTL); locus and lead variant overlap with those same molecular phenotypes; Multi-marker Analysis of GenoMic Annotation (MAGMA) 27 gene-based analysis, and variant annotation, including the nearest gene and genes with protein-altering variants in our GWS loci.We generated scores for all genomic features (genes, long non-coding RNA, and pseudogenes) within 1MB anks of the top hit in each locus according to the guidelines in Table S6.Genes had to pass a minimum score threshold of four (≥ 4) to be nominated.At each locus, genes that had a ≥ 20% relative difference compared to the top scoring in the locus were not prioritized.
To internally validate our prioritization methods, we applied our scoring strategy to previously identi ed AD loci to see if we would prioritize the same functional gene as is suggested by literature.That is, we tested to see if our methods would prioritize CR1, BIN1, TREM, MS4A4A/MS4A6A, PICALM and APOE as literature has suggested these genes to be the functional gene at their respective loci.With regards to the LILRA5 locus, Bellenguez et al., 18 prioritized LILRB2 as a tier 2 gene for the locus based on being the nearest gene and enrichment for expression in microglia.Given the lack of any other GWAS or functional analyses strongly supporting this idea, we did not use this locus as a basis of whether our prioritization methods were successful.
For the novel loci we were able to prioritize four genes for four out of eight loci (Fig. 2A).Speci cally, we prioritized CDH12 (chr5p14.3),FOLH1 (chr11p11.12),ALG10B (chr12q12), and LRRC25 (chr19p13.11).Notably, FOLH1 and LRRC25 are prioritized within our two novel trans-ancestry loci.CDH12 is a cell adhesion gene expressed in vesicles and plasma membrane that functions in cell adhesion 28,29 .CDH12 was the sole gene prioritized in the novel chr5:21360122 locus based on evidence from QTL mapping and annotation.Speci cally, the GWAS locus overlapped signi cant cis-pQTL (Fig. 2A, Table S9) and cis-eQTL (Fig. 2A, Table S10) loci, as well as it is mapped as the nearest gene to at least one genome-wide variant in the locus by FUMA.FOLH1 is expressed extracellularly and in plasma membrane.It has speci c brain functions for the modulation of excitatory neurotransmission, particularly as it relates glutamate activity 28,29 .It was the second highest scoring gene of the four, having MetaBrain eQTLs that colocalized (Fig. 2A, Table S11) with the GWAS locus (chr11:49032935).While the colocalization evidence alone was enough to prioritize this gene for this locus, it is additionally strengthened by having genome-wide signi cant loci for all three assessed molecular QTLs overlapping the GWAS locus (Table S12).ALG10B is primarily localized to the endoplasmic reticulum (ER) and functions in transferase activity and protein metabolism 28,29 .Like FOLH1, this gene was supported by GTEx eQTLs that colocalize with the GWAS locus, which alone was su cient for its prioritization.The only other evidence for this gene was an overlapping locus with MetaBrain eQTLs.Lastly, there is LRRC25 which functions in NF-kappa-B signaling pathway and type I interferon signaling pathway inhibitions 28,29 .Its expression is localized to the cytosol, ER, and microtubules.This gene had the most supporting evidence for its prioritization.Functional evidence for this gene includes GTEx and MetaBrain eQTLs colocalizing with the GWAS locus, a genome-wide signi cant eQTLs within the locus, and the lead SNP (chr19:18422832) is itself a genome-wide signi cant eQTL in GTEx.The gene, ELL, was also a candidate for prioritization as GTEx eQTLs for the gene colocalized with the GWAS locus (Table S8, Table S10).It does not have as much supporting evidence, but the colocalization result suggests it could be a secondary actor in the locus for in uencing EOAD.

Biological insights for GWAS loci and prioritized genes
To further disentangle the biological relevance of these of prioritized genes in our GWAS loci, we leveraged data from ve broad categories: I) brain cell-type speci city, II) expression in AD vs Controls, III) protein-based evidence, IV) associations with other phenotypes, and V) enrichment in biological pathways.The ndings of these analyses are summarized in Fig. 2C-D.
2.5.1.Evidence for microglia activation and autophagy in novel prioritized genes We used cell-type speci city data from Western et al., 30 to query our prioritized genes for a primary cell-type in brain tissues (Methods 4.7.1).We particularly took note of any genes which were enriched in microglial expression because several genes known to affect AD (e.g., CR1, MS4A, HLA, and TREM2) 2,31-34 are microglia genes involved in pathways related to amyloid beta (Aβ) clearance or autophagy, as is LRRC25.In addition to its functional role described above, LRRC25 regulates virally induced autophagy.It was also highlighted as an AD risk gene by Kosoy et al., 35 , so it may serve a similar phagocytotic function as known microglia-speci c AD risk genes.To investigate this hypothesis, we leveraged STRING 36 protein-protein interaction (PPI) analysis and brain tissue expression data from Agora to identify a functional relationship between LRRC25 and known AD microglial genes.In our PPI analysis, we queried genes which passed the score threshold for prioritization (Table S8) in known and novel GWAS loci and used all protein coding genes as the background set.However, due to pleiotropy within the APOE locus, genes from that locus were not included.STRING identi ed signi cantly more interactions than expected by chance (PPI enrichment P = 2.36×10 − 10 ).Notably there were text-mining, experimentally determined, and co-expression edges shared between LRRC25 and CR1/CR2, TREM2, LILRA5 and LILRB2, all of which have microglia related functions in AD (Fig. 2C, Figure S16A).It should be noted that we see similar edges between ELL and CR1/CR2, we do not, at this point, suspect it functionally affects AD through microglia-mediated functions.In support of the PPI ndings, we also found from Agora expression data that the Log2-fold change (L2FC) for LRRC25 (0.47) was nearly identical to that of CR1 (0.36) in AD cases compared to controls in two brain tissues-the inferior frontal gyrus (IFG) and parahippocampal gyrus (PHG), which is not observed novel-known gene pairing in the same tissue (Table S13).This result supports the PPI nding of a co-expression relationship between the two genes as well as add tissue localization context for their relationship.Further supporting the role of microglia and autophagy in EOAD are the results from our colocalization between EOAD GWAS loci and trans-pQTLs.This analysis was used to determine which protein might have a distal functional relationship with our GWAS loci as was done for MS4A and sTREM2 by Deming et al., 32 .The results of our analysis exclude pQTLs which colocalize with the APOE locus, as it is known to have broad pleiotropic effects.
We found EOAD variants in the BIN1 locus colocalized with a trans-pQTLs for TMEM132C (Table S14).TMEM132C codes for trans-membrane protein 132C whose expression is enriched in oligodendrocytes 37 and localized to cytosol and centrosome 38 according to data sourced from proteinatlas.org.There are no functional annotations which elucidate how the BIN1 locus might be impacting AD through regulation or modi cation of TMEM132C, however, TMEM106B is a known dementia-associated gene which has functions in lysosome function and homeostasis.Being from the same family of genes, TMEM132C may work alongside BIN1 in autophagy pathways 39 .Speci cally, one study has found that BIN1 contributes to early endosome size deregulation 40 , and that may be due attributed to pQTLs near BIN1 modifying TMEM132C function.In addition to microglia and autophagy pathways, animal studies have found that BIN1 modulates glutamate activity 41 , which we believe is an alternative pathway contributing to EOAD.

Novel prioritized gene involvement in glutamate dysregulation
In addition to microglia-mediated autophagy pathways, neurotransmitter dysregulation can also impact AD onset and progression.Speci cally, glutamate is an excitatory neurotransmitter which can be neurotoxic at high levels and is found to be dysregulated in AD 8, [42][43][44] .Information from public resources suggest that FOLH1 contributes to the modulation of glutamate so we leverage proteomic data, mRNA brain expression information from Agora, pathway analyses, and associations with other phenotypes to nd evidence supporting the hypothesis that FOLH1 affects EOAD through its modulation of glutamate.In proteomic analyses we use pQTL data from Western et al., 30 for colocalization analsyes, proteome-wide association study (PWAS) and Mendelian randomization on our prioritized genes.While the former two analyses provided no signi cant results for any novel prioritized genes (Figure S16B), we do nd that FOLH1 was signi cant in MR analysis when using pQTL as instrument variables (P = 8.00×10 − 07 ; Table E2) and thus have a likely causal effect on EOAD.Based on data from Agora, it is signi cantly upregulated in parahippocampal gyrus (PHG) of cases compared to control.Because of the positive relationship between FOLH1 and glutamate -that is, it hydrolyzes N-aceylaspartylglutamate to release glutamate 45 -this result suggests that FOLH1 is causing more glutamate to be excreted in that brain region and causing a neurotoxic effect.Notably, FOLH1 is signi cantly downregulated in cases compared to controls in anterior cingulate cortex (ACC) and cerebellum (CBE).We can speculate that this difference indicates differences in Amyloid beta (Aβ) presence in those tissues.We suspect that is the case because TREM2, which recruits microglia for Aβ clearance 46 , is also downregulated in CBE in cases compared to controls, while it is upregulated in PHG.Studies have suggested that Aβ contributes to synaptic failure through modifying the glutamate related systems 44 .Thus, lower levels of FOLH1 and TREM2 may indicate a lack of Aβ accumulation in those tissues.These speculations are consistent with our pathway analysis results which found that FOLH1 is enriched in disease-gene network analysis for memory impairment, amyloid plaque, memory loss, presenile dementia, and memory impairment (Fig. 2D, Table S15).
Our biological analyses also implicate ion ow dysregulation as a potential pathway driving EOAD.Speci cally, the data suggests calcium ion (Ca 2+ ) 47 and potassium ion (K + ) 48 dysregulation may contribute to glutamate dysregulation and subsequent neurotoxicity.This is supported by several points evidence.
First, we see calcium transport, calmodulin binding, and calmodium-dependent signaling pathway enrichments for known AD genes BIN1, TREM2, LILRA5, and LILRB2 (table S15).Calmodulin acts downstream of glutamate production by binding Ca 2+ ions which pass through the glutamate activated NMDA receptor 47,49 .NMDA receptors are expressed on glial cells and can induce pro-in ammatory responses 50 .This increased activation NMDA receptors may additionally lead to downstream signaling changes such as inactivating extracellular signal-regulated kinases (ERK) 49 , which may explain our CDH12 association.CDH12 codes for the Cadherin 12 protein which functions in calcium-dependent cell adhesion and is implicated in synaptogenesis, cell junction organization, and ERK signaling 28,29 .According to data from Agora, CDH12 is signi cantly downregulated in seven AD-relevant brain regions (Table S13), so this is consistent with expectations given lower ERK signaling 51 .Regarding K + dysregulation, we also see enrichment in potassium transport pathways implicating TREM2, BIN1, and our novel prioritized gene ALG10B (Table S15).Glutamate transporter genes, such as GLT1, function by co-transporting three sodium ions and one hydrogen ion into the cell, while counter-transporting one potassium ion 52 .Dysfunction of ALG10B, which normally functions in glycosylation 53 and inward recti er potassium channel regulation (Fig. 2D, Table S15, Figure S16), may lead to an inability of glial cells to transport glutamate out of the post synaptic space, leading to neurotoxic accumulations.

Protein-proteins interactions for novel prioritized genes in EOAD
To further elucidate potential latent pathways associated with EOAD, we also tested for complex gene interaction and if any modi able risk factors are associated with our prioritized genes and with AD.To support this direction, we look at results from trans-pQTL colocalization, GWAS catalog associations with AD risk factors, and metabolite QTL (metabQTL 54 ) colocalization.
Next, we queried the GWAS catalog (https://www.ebi.ac.uk/gwas/) 55 for other phenotypes that implicate prioritized genes in our novel loci.Speci cally, we selected all genes which were within our novel GWAS locus and which had a prioritization score greater than or equal to four (≥ 4).We then searched for those genes within "mapped gene" and "reported gene" elds from the GWAS catalog results 55 .We additionally noted if any of those phenotypes were in one of four AD relevant categories: I) dementia; II) AD biomarkers such as TREM2 or Tau; III) other neurological or psychiatric traits such as depression, stroke, or schizophrenia; and IV) AD risk factors such as, cancer, heart disease, or diet.Starting with dementia associations, we found that variants mapped to LRRC25 were previously reported to be associated with AD-related phenotypes (Table S16), speci cally with GWAS of upper vs lower quantiles of AD PRS (rs754032589) 56 and a GWAS of AD and gastroesophageal re ux disease (rs3859570; Table S16) 57 .Looking at AD biomarker associations, variants in the CDH12 gene region have been reported to be associated with PHF-tau measurement (rs2516478/rs1261246 and rs10805748/rs2250562) 58 .Next, we found that rs61350355 in the FOLH1 locus is associated with amygdala volume change rate 59 , rs147303113 (LRRC25) is associated with brain shape measurement 60 , and variants mapped to CDH12 are associated with cognitive domain measurement (chr5:22528391) 61 , cognitive function measurement (rs183856) 62 , and schizophrenia (rs11738207 and rs2680786 63 ; Table S16).Finally, variants mapped to each of the novel loci prioritized genes are associated with 21 phenotypes across 10 unique modi able AD risk factors including alcohol consumption, blood pressure, cancer, cardiovascular disease, diabetes, diet, educational attainment, infection, lipid measurements, and body weight (Table S16).GWAS catalog hits mapped to LRRC25 and FOLH1 55 are correlated with our GWAS signal, adding con dence that these functions are linked for those genes.We did not nd similar correlations for CDH12 or ALG10B mapped variants, so more work needs to be done to validate the function of these loci to EOAD.
Lastly, colocalization with CSF metabolites QTLs 54 identi ed 13 unique metabolites which had metabQTLs colocalize with three EOAD loci.Eleven of the colocalization results were driven by the APOE locus (Table S17).The remaining two results were colocalization between myo-inositol metabQTLs and the SSBP4 locus, and colocalization between metabQTLs for an unnamed metabolite internally referred to as "X-13684 with the chr5:21360122 locus using NHW summary statistics (Table S17).The colocalization result with myo-inositol is likely driven by ELL, which is linked to myo-inositol levels in the GWAS catalog 55 (Table S16).Myo-inositol is synthesized in high concentrations in the brain where it plays a role in facilitating the binding of neurotransmitter and some steroid hormones to their receptors.This may suggest that ELL plays a similar role to FOLH1 in EOAD.Furthermore, a query of The Human Metabolon Database found that myo-inositol levels are reduced in patients with depression 64 , which can be a symptom of AD.One other study has previously investigated a link between myo-inositol levels and AD but did not nd a signi cant difference in the metabolite levels between cases and controls 65 .Our results may indicate that the relationship between myo-inositol and AD is earlier-onset speci c.

Discussion
The goal of this study was to identify novel genes and genomic loci associated with EOAD and to determine the degree of overlap between the genetic architectures of EOAD and LOAD.Using genetic and phenotypic data from 19,668 NHW, 4,445 AA, and 1,213 Asian participants from ADGC and Knight-ADRC, we performed the largest multi-ancestry GWAS of EOAD to date.Our analysis identi ed eight novel loci, including two of which were only identi ed in transancestry meta-analysis (Fig. 1B).
A total of seven loci were already previously identi ed in at least one of the LOAD analyses by Bellenguez 18 , Kunkle 19 or Shigemizu 21 .On the other hand, ve of the eight novel EOAD loci were not observed even at nominal signi cance in previous LOAD analyses (Table S3).Those which were at nominal signi cance include the trans-ancestry loci near TRIM49B and SSBP4, which were nominally signi cant in Bellenguez 18 (P = 0.022 and P = 0.043, respectively), and the NHW locus near AC10635.7,which was nominally signi cant in Kunkle (P = 0.029) 19 .This suggests that EOADs underlying genetic architecture is at least partially unique from LOAD and genes in these novel loci either are not contributing to LOAD risk or current studies have not been able to capture them.The lack of overlap between LOAD and EOAD is possibly a consequence of LOAD having more age-related concomitant health issues and overlap with other lateonset dementias than EOAD.
Through colocalization results, we found overlap at previously identi ed loci near CR1, BIN1, TREM2, PICALM and APOE, suggesting they are likely to share causal variants in EOAD and LOAD.The colocalization at these loci is likely to be driving the overall genetic overlap between LOAD and EAOD (LDSC rg = 0.791).At the same time, this nding further suggests that our novel loci indicate that EOAD has a partially distinctive underlying genetic architecture from LOAD.Additionally, we also saw strong correlation between EOAD status and LOAD PRS in NHW, even without APOE (Fig. 2B), which was expected given the number of overlapping loci and the LDSC results.However, LOAD PRS results for AA and Asian suggests that Japanese and AA LOAD risk studies do not have su cient power to detect many loci that can be used for PRS.Taking from the results in NHW, we propose that genes in previously identi ed loci are likely strong drivers of all AD, but novel loci are likely contributing to differences seen between the two AD types.Future EOAD-speci c analyses and larger trans-ancestry analyses will be important to con rm this nding.
ancestry LOAD analyses by Lake et al., 25 and Rajabli et al., 24 .Lake et al., were able to identify two novel trans-ancestry LOAD loci by meta-analysis of summary statistics from ve recent LOAD GWASs; from Bellenguez 18 , FinnGen Release 6 (https://www.nngen./en/access_results), Shigemizu 21 , Kunkle 19 , and a de novo GWAS of Caribbean Hispanics (Ncases = 54,233, Nproxy-ADD = 46,828, Ncontrols = 543,127).We checked to see if any peaks from Lake et al., or Rajabli et al., overlapped with this EOAD analyses, but none of the top hits from those analyses passed nominal signi cance in EOAD analyses.This reinforces the premise of EOAD and LOAD having at least partially distinct genetics and adds that this is consistent within and across ancestry.However, while there were no GWAS results in common with our analysis, we do nd some gene and pathway overlap.Speci cally, Rajabli et al., propose the nearest gene to one of their GWAS loci, LRRC4C as a potential AD functional gene.LRRC4C is part of the same leucine-rich repeat family of genes as our prioritized gene LRRC25.LRRC4C has been found to be implicated in neurodevelopmental disorders including developmental epileptic encephalopathy 28,29 .This may implicate this family of genes as having important latent functions which impact AD.Additionally, previous studies implicate insulin receptor activity regulation by GBR14 24 .This is in line with our ndings of a relationship between ELL and myo-inositol in EOAD (Table S17), as myo-inositol also modulates insulin-mediated signaling.We believe it would be di cult, if not impossible, to identify these genes and pathways by large-scale analysis of NHW populations alone or without comparisons of EOAD and LOAD.This reinforces the necessity for more and larger GWASs of EOAD vs LOAD and broad ancestral origins.
Lastly, we nominated what we believe are the likely functional genes in four novel GWAS loci; CDH12 at the chr5:21360122 locus, FOLH1 at the chr11:49032935 locus, ALG10B at the chr12:37412586 locus, and LRRC25 at the chr19:18422832 locus.The combination of our genetic and biological results suggests that FOLH1 and LRRC25 are our strongest candidate genes for affecting EOAD (Fig. 2A), but we CDH12 and ALG10B are important downstream elements which exacerbate neurodegeneration.Both LRRC25 and FOLH1 had strong genetic evidence from colocalization and shared loci GTEx and MetaBrain eQTLs.Biologically, LRRC25 is enriched for expression in microglia, a cell type which is known to be functionally relevant in AD 18,35,66 .It has functions in NF-kappa-B and type I interferon signaling pathway inhibition and virally induced autophagy, so it is likely contributing to EOAD through known microglial activation and autophagy pathways.This is supported by results from STRING, which suggests that it is enriched for interactions with CR1, TREM2, LILRA, and LILRB2, all of which function in similar pathways.Additionally, LRRC25 follows a similar expression pattern to CR1 in parahippocampal gyrus (PHG) and superior temporal gyrus (STG; Table S13), so it is likely affecting EOAD through its relationship to CR1 in those brain regions.Lastly, it may also be affecting EOAD through a functional relationship with ELL, which is enriched in LOAD and tauopathy in STRING human phenotypes.On the other hand, its interaction with myo-inositol, as suggested by metabQTL colocalization (Table S17) and GWAS catalog results (Table S16), indicates it may affect EOAD through its relationship with neurotransmitters as some can be neurotoxic in high concentrations.That neurotransmitter interaction may also mean ELL is related to FOLH1.As described in the results, FOLH1 is enriched for expression in oligodendrocytes and functions in modulating excitatory neurotransmission, speci cally glutamate activity 45 which is neurotoxic in levels.FOLH1 is signi cantly higher expressed in PHG of AD cases compared to controls so this may be what drives the relationship between this gene and AD.Furthermore, it is enriched in disease-gene networks for memory loss and amyloid plaque, as well as it is associated with multiple glutamate phenotypes and amygdala volume change rate in the GWAS catalog 55 , giving us high con dence that this gene is associated with AD via this pathway (Table S15).Finally, our analyses suggest that glutamate dysregulation may also be driving CA 2+ and K + ion ow dysregulation.Glutamate release into the synaptic space drives CA 2+ transport into NMDA receptors 47,49,50,67,68 , which in turns modulates calmodulin-binding and signaling pathways which are downstream of it 47,49 .These downstream pathways include ERK signaling, which is modulated in part by CDH12 51 and NF-kappa-B signaling, which is modulated in part by LRRC25 69 .Glia cells may also fail to transport glutamate out of the synaptic space because of ALG10B dysfunction and subsequent dysregulation of potassium channels that enable glutamate uptake 48 .This study does have some weaknesses that need to be acknowledged.Firstly, despite a similar sample size and single variant analysis model to the work done by Rajabli et al., 24 , this study identi ed more novel associations with AD.Given the limited availability of large-scale EOAD analyses for comparison, it is challenging to entirely dismiss the possibility of false positives arising a consequence of our study design.Speci cally, one might suspect that our GWAS results are a consequence of genetic variability from the number of cohorts included in the single variant analysis for each ancestry.However, when array, which functions as a partial proxy for cohort, was included as a covariate in our model, there was very little genomic in ation in our single variant analysis (λ NHW = 1.045, λ AA = 1.014, λ Asian = 1.001), and we found no evidence of heterogeneity within novel loci as well as many cohorts contributed to each signal (Figure S4-S6).This, in addition to the fact that we were able to identify prioritized genes for 50% of our GWAS loci, suggests that including ten genetic PCs in the model as well including array as a covariate was su cient to control variability and gives us con dence that most, if not all, of our signals are likely real.We also must acknowledge that although we aimed to add to the eld's understanding of AD biology by including multiple ancestries, we lacked the power to identify novel loci speci c to AA or Asian.However, including those ancestries allowed us to identify two novel trans-ancestry signals and we had su cient evidence from molecular traits to prioritize strong candidate genes for those loci.
In summary, we performed the largest trans-ancestry GWAS of EOAD to date, identifying eight novel ancestry-speci c and trans-ancestry loci for this form of AD.We established lines of similarity and distinction between EOAD and LOAD based on ours and other large-scale AD GWAS analyses.We were able to use proteomic, transcriptomic, and metabolomic data to identify likely functional genes in our GWAS loci and we highlight FOLH1 and LRRC25 as likely driving EOAD through microglia activation/autophagy pathways, and dysregulated excitatory neurotransmitter function, respectively.

Cohorts
This project used genotype and phenotype data of participants (n=70,620) who self-identi ed as either Non-Hispanic White (NHW, n=50,180), African American (AA, n=8,563), Asian (n=4,742), or Hispanic (HIS, n=2,292) from the Alzheimer's Disease Genetics Consortium (ADGC) as well as participants from the Knight Alzheimer's Disease Research Center (Knight-ADRC)(n=4,843).The ADGC collects data from multiple genotyping rounds from several studies.4.5 Investigating the shared genetic architecture of EOAD and LOAD Methods to investigate overlap between EOAD and LOAD include, cross-checking the sentinel hit in GWAS loci, correlation between LOAD PRS and EOAD status, genetic covariance using Linkage Disequilibrium SCore (LDSC) 74,75 regression local covariance using SUPERGNOVA 76 , and nally, colocalization.

Cross-check of lead SNP in GWAS loci
To cross-check sentinel hits in GWAS loci, we directly compared the P values of the sentinel variant from our EOAD GWAS with the same variant in LOAD summary statistics from the most recent LOAD GWASs of NHW, AA, and Japanese participants.4.5.2Correlation of EOAD status and LOAD PRS Polygenic risk scores (PRS) were calculated using PRSice2 77 with LOAD summary statistics as the base GWAS and EOAD plink les for the same ancestry as the target.Risk scores were generated for two different models, one including the APOE region (chr19; 44,000,009-47,999,435 bp) and one excluding it.
PRSice2 was run using LOAD SNPs at the following p-value thresholds; 5×10 -08 , 5×10 -05 , 5×10 -02 , and 5×10 -01 in each model.The clumping p, R 2 , and kb thresholds were set at 1, 0.1, and 250, respectively.We also included a phenotype le containing EOAD status in the PRSice input and used the software to perform logistic regression (EOAD status ~ LOAD PRS) for each P threshold in each APOE model.

Local genetic covariance
Local covariance was calculated using Super GeNetic cOVariance Analyzer (SUPERGNOVA) 76 .LOAD and EOAD summary statistics were prepared by using the munge.pyprogram from LD SCore (LDSC).hg19 formatted summary statistics were used for these steps as they were already prepared including the rsID for each variant as well as this format matched the b les and partitioned .bedles provided by the SUPERGNOVA tutorial also in hg19 format.

EOAD and LOAD colocalization
Colocalization between EOAD and LOAD was performed using the coloc.abffunction in the Coloc R package.Loci of interest for comparing the two datasets were identi ed by taking 1MB upstream and downstream of the sentinel SNP in each EOAD GWAS locus.Within each locus, the set of overlapping variants between the two datasets were used to run colocalization.Loci were only considered to colocalize if hypothesis four (variant is causal for both phenotypes) had a posterior probability greater than or equal to 0.8 (PP.H4 ≥ 0.8).

Gene Prioritization
To identify likely causal genes within GWAS loci, we employed a summation of scores from a weighted presence-absence matrix for all genes within 1MB anks of the sentinel variant in GWAS loci.The presence-absence matrix is lled based on evidence of overlap with molecular quantitative loci (QTL), if a gene is signi cant in gene-based analysis, if a gene is the nearest gene to a signi cant variant, and if a signi cant variant is a non-synonymous exonic variant for a gene.Overlap with molecular QTLs was assessed for protein QTLs (pQTL) and expression QTLs (eQTL) in three categories: QTLs for a given gene colocalize with EOAD at a GWAS locus, a given gene has a shared locus with EOAD, or the lead variant in a GWAS locus is a genome-wide signi cant QTL for a gene.A shared locus is de ned as any genome-wide signi cant QTL within 1MB anks of the lead variant for a GWAS locus.The speci c scoring framework is described by supplementary Table S7.

Generating and accessing QTL data
Cerebrospinal uid (CSF) pQTL data was generated using CSF levels of 7,584 unique aptamers (6,179 unique proteins) measured on the SOMAscan7k platform for 4,223 participants from six dementia relevant cohorts 30 .pQTL summary statistics used for post-GWAS analyses is generated by meta-analysis of the discovery (n=1,912) and replication (n=1,195) datasets.Cis-pQTLs were variant-aptamer associations within 1MB in either direction of the target protein-coding gene's hg38 coordinate-based transcription start site.eQTL data were downloaded from GTEx Portal v8 for all Brain tissues, accessed on June 29, 2022.Additional cis-eQTL data was downloaded from the most recent MetaBrain analysis from the MetaBrain website (https://www.metabrain.nl) on downloaded February 24, 2023.CSF metabQTL data was generated using metabolite levels measure by Metabolon for 1,224+1,087 (discovery and replication) participants from ve dementia relevant cohorts, while brain metabQTLs are generated using parietal, prefrontal cortex, and temporal cortex post-mortem brain biopsies from 1,172 participants from four dementia-relevant cohorts 54 .CSF metabQTLs analysis is run separately in the discovery and replication sets, then meta-analyzed for the nal summary statistics.Further information on pQTL methods can be found in the source material Western et al., 30 and metabQTL methods can be found in source material from Wang et al., 54 .

Colocalization with molecular QTLs
Colocalization with molecular QTLs was performed identically to as described in methods section 4.5.4.All variants within 1MB anks of the sentinel SNP in each GWAS locus were subset in EOAD summary statistics.QTL data was subset similarly per tissue, per gene.Overlapping variants between the two datasets were used to run colocalization.pQTLs were only considered to colocalize if hypothesis four (variant is causal for both phenotypes) had a posterior probability greater than or equal to 0.8 (PP.H4 ≥ 0.8).

Overlap with molecular QTLs
Overlap with molecular QTLs was queried as described in Methods section 4.6.For each GWAS locus, we determine bounds by taking 1MB upstream and downstream anks to the lead SNP in the locus.We then subset the molecular QTLs for each gene for each tissue to those loci and query for any genome-wide signi cant SNPs.

Annotation and gene-based analysis
Annotation and gene-based analyses were performed for each single ancestry and meta-analysis using FUMA software as described in Methods section 4.4.For the purposes of gene prioritization, two annotation items were analyzed.First, a gene was scored if it was the nearest gene to a genome-wide signi cant variant in the locus.Second, a gene was scored if a genome-wide signi cant variant in the locus was a non-synonymous exonic (protein altering) variant.For MAGMA gene-based analysis, a gene was scored simply if was signi cant in gene-based analysis for any ancestry or meta-analysis after multiple testing correction (Bonferroni adjusted P = 0.05 / n genes).The speci c cutoff for each ancestry varied, but the number of genes was ~19,000 for each, leading to an approximate cutoff of 2.63×10 -06 .4.7 Biological inference of prioritized genes Prioritized genes were assessed in four main categories: Primary brain cell type, differential expression in AD brain tissues, association with related phenotypes, biological pathways, and protein-protein interaction.In each we used online methods, protein-based analysis, or pathway analysis, which will be further described below.

Primary Brain cell type
Cell-type speci city analysis was performed by Western et al., 30 at the NeuroGenomics and Informatic center at Washington University in St. Louis.The methods are described in full in their publication.As an overview, gene expression data for human astrocytes, neurons, oligodendrocytes, microglia/macrophages, and endothelial cells were downloaded and used to determine a primary cell-type for all SOMAscan 7K panel proteins.Gene expression was averaged across participants for each cell type, then summed for a total expression level for each gene across cell types.Following this, the percentage to which each cell type contributed to the total expression was calculated.A gene was reported to have a speci c cell-type if the highest contributing cell type was 1.5× higher than the second highest contributing cell type.For example, if for Gene X, the highest contributing cell type, Cell-type A, was 45% of the total expression, and the next highest cell type, Cell-type B, was 30% or less of the total expression, then Gene X would be Cell-type A speci c.Protein in the SOMAscan7K platform were matched to their Entrez gene symbol using the SOMAlogic provided documentation and the ratio-based strategy above was used to determine cell-type speci city for all proteins.EOAD prioritized genes were subset from this to determine cell-type speci city.If cell-type speci city was unable to be determined by this method, the human proteome atlas (https://www.proteinatlas.org/) was also queried to determine if any cell type speci city information was available for a given gene.

Differential Expression in AD brain tissues
The Agora knowledge portal is a database powered by AMP-AD research which hosts evidence for AD relevant genes.The Agora gene comparison tool allows for query and simultaneous comparison of a set of genes' differential expression as measured by RNA or protein across nine brain regions: anterior cingulate cortex, cerebellum, dorsolateral prefrontal cortex, frontal pole, inferior frontal gyrus, posterior cingulate cortex, parahippocampal gyrus, superior temporal gyrus, and temporal cortex.The tool shows graphically whether any genes in the queried set are signi cantly upregulated or downregulated in AD diagnosed participants compared to controls.Additionally, the output also provides a genetic score, multi-omics score, and a risk score which sums the multi-omics and genetic score to allow for a numerical assessment of how much a given gene contributes to AD. Agora and AMP-AD methods are further described here https://help.adknowledgeportal.org/apd/AD-Risk-Scores-Data-and-Methods.2826043399.html.

Association with related phenotypes
We used online resources GeneCards 28,29 (https://www.genecards.org/), the Human Protein Atlas (https://www.proteinatlas.org) and the GWAS catalog 55 (https://www.ebi.ac.uk/gwas/) to search for our prioritized genes and understand what other phenotypes they are known to affect.Particularly we looked to if they were associated with or known to affect other neurological, dementia and neurodegenerative, cancer, and other health risk factors (e.g., BMI, cardiovascular risk factors, education attainment, etc.).

Biological pathways
Pathway analyses were performed with the enrichGO 78 , enrichDO 79 DisGeNet and enrichKEGG functions from the R package ClusterPro ler v4.09 80 as well as the Gene2Func tool from FUMA 72 and enrichment analyses from STRING 36 .In each case, the input set was all threshold passing genes (Total prioritization score ≥ 4 points) except those from the APOE region, and all genes were included as the background.

Protein-protein interactions
Protein-protein interactions were empirically investigated using STRING (https://string-db.org/)database to see if our prioritized genes (excluding the APOE region) were enriched for any interactions.Additionally, we performed proteome-wide association (PWAS), which integrates GWAS summary statistics and pQTL data to resolve causal protein in GWAS loci.PWAS was performed with FUSION 81 transcriptome-wide association study software.Variant weights were calculated for each protein and for each association separately for those aptamers with at least one study-wide pQTL.Protein levels were estimated based on these weights and correlated with EOAD using NHW summary statistic on chromosomes where genome wide signi cant EOAD GWAS loci were identi ed.Finally, we performed colocalization between EOAD and trans-pQTLs to see if there was potential evidence of co-regulation of genes in EOAD loci.

1
Figure 1 Project work ow and Meta-analysis This project analyzes three ancestries; non-Hispanic Whites (NHW), African Americans (AA), and Asians to identify genes and genomic loci associated with earlier onset Alzheimer disease following the work ow in panel A. Shown in panel B is the Manhattan plot of the trans-ancestry meta-analysis.Loci are annotated with the nearest gene to the top hit.The blue-horizontal line represents the suggestive p-value threshold of 1×10 -05 , and the red-horizontal line is the genome-wide signi cance threshold of 5×10 -08 .Novel loci are highlighted by red text.

Table 1
Summary of GWAS loci from single variant and meta analyses *This table summarizes the nearest gene, the prioritized gene, and summary statistics from each ancestry's single variant analysis, meta-analysis and LOAD f GWAS locus.Known AD loci are shown at the top of the table, novel loci are in the lower half of the table.LOAD_COLOC is the best PPH4 value from doing colocalization between EOAD and LOAD at EOAD loci, while LOAD_P is the best p value for the lead SNP in LOAD analyses.* = best is from Bellenguez et al., from Shigemizu et al., *** = best is from Kunkle et al.,