Genetic meta-analysis of diagnosed Alzheimer’s disease identifies new risk loci and implicates Aβ, tau, immunity and lipid processing

Abstract

Risk for late-onset Alzheimer’s disease (LOAD), the most prevalent dementia, is partially driven by genetics. To identify LOAD risk loci, we performed a large genome-wide association meta-analysis of clinically diagnosed LOAD (94,437 individuals). We confirm 20 previous LOAD risk loci and identify five new genome-wide loci (IQCK, ACE, ADAM10, ADAMTS1, and WWOX), two of which (ADAM10, ACE) were identified in a recent genome-wide association (GWAS)-by-familial-proxy of Alzheimer’s or dementia. Fine-mapping of the human leukocyte antigen (HLA) region confirms the neurological and immune-mediated disease haplotype HLA-DR15 as a risk factor for LOAD. Pathway analysis implicates immunity, lipid metabolism, tau binding proteins, and amyloid precursor protein (APP) metabolism, showing that genetic variants affecting APP and Aβ processing are associated not only with early-onset autosomal dominant Alzheimer’s disease but also with LOAD. Analyses of risk genes and pathways show enrichment for rare variants (P = 1.32 × 10⁻⁷), indicating that additional rare variants remain to be identified. We also identify important genetic correlations between LOAD and traits such as family history of dementia and education.

Our previous work identified 19 genome-wide-significant common variant signals in addition to APOE that influence risk for LOAD (onset age > 65 years)¹. These signals, combined with ‘subthreshold’ common variant associations, account for ~31% of the genetic variance of LOAD², leaving the majority of genetic risk uncharacterized³. To search for additional signals, we conducted a GWAS meta-analysis of non-Hispanic Whites (NHW) by using a larger Stage 1 discovery sample (17 new, 46 total datasets; n = 21,982 cases, 41,944 cognitively normal controls) from our group, the International Genomics of Alzheimer’s Project (IGAP) (composed of four consortia: Alzheimer Disease Genetics Consortium (ADGC), Cohorts for Heart and Aging Research in Genomic Epidemiology Consortium (CHARGE), The European Alzheimer’s Disease Initiative (EADI), and Genetic and Environmental Risk in AD/Defining Genetic, Polygenic and Environmental Risk for Alzheimer’s Disease Consortium (GERAD/ PERADES) (Supplementary Tables 1 and 2, and Supplementary Note). To sample both common and rare variants (minor allele frequency (MAF) ≥ 0.01 and MAF < 0.01, respectively), we imputed the discovery datasets by using a 1,000 Genomes reference panel consisting of 36,648,992 single-nucleotide polymorphisms (SNPs), 1,380,736 insertions/deletions, and 13,805 structural variants. After quality control, 9,456,058 common variants and 2,024,574 rare variants were selected for analysis. Genotype dosages were analyzed within each dataset, and then combined with meta-analysis (Supplementary Fig. 1 and Supplementary Tables 1–3).

Results

Meta-analysis of Alzheimer’s disease GWAS.

The Stage 1 discovery meta-analysis produced 12 loci with genome-wide significance (P ≤ 5 × 10⁻⁸) (Table 1), all of which are previously described^1,4–11. Genomic inflation factors (λ) were slightly inflated (λ median = 1.05; λ regression = 1.09; see Supplementary Figure 2 for a quantile–quantile (QQ) plot); however, univariate linkage disequilibrium score (LDSC) regression^12,13 estimates indicated that the majority of this inflation was due to a polygenic signal, with the intercept being close to 1 (1.026, s.e.m. = 0.006). The observed heritability (h²) of LOAD was estimated at 0.071 (0.011) using LDSC. Stage 1 meta-analysis was first followed by Stage 2, using the I-select chip we previously developed in Lambert et al.¹ (including 11,632 variants, n = 18,845; Supplementary Table 4) and finally Stage 3A (n = 11,666) or Stage 3B (n = 30,511) (for variants in regions not well captured in the I-select chip) (see Supplementary Figure 1 for the workflow). The final sample was 35,274 clinical and autopsy-documented Alzheimer’s disease cases and 59,163 controls.

Table 1 |.

Summary of discovery Stage 1, Stage 2 and overall meta-analyses results for identified loci reaching genome-wide significance after Stages 1 and 2

						Stage 1 discovery (n = 63,926)			Stage 2 (n = 18,845)			Overall stage 1 + stage 2 (n = 82,771)
Variant a	chr.	Positionb	Closest genec	Major/minor alleles	MAFd	OR	95% CIe	P	OR	95% CIe	P	OR	95% CIe	Meta P	I2 (%), Pf
Previous genome-wide-significant loci still reaching significance
rs4844610	1	207802552	CR1	C/A	0.187	1.16	1.12–1.20	8.2 × 10−16	1.20	1.13–1.27	3.8 × 10−10	1.17	1.13–1.21	3.6 × 10−24	0, 8 × 10−1
rs6733839	2	127892810	BIN1	C/T	0.407	1.18	1.15–1.22	4.0 × 10−28	1.23	1.18–1.29	2.0 × 10−18	1.20	1.17–1.23	2.1 × 10−44	15, 2 × 10−1
rs10933431	2	233981912	INPP5D	C/G	0.223	0.90	0.87–0.94	2.6 × 10−7	0.92	0.87–0.97	3.2 × 10−3	0.91	0.88–0.94	3.4 × 10−9	0, 8 × 10−1
rs9271058	6	32575406	HLA-DRB1	T/A	0.270	1.10	1.06–1.14	5.1 × 10−8	1.11	1.06–1.17	5.7 × 10−5	1.10	1.07–1.13	1.4 × 10−11	10, 3 × 10−1
rs75932628	6	41129252	TREM2	C/T	0.008	2.01	1.65–2.44	2.9 × 10−12	2.50	1.56–4.00	1.5 × 10−4	2.08	1.73–2.49	2.7 × 10−15	0, 6 × 10−1
rs9473117	6	47431284	CD2AP	A/C	0.280	1.09	1.05–1.12	2.3 × 10−7	1.11	1.05–1.16	1.0 × 10−4	1.09	1.06–1.12	1.2 × 10−10	0, 6 × 10−1
rs12539172	7	100091795	NYAP1g	C/T	0.303	0.93	0.91–0.96	2.1 × 10−5	0.89	0.84–0.93	2.1 × 10−6	0.92	0.90–0.95	9.3 × 10−10	0, 8 × 10−1
rs10808026	7	143099133	EPHA1	C/A	0.199	0.90	0.87–0.94	3.1 × 10−8	0.91	0.86–0.96	1.1 × 10−3	0.90	0.88–0.93	1.3 × 10−10	0, 5 × 10−1
rs73223431	8	27219987	PTK2B	C/T	0.367	1.10	1.07–1.13	8.3 × 10−10	1.11	1.06–1.16	1.5 × 10−5	1.10	1.07–1.13	6.3 × 10−14	0, 6 × 10−1
rs9331896	8	27467686	CLU	T/C	0.387	0.88	0.85–0.91	3.6 × 10−16	0.87	0.83–0.91	1.7 × 10−9	0.88	0.85–0.90	4.6 × 10−24	3, 4 × 10−1
rs3740688	11	47380340	SPI1h	T/G	0.448	0.91	0.89–0.94	9.7 × 10−11	0.93	0.88–0.97	1.2 × 10−3	0.92	0.89–0.94	5.4 × 10−13	4, 4 × 10−1
rs7933202	11	59936926	MS4A2	A/C	0.391	0.89	0.86–0.92	2.2 × 10−15	0.90	0.86–0.95	1.6 × 10−5	0.89	0.87–0.92	1.9 × 10−19	27, 5 × 10−2
rs3851179	11	85868640	PICALM	C/T	0.356	0.89	0.86–0.91	5.8 × 10−16	0.85	0.81–0.89	6.1 × 10−11	0.88	0.86–0.90	6.0 × 10−25	0, 8 × 10−1
rs11218343	11	121435587	SORL1	T/C	0.040	0.81	0.76–0.88	2.7 × 10−8	0.77	0.68–0.87	1.8 × 10−5	0.80	0.75–0.85	2.9 × 10−12	7, 3 × 10−1
rs17125924	14	53391680	FERMT2	A/G	0.093	1.13	1.08–1.19	6.6 × 10−7	1.15	1.06–1.25	5.0 × 10−4	1.14	1.09–1.18	1.4 × 10−9	8, 3 × 10−1
rs12881735	14	92932828	SLC24A4	T/C	0.221	0.92	0.88–0.95	4.9 × 10−7	0.92	0.87–0.97	4.3 × 10−3	0.92	0.89–0.94	7.4× 10−9	0, 6 × 10−1
rs3752246	19	1056492	ABCA7	C/G	0.182	1.13	1.09–1.18	6.6 × 10−10	1.18	1.11–1.25	4.7 × 10−8	1.15	1.11–1.18	3.1 × 10−16	0, 5 × 10−1
rs429358	19	45411941	APOE	T/C	0.216	3.32	3.20–3.45	1.2 × 10−881	APOE region not carried forward to replication stage
rs6024870	20	54997568	CASS4	G/A	0.088	0.88	0.84–0.93	1.1 × 10−6	0.90	0.82–0.97	9.0 × 10−3	0.88	0.85–0.92	3.5 × 10−8	0, 9 × 10−1
New genome-wide-significant loci reaching significance
rs7920721	10	11720308	ECHDC3	A/G	0.389	1.08	1.05–1.11	1.9 × 10−7	1.07	1.02–1.12	3.2 × 10−3	1.08	1.05–1.11	2.3 × 10−9	0,8 × 10−1
rs138190086	17	61538148	ACE	G/A	0.020	1.29	1.15–1.44	7.5 × 10−6	1.41	1.18–1.69	1.8 × 10−4	1.32	1.20–1.45	7.5 × 10−9	0, 9 × 10−1
Previous genome-wide-significant loci not reaching significance
rs190982	5	88223420	MEF2C	A/G	0.390	0.95	0.92–0.97	2.8 × 10−4	0.93	0.89–0.98	2.7 × 10−3	0.94	0.92–0.97	2.8 × 10−6	0, 6 × 10−1
rs4723711	7	37844263	NME8	A/T	0.356	0.95	0.92–0.98	2.7 × 10−4	0.91	0.87–0.95	1.0 × 10−4	0.94	0.91–0.96	2.8 × 10−7	0, 5 × 10−1

^aVariants showing the best level of association after meta-analysis of Stages 1 and 2.

^bBuild 37, assembly hg19.

^cBased on position of top SNP in reference to the RefSeq assembly.

^dAverage in the discovery sample.

^eCalculated with respect to the minor allele.

^fCochran’s Q test

^gPreviously the ZCWPW1 locus.

^hPreviously the CELF1 locus. Chr., chromosome; CI, confidence interval; OR, odds ratio; I², heterogeneity estimate.

Meta-analysis of Stages 1 and 2 produced 21 genome-wide-significant associations (P ≤ 5 × 10⁻⁸) (Table 1 and Fig. 1), 18 of which were previously reported as genome-wide significant in Lambert et al.¹. Three other signals were not initially described in the initial IGAP GWAS: the rare R47H TREM2 coding variant previously reported by others^7,8,14; ECHDC3 (rs7920721; NC_000010.10: g.11720308A>G), which was recently identified as a potential genome-wide-significant Alzheimer’s disease risk locus in several studies^15–17, and ACE (rs138190086; NC_000017.10: g.61538148G>A) (Supplementary Figs. 3 and 4). In addition, seven signals showed suggestive association with P < 5 × 10⁻⁷ (closest genes: ADAM10, ADAMTS1, ADAMTS20, IQCK, MIR142/ TSPOAP1-AS1, NDUFAF6, and SPPL2A) (Supplementary Figs. 5–11). Stage 3A and meta-analysis of all three stages for these nine signals (excluding the TREM2 signal; see Supplementary Table 5 for the variant list) identified five genome-wide-significant loci. In addition to ECHDC3, this included four new genome-wide Alzheimer’s disease risk signals at IQCK, ADAMTS1, ACE, and ADAM10 (Table 2). ACE and ADAM10 were previously reported as Alzheimer’s disease candidate genes^18–22 but were not replicated in some subsequent studies^23–25. A recent GWAS using family history of Alzheimer’s disease or dementia as a proxy²⁶ also identified these two risk loci, suggesting that while use of proxy Alzheimer’s disease/dementia cases introduces less sensitivity and specificity for true Alzheimer’s disease signals overall in comparison to clinically diagnosed Alzheimer’s disease, proxy studies can identify disease-relevant associations. Two of the four other signals approached genome-wide significance: miR142/TSPOAP1-AS1 (P = 5.3 × 10⁻⁸) and NDUFAF6 (P = 9.2 × 10⁻⁸) (Table 2). Stage 3A also extended the analysis of two loci (NME8 and MEF2C) that were previously genome-wide significant in our 2013 meta-analysis. These loci were not genome-wide significant in our current study and will deserve further investigation (NME8: P = 2.7 × 10⁻⁷; MEF2C: P = 9.1 × 10⁻⁸; Supplementary Figs. 12 and 13). Of note, GCTA COJO²⁷ conditional analysis of the genome-wide loci indicates that TREM2 and three other loci (BIN1, ABCA7, and PTK2B/CLU) have multiple independent LOAD association signals (Supplementary Table 6), suggesting that the genetic variance associated with some GWAS loci is probably underestimated.

Fig. 1 |. Manhattan plot of meta-analysis of stage 1, 2, and 3 results for genome-wide association with Alzheimer’s disease.

The threshold for genomewide significance (P < 5 × 10⁻⁸) is indicated by the red line, while the blue line represents the suggestive threshold (P < 1 × 10⁻⁵). Loci previously identified by the Lambert et al.¹ IGAP GWAS are shown in blue and newly associated loci are shown in red. Loci are named for the closest gene to the sentinel variant for each locus. Diamonds represent variants with the smallest P values for each genome-wide locus.

Table 2 |.

Summary of discovery Stage 1, Stage 2, Stage 3 (A and B), and overall meta-analysis results of potential novel loci

Stage 3A						Stage 1 + 2 (n = 82,771)			Stage 3A (n = 11,666)			Overall (n = 94,437)
SNPa	Chr.	Positionb	Closest genec	Major/ minor allele	MAFe	OR	95% CIf	P	OR	95% CIf	P	OR	95% CIf	Meta P
rs4735340	8	95976251	NDUFAF6	T/A	0.476	0.94	0.92–0.96	3.4 × 10−7	0.92	0.83–1.02	9.7 × 10−2	0.94	0.92–0.96	9.2 × 10−8
rs7920721g	10	11720308	ECHDC3	A/G	0.390	1.08	1.05–1.11	2.3 × 10−9	1.11	1.04–1.18	1.5 × 10−3	1.08	1.06–1.11	1.8 × 10−11
rs7295246	12	43967677	ADAMTS20	T/G	0.413	1.07	1.04–1.09	2.7 × 10−7	1.02	0.96–1.09	4.5 × 10−1	1.06	1.04–1.08	3.9 × 10−7
rs10467994	15	51008687	SPPL2A	T/C	0.333	0.94	0.91–0.96	3.9 × 10−7	0.97	0.87–1.08	6.2 × 10−1	0.94	0.92–0.96	4.3 × 10−7
rs593742	15	59045774	ADAM10	A/G	0.295	0.93	0.91–0.96	1.3 × 10−7	0.91	0.85–0.98	1.5 × 10−2	0.93	0.91–0.95	6.8 × 10−9
rs7185636	16	19808163	IQCK	T/C	0.180	0.92	0.89–0.95	8.4 × 10−8	0.94	0.86–1.01	1.1 × 10−1	0.92	0.89–0.95	2.4 × 10−8
rs2632516	17	56409089	MIR142/TSPOAP1-AS1d	G/C	0.440	0.94	0.92–0.96	2.3 × 10−7	0.91	0.82–1.01	7.5 × 10−2	0.94	0.91–0.96	5.3 × 10−8
rs138190086	17	61538148	ACE	G/A	0.020	1.32	1.20–1.45	7.5 × 10−9	1.17	0.92–1.48	2.1 × 10−1	1.30	1.19–1.42	5.3 × 10−9
rs2830500	21	28156856	ADAMTS1	C/A	0.308	0.93	0.91–0.96	7.4 × 10−8	0.95	0.89–1.02	1.3 × 10−1	0.93	0.91–0.96	2.6 × 10−8
Stage 3B						Stage 1 (n = 63,926)			Stage 3B (n = 30,511)h			Overall (n = 94,437)h
SNPa	Chr.	Positionb	Closest genec	Major/ minor allele	MAFe	OR	95% CIf	P	OR	95% CIf	P	OR	95% CIf	Meta P
rs71618613	5	29005985	SUCLG2P4	A/C	0.010	0.68	0.57–0.80	9.8 × 10−6	0.76	0.63–0.93	6.8 × 10−3	0.71	0.63–0.81	3.3 × 10−7
rs35868327	5	52665230	FST T/A	A/C	0.013	0.69	0.59–0.80	7.8 × 10−7	0.58	0.29–1.17	1.2 × 10−1	0.68	0.59–0.79	2.6 × 10−7
rs114812713	6	41034000	OARD1	G/C	0.030	1.35	1.24–1.47	4.5 × 10−12	1.23	1.06–1.42	7.2 × 10−3	1.32	1.22–1.42	2.1 × 10−13
rs62039712	16	79355857	WWOX	G/A	0.116	1.17	1.10–1.23	1.2 × 10−7	1.14	0.96–1.36	1.3 × 10−1	1.16	1.10–1.23	3.7 × 10−8

Novel loci were defined as loci not reported in Lambert et al.¹ with (1) a Stage 1 + 2 meta P < 5 × 10⁻⁷ (nine variants after excluding TREM2) (Stage 3A) or (2) a MAF < 0.05 and Stage 1 P < 1 × 10⁻⁵ or MAF ≥ 0.05 and Stage 1 P < 5 × 10⁻⁶ for genome regions not covered on the Stage 2 custom array (Stage 3B).

^aSNPs showing the best level of association after meta-analysis of Stages 1, 2 and 3.

^bBuild 37, assembly hg19.

^cBased on position of top SNP in reference to the RefSeq assembly.

^dVariant is annotated to both gene features.

^eAverage in the discovery sample.

^fCalculated with respect to the minor allele.

^gRecently identified as a LOAD locus in two separate 2017 studies.

^hSample sizes for these loci are smaller (overall n = 89,769 for SUCLG2P4, 65,230 for FST, and n = 69,898 for WWOX).

We also selected 33 variants from Stage 1 (28 common and 5 rare variants in loci not well captured in the I-select chip; see Methods for full selection criteria) for genotyping in Stage 3B (including populations of Stage 2 and Stage 3A). We nominally replicated a rare variant (rs71618613; NC_000005.9: g.29005985A>C) within an intergenic region near SUCLG2P4 (MAF = 0.01; P = 6.8 × 10⁻³; combined P = 3.3 × 10⁻⁷) and replicated a low-frequency variant in the TREM2 region (rs114812713; NC_000006.11: g.41034000G>C, MAF = 0.03, P = 7.2 × 10⁻³; combined P = 2.1 × 10⁻¹³) in the gene OARD1 that may represent an independent signal according to our conditional analysis (Table 2, Supplementary Figs. 14 and 15, Supplementary Tables 6 and 7). In addition, rs62039712 (NC_000016.9: g.79355857G>A) in the WWOX locus reached genome-wide significance (P = 3.7 × 10⁻⁸), and rs35868327 (NC_000005.9: g.52665230T>A) in the FST locus reached suggestive significance (P = 2.6 × 10⁻⁷) (Table 2 and Supplementary Figs. 16 and 17). WWOX may play a role in Alzheimer’s disease through its interaction with tau^28,29, and it is worth noting that the sentinel variant (defined as the variant with the lowest P value) is just 2.4 mega-bases from PLCG2, which contains a rare variant that we recently associated with Alzheimer’s disease¹⁴. Since both rs62039712 and rs35868327 were only analyzed in a restricted number of samples, these loci deserve further attention.

Candidate gene prioritization at genome-wide loci.

To evaluate the biological significance and attempt to identify the underlying risk genes for the newly identified genome-wide signals (IQCK, ACE, ADAM10, ADAMTS1, and WWOX) and those found previously, we pursued five strategies: (1) annotation and gene-based testing for deleterious coding, loss-of-function (LOF) and splicing variants; (2) expression-quantitative trait loci (eQTL) analyses; (3) evaluation of transcriptomic expression in LOAD clinical traits (correlation with the BRAAK stage³⁰ and differential expression in Alzheimer’s disease versus control brains³¹); (4) evaluation of transcriptomic expression in Alzheimer’s disease–relevant tissues^32–34; and (5) gene cluster/pathway analyses. For the 24 signals reported here, other evidence indicates that APOE^35,36, ABCA7 (refs. ^37–40), BIN1 (ref. ⁴¹), TREM2 (refs. ^7,8), SORL1 (refs. ^42,43), ADAM10 (ref. ⁴⁴), SPI1 (ref. ⁴⁵), and CR1 (ref. ⁴⁶) are the true Alzheimer’s disease risk gene, although there is a possibility that multiple risk genes exist in these regions⁴⁷. Because many GWAS loci are intergenic, and the closest gene to the sentinel variant may not be the actual risk gene, in these analyses we considered all protein-coding genes within ±500 kilobases (kb) of the sentinel variant linkage disequilibrium (LD) regions (r² ≥ 0.5) for each locus as a candidate Alzheimer’s disease gene (n = 400 genes) (Supplementary Table 8).

We first annotated all sentinel variants for each locus and variants in LD (r² > 0.7) with these variants in a search for deleterious coding, LOF or splicing variants. In line with findings that most causal variants for complex disease are non-coding⁴⁸, only 2% of 1,073 variants across the 24 loci (excluding APOE) were exonic variants, with a majority (58%) being intronic (Supplementary Fig. 18 and Supplementary Table 9). Potentially deleterious variants include the rare R47H missense variant in TREM2, common missense variants in CR1, SPI1, MS4A2, and IQCK, and a relatively common (MAF = 0.16) splicing variant in IQCK. Using results of a large whole-exome-sequencing study conducted in the ADGC and CHARGE sample⁴⁹ (n = 5,740 LOAD cases and 5,096 controls), we also identified ten genes located in our genome-wide loci as having rare deleterious coding, splicing or LOF burden associations with LOAD (false discovery rate (FDR) P < 0.01), including previously implicated rare-variant signals in ABCA7, TREM2, and SORL1 (refs. ^14,49–55), and additional associations with TREML4 in the TREM2 locus, TAP2 and PSMB8 in the HLA-DRB1 locus, PIP in the EPHA1 locus, STYX in the FERMT2 locus, RIN3 in the SLC24A4 locus, and KCNH6 in the ACE locus (Supplementary Table 10).

For eQTL analyses, we searched existing eQTL databases and studies for cis-acting eQTLs in a prioritized set of variants (n = 1,873) with suggestive significance or in LD with the sentinel variant in each locus. Of these variants, 71–99% have regulatory potential when considering all tissues according to RegulomeDB⁵⁶ and HaploReg⁵⁷, but restricting to Alzheimer’s disease–relevant tissues (via Ensembl Regulatory Build⁵⁸ and GWAS4D⁵⁹) appears to aid in regulatory variant prioritization, with probabilities for functional variants increasing substantially when using GWAS4D cell-dependent analyses with brain or monocytes, for instance (these and other annotations are provided in Supplementary Table 11). Focusing specifically on eQTLs, we found overlapping cis-acting eQTLs for 153 of the 400 protein-coding genes, with 136 eQTL-controlled genes in Alzheimer’s disease–relevant tissues (that is, brain and blood/immune cell types; see Methods for details) (Supplementary Tables 12 and 13). For our newly identified loci, there were significant eQTLs in Alzheimer’s disease–relevant tissue for ADAM10, FAM63B, and SLTM (in the ADAM10 locus); ADAMTS1 (ADAMTS1 locus); and ACSM1, ANKS4B, C16orf62, GDE1, GPRC5B, IQCK, and KNOP1 (IQCK locus). There were no eQTLs in Alzheimer’s disease–relevant tissues in the WWOX or ACE locus, although several eQTLs for PSMC5 in coronary artery tissue were found for the ACE locus. eQTLs for genes in previously identified loci include BIN1 (BIN1 locus), INPP5D (INPP5D locus), CD2AP (CD2AP locus), and SLC24A4 (SLC24A4 locus). Co-localization analysis confirmed evidence of a shared causal variant affecting expression and disease risk in 66 genes over 20 loci, including 31 genes over 13 loci in LOAD-relevant tissue (see Supplementary Table 14 and 15 for complete lists). Genes implicated include CR1 (CR1 locus), ABCA7 (ABCA7 loci), BIN1 (BIN1 locus), SPI1 and MYBPC3 (SPI1 locus), MS4A2, MS4A6A, and MS4A4A (MS4A2 locus), KNOP1 (IQCK locus), and HLA-DRB1 (HLA-DRB1 locus) (Supplementary Table 12).

To study the differential expression of genes in brains of patients with Alzheimer’s disease versus controls, we used 13 expression studies³¹. We found that 58% of the 400 protein-coding genes within the genome-wide loci had evidence of differential expression in at least one study (Supplementary Table 16). Additional comparisons to Alzheimer’s disease related gene expression sets revealed that 62 genes were correlated with pathogenic stage (BRAAK) in at least one brain tissue³⁰ (44 genes in prefrontal cortex, the most relevant LOAD tissue; 36 in cerebellum and 1 in visual cortex). Finally, 38 genes were present in a set of 1,054 genes preferentially expressed in aged microglial cells, a gene set shown to be enriched for Alzheimer’s disease genes (P = 4.1 × 10⁻⁵)³⁴. We also annotated our list of genes with brain RNA-seq data, which showed that 80% were expressed in at least one type of brain cell, and the genes were most highly expressed in fetal astrocytes (26%), followed by microglia/ macrophages (15.8%), neurons (14.8%), astrocytes (11.5%), and oligodendrocytes (6.5%). When not considering fetal astrocytes, mature astrocytes (21%), and microglial cells (20.3%), the resident macrophage cells of the brain thought to play a key role in the pathologic immune response in LOAD^8,14,60, became the highest expressed cell types in the genome-wide set of genes, with 5.3% of the 400 genes showing high microglial expression (Supplementary Table 17; see Supplementary Table 18 for the highly expressed gene list by cell type).

We conducted pathway analyses (MAGMA⁶¹) separately for common (MAF > 0.01) and rare variants (MAF < 0.01). For common variants, we detected four function clusters including (1) APP metabolism/Aβ formation (regulation of Aβ formation: P = 4.56 × 10⁻⁷ and regulation of APP catabolic process: P = 3.54 × 10⁻⁶); (2) tau protein binding (P = 3.19 × 10⁻⁵); (3) lipid metabolism (four pathways including protein−lipid complex assembly: P = 1.45 × 10⁻⁷); and (4) immune response (P = 6.32 × 10⁻⁵) (Table 3 and Supplementary Table 19). Enrichment of the four clusters remained after removal of genes in the APOE region. When APOE-region genes and genes near genome-wide-significant genes were removed, tau showed moderate association (P = 0.027), and lipid metabolism and immune-related pathways showed strong associations (P < 0.001) (Supplementary Table 20). Genes driving these enrichments (that is, having a gene-wide P < 0.05) included SCNA, a Parkinson’s risk gene that encodes alpha-synuclein, the main component of Lewy bodies, whch may play a role in tauopathies^62,63, for the tau pathway; apolipoprotein genes (APOM, APOA5) and ABCA1, a major regulator of cellular cholesterol, for the lipid metabolism pathways; and 52 immune pathway genes (Supplementary Table 21). While no pathways were significantly enriched for rare variants, lipid and Aβ pathways did reach nominal significance in rare-variant-only analyses. Importantly, we also observed a highly significant correlation between common and rare pathway gene results (P = 1.32 × 10⁻⁷), suggesting that risk Alzheimer’s disease genes and pathways are enriched for rare variants. In fact, 50 different genes within tau, lipid, immunity and Aβ pathways showed nominal rare-variant driven associations (P < 0.05) with LOAD.

Table 3 |.

Significant pathways (q value ≤ 0.05) from MAGMA pathway analysis for common and rare variant subsets

Pathway	No. of genes in the pathway in the dataset	Common variant Pa	Common variant q value	Rare variant Pa	Rare variant q value	Pathway description
GO:65005	20	1.4 × 10−7a	9.5 × 10−4	6.7 × 10−2	8.4 × 10−1	Protein−lipid complex assembly
GO:1902003	10	4.5 × 10−7a	1.4 × 10−3	4.9 × 10−2	8.4 × 10−1	Regulation of Aβ formation
GO:32994	39	1.1 × 10−6a	2.5 × 10−3	1.7 × 10−2	8.1 × 10−1	Protein−lipid complex
GO:1902991	12	3.5 × 10−6a	5.8 × 10−3	5.6 × 10−2	8.4 × 10−1	Regulation of amyloid precursor protein catabolic process
GO:43691	17	5.5 × 10−6a	6.7 × 10−3	3.0 × 10−2	8.1 × 10−1	Reverse cholesterol transport
GO:71825	35	6.1 × 10−6a	6.7 × 10−3	1.2 × 10−1	8.4 × 10−1	Protein−lipid complex subunit organization
GO:34377	18	1.6 × 10−5a	1.5 × 10−2	1.8 × 10−1	8.4 × 10−1	Plasma lipoprotein particle assembly
GO:48156	10	3.1 × 10−5a	2.6 × 10−2	7.7 × 10−1	8.5 × 10−1	Tau protein binding
GO:2253	382	6.3 × 10−5a	4.6 × 10−2	2.0 × 10−1	8.4 × 10−1	Activation of immune response

^aSignificant after FDR correction (q value ≤ 0.05).

To further explore the APP/Aβ pathway enrichment, we analyzed a comprehensive set of 335 APP metabolism genes⁶⁴ curated from the literature. We observed significant enrichment of this gene set in common variants (P = 2.27 × 10⁻⁴; P = 3.19 × 10⁻⁴ excluding APOE), with both ADAM10 and ACE nominally significant drivers of this result (Table 4 and Supplementary Tables 22 and 23). Several ‘sub-pathways’ were also significantly enriched in the common variants, including ‘clearance and degradation of Aβ’, and ‘aggregation of Aβ’, along with its subcategory ‘microglia’, the latter supporting microglial cells suspected role in response to Aβ in LOAD⁶⁵. Nominal enrichment for risk from rare variants was found for the pathway ‘aggregation of Aβ: chaperone’ and 23 of the 335 genes.

Table 4 |.

Top results of pathway analysis of the Aβ-centered biological network from Campion et al.⁶⁴ (see Supplementary Table 12 for full results)

Category	Subcategory	No. of genes	Common variant P 0 kb	Common variant P 35–10 kb	Rare variant P 0 kb	Rare variant P 35–10 kb
Aβ-centered biological network (all genes)	–	331	2.2 × 10−4a	1.5 × 10−4a	8.2 × 10−1	5.1 × 10−1
Clearance and degradation of Aβ	–	74	2.1 × 10−4a	3.2 × 10−3	3.1 × 10−1	5.1 × 10−1
Clearance and degradation of Aβ	Microglia	47	2.2 × 10−4a	1.8 × 10−2	2.4 × 10−1	6.8 × 10−1
Aggregation of Aβ	–	35	7.0 × 10−4a	9.9 × 10−3	9.0 × 10−2	1.6 × 10−1
Aggregation of Aβ	Miscellaneous	21	1.0 × 10−3a	3.3 × 10−2	9.5 × 10−2	1.9 × 10−1
APP processing and trafficking	Clathrin/caveolin-dependent endocytosis	10	1.1 × 10−3	1.1 × 10−2	3.6 × 10−1	1.8 × 10−1
Mediator of Aβ toxicity	–	51	3.8 × 10−2	4.6 × 10−2	5.8 × 10−1	5.7 × 10−1
Mediator of Aβ toxicity	Calcium homeostasis	6	6.9 × 10−2	1.2 × 10−1	3.9 × 10−1	2.5 × 10−1
Mediator of Aβ toxicity	Miscellaneous	3	7.6 × 10−2	2.3 × 10−2	9.7 × 10−1	7.6 × 10−1
Clearance and degradation of Aβ	Enzymatic degradation of Aβ	15	7.7 × 10−2	2.6 × 10−2	6.1 × 10−1	2.9 × 10−1
Mediator of Aβ toxicity	Tau toxicity	20	9.0 × 10−2	3.4 × 10−1	7.1 × 10−1	6.8 × 10−1
Aggregation of Aβ	Chaperone	9	1.5 × 10−1	3.0 × 10−1	1.9 × 10−1	1.1 × 10−2

^aSignificant after Bonferroni correction for 33 pathway sets tested.

To identify candidate genes for our novel loci, we combined results from our five prioritization strategies in a priority ranking method similar to that of Fritsche et al.⁶⁶ (Fig. 2 and Supplementary Table 24). ADAM10 was the top ranked gene of the 11 genes within the ADAM10 locus. ADAM10, the most important α-secretase in the brain, is a component of the non-amyloidogenic pathway of APP metabolism⁶⁷ and sheds TREM2 (ref. ⁶⁸), an innate immunity receptor expressed selectively in microglia. Overexpression of ADAM10 in mouse models can halt Aβ production and subsequent aggregation⁶⁹. In addition, two rare ADAM10 alterations segregating with disease in LOAD families increased Aβ plaque load in ‘Alzheimer-like’ mice, with diminished α-secretase activity from the alterations probably the causal mechanism^19,44. For the IQCK signal, which is also an obesity locus^70,71, IQCK, a relatively uncharacterized gene, was ranked top, although four of the other 11 genes in the locus have a priority rank ≥ 4, including KNOP1 and GPRC5B, the latter being a regulator of neurogenesis^72,73 and inflammatory signaling in obesity⁷⁴. Of the 22 genes in the ACE locus, PSMC5, a key regulator of major histocompatibility complex (MHC)^75,76, has a top score of 4, while DDX42, MAP3K3, an important regulator of macrophages and innate immunity^77,78, and CD79B, a B lymphocyte antigen receptor subunit, each have a score of 3. Candidate gene studies have associated ACE variants with Alzheimer’s disease risk^20,22,79, including a strong association in the Wadi Ara, an Israeli Arab community with high risk of Alzheimer’s disease²¹. However, these studies yielded inconsistent results²³, and our work reports a clear genome-wide association in NHW at this locus. While ACE was not prioritized, it should not be rejected as a candidate gene, as its expression in Alzheimer’s disease brain tissue is associated with Aβ load and Alzheimer’s disease severity⁸⁰. Furthermore, cerebrospinal fluid (CSF) levels of the angiotensin-converting enzyme (ACE) are associated with Aβ levels⁸¹ and LOAD risk⁸², and studies show ACE can inhibit Aβ toxicity and aggregation⁸³. Finally, angiotensin II, a product of ACE function, mediates a number of neuropathological processes in Alzheimer’s disease⁸⁴ and is now a target for intervention in phase II clinical trials of Alzheimer’s disease⁸⁵. Another novel genome-wide locus reported here, ADAMTS1, is within 665 kb of APP on chromosome 21. Of three genes at this locus, our analyses nominate ADAMTS1 as the likely risk gene, although we cannot rule out that this signal is a regulatory element for APP. ADAMTS1 is elevated in Down’s syndrome with neurodegeneration and Alzheimer’s disease⁸⁶, and it is a potential neuroprotective gene^87–89 or a neuroinflammatory gene important to microglial response⁹⁰. Finally, WWOX and MAF, which surround an intergenic signal in an obesity associated locus⁹¹, were both prioritized for the WWOX locus, with MAF, another important regulator of macrophages^92,93, being highly expressed in microglia in the Brain RNA-seq database, and WWOX, a high-density-lipoprotein cholesterol and triglyceride–associated gene^94,95, being expressed most highly in astrocytes and neurons. WWOX has been implicated in several neurological phenotypes⁹⁶; in addition, it binds tau and may play a critical role in regulating tau hyper-phosphorylation, neurofibrillary formation and Aβ aggregation^28,29. Intriguingly, treatment of mice with its binding partner restores memory deficits⁹⁷, hinting at its potential in neurotherapy.

Fig. 2 |. Top prioritized genes of 400 genes located in genome-wide-significant loci.

The criteria include: (1) deleterious coding, LOF or splicing variant in the gene; (2) significant gene-based tests; (3) expression in a tissue relevant to Alzheimer’s disease (astrocytes, neurons, microglia/macrophages, oligodendrocytes); (4) a HuMi microglial-enriched gene; (5) having an eQTL effect on the gene in any tissue, in Alzheimer’s disease–relevant tissue, and/ or a co-localized eQTL; (6) being involved in a biological pathway enriched in Alzheimer’s disease (from the current study); (7) expression correlated with the BRAAK stage; and (8) differential expression in a 1 + Alzheimer’s disease (AD) study. Novel genome-wide loci from the current study are listed first, followed by known genome-wide loci. Each category is assigned an equal weight of 1, with the priority score equaling the sum of all categories. Colored fields indicate that the gene meets the criteria. Genes with a priority score ≥ 4 are listed for each locus. If no gene reached a score of ≥ 5 in a locus, then the top ranked gene(s) is listed.

For previously reported loci, applying the same prioritization approach highlights several genes, as described in Fig. 2, some of which are involved in APP metabolism (FERMT2, PICALM) or tau toxicity (BIN1, CD2AP, FERMT2, CASS4, PTK2B)^98–101. Pathway, tissue and disease trait enrichment analyses support the utility of our prioritization method, as the 53 prioritized genes with a score ≥ 5 are (1) enriched in substantially more Alzheimer’s disease–relevant pathways, processes and dementia-related traits; (2) enriched in candidate Alzheimer’s disease cell types such as monocytes (adjusted P = 9.0 × 10⁻⁶) and macrophages (adjusted P = 5.6 × 10⁻³); and (3) more strongly associated with dementia-related traits and Alzheimer’s disease–relevant pathways (Supplementary Table 25 and 26; see Supplementary Fig. 19 for the interaction network of these prioritized genes). To further investigate the cell types and tissues the prioritized genes are expressed in, we performed differentially expressed gene (DEG) set enrichment analysis of the prioritized genes by using GTEx¹⁰² tissues, and we identified significant differential expression in several potentially relevant Alzheimer’s disease tissues including immune-related tissues (upregulation in blood and spleen), obesity-related tissue (upregulation in adipose), heart tissues (upregulation in left ventricle and atrial appendage), and brain tissues (dowregulation in cortex, cerebellum, hippo-campus, basal ganglia, and amygdala). Furthermore, the 53 genes are overexpressed in ‘adolescence’ and ‘young adult’ brain tissues in BrainSpan¹⁰³, a transcriptomic atlas of the developing human brain, which is consistent with accumulating evidence suggesting Alzheimer’s disease may start decades before the onset of disease^104,105 (Supplementary Fig. 20; see Supplementary Fig. 21 for a tissue expression heat map for the 53 genes).

Fine-mapping of the HLA region.

The above approach prioritized HLA-DRB1 as the top candidate gene in the MHC locus, known for its complex genetic organization and highly polymorphic nature (see Supplementary Fig. 22 for a plot of the region of the Stage 1 results). Previous analyses in the ADGC (5,728 Alzheimer’s disease cases and 5,653 controls) have linked both HLA class I and II haplotypes with Alzheimer’s disease risk¹⁰⁶. In order to further investigate this locus in a much larger sample, we used a robust imputation method and fine-mapping association analysis of alleles and haplotypes of HLA class I and II genes in 14,776 cases and 23,047 controls from our datasets (Supplementary Table 27). We found risk effects of HLA-DQA1^*01:02 (FDR P = 0.014), HLA-DRB1^*15:01 (FDR P = 0.083), and HLA-DQB1^*06:02 (FDR P = 0.010) (Supplementary Table 28). After conditioning on the sentinel meta-analysis variant in this region (rs78738018), association signals were lost for the three alleles, suggesting that the signal observed at the variant level is due to the association of these three alleles. These alleles form the HLA-DQA1^*01:02~HLA-DQB1^*06:02~HLA-DRB1^*15:01 (DR15) haplotype, which is also associated with Alzheimer’s disease in our sample (FDR P = 0.013) (Supplementary Table 29). Taken together, these results suggest a central role of the DR15 haplotype in Alzheimer’s disease risk, a finding originally discovered in a small study in the Tunisian population¹⁰⁷ and more recently in a large ADGC analysis¹⁰⁶. Intriguingly, the DR15 haplotype and its component alleles also associate with protection against diabetes¹⁰⁸, a high risk for multiple sclerosis^109,110, and risk or protective effects with many other immune-mediated diseases (Supplementary Table 30). Moreover, the associated diseases include a large number of traits queried from an HLA-specific Phewas¹¹¹, including neurological diseases (for example, Parkinson’s disease^112,113) and diseases with risk factors for Alzheimer’s disease (for example, hyperthyroidism¹¹⁴), pointing to potential shared and/or interacting mechanisms and co-morbidities, a common paradigm in the MHC locus¹¹⁵. Two additional alleles, HLA-DQA1^*03:01 and HLA-DQB1^*03:02, belonging to another haplotype, show a protective effect on Alzheimer’s disease, but their signal was lost after conditioning on HLA-DQA1^*01:02, and the HLA-DQA1^*03:01~HLA-DQB1^*03:02 haplotype is not associated with Alzheimer’s disease (FDR P = 0.651).

Genetic correlations with Alzheimer’s disease.

As described above, several of our genome-wide loci have potentially interesting co-morbid or pleiotropic associations with traits that may be relevant to the pathology of Alzheimer’s disease. To investigate the extent of LOAD’s shared genetic architecture with other traits, we performed LD-score regression to estimate the genetic correlation between LOAD and 792 human diseases, traits and behaviors^12,116 (Supplementary Table 31). The common variant genetic architecture of LOAD was positively correlated with a maternal family history of Alzheimer’s disease/dementia (r_g for the genetic correlation of two traits = 0.81; FDR P = 2.79 × 10⁻⁷), similar to the Marioni et al. family proxy analyses²⁶, which found maternal genetic correlation with Alzheimer’s disease to be higher than that for paternal Alzheimer’s disease (r_g = 0.91 and 0.66, respectively). There is substantial overlap between these estimates, as the Marioni et al. analyses include the 2013 IGAP summary statistics and employed the same UK Biobank variable that we used for r_g estimates with maternal history of dementia. We also find significant negative correlation between Alzheimer’s disease and multiple measures of educational attainment (for example, college completion, r_g = −0.24; years of schooling, r_g range = −0.19 to −0.24; cognitive scores, r_g = −0.24 and −0.25) (FDR P < 0.05), supporting the theory that a greater cognitive reserve could help protect against development of LOAD¹¹⁷. The extent to which socioeconomic, environmental, or cultural factors contribute to the correlation between educational attainment and risk for Alzheimer’s disease is unknown, but research shows dementia risk to be associated with lower socio-economic status, independently of education status^118,119. We also found negative correlations at P < 0.05 with multiple measures of cardiovascular health (that is, family history of high blood pressure and heart disease and vascular/heart problems) and diabetes (that is, fasting proinsulin, basal metabolic rate and fasting insulin), supporting previous research suggesting that use of blood pressure and diabetic medications may reduce the risk of Alzheimer’s disease¹²⁰. In fact, use of blood pressure medication does show a negative genetic correlation with Alzheimer’s disease in our study (r_g = −0.12; P = 0.035), although this result does not survive FDR correction. These and other top results from this analysis (for example, body mass index, height; see Supplementary Table 31 for a full list of other nominally significant correlations) have been linked to Alzheimer’s disease previously^{116,120–127}, either through suggestive or significant genetic or epidemiological associations (see Kuzma et al.¹²⁸ for a recent review), but the multiple measures here support and emphasize their genetic correlation with LOAD and highlight the possible genetic pleiotropy or co-morbidity of these traits with pathology of LOAD.

Discussion

In conclusion, our work identifies five new genome-wide associations for LOAD and shows that GWAS data combined with high-quality imputation panels can reveal rare disease risk variants (for example, TREM2). The enrichment of rare variants in pathways associated with Alzheimer’s disease indicates that additional rare variants remain to be identified, and larger samples and better imputation panels will facilitate identifying them. While these rare variants may not contribute substantially to the predictive value of genetic findings, they will enhance the understanding of disease mechanisms and potential drug targets. Discovery of the risk genes at genome-wide loci remains challenging, but we demonstrate that converging evidence from existing and new analyses can prioritize risk genes. We also show that APP metabolism is associated with not only early-onset Alzheimer’s disease but also LOAD, suggesting that therapies developed by studying early-onset families could also be applicable to the more common late-onset form of the disease. Pathway analysis showing that tau is involved in LOAD supports recent evidence that tau may play an early pathological role in Alzheimer’s disease^129–131 and confirms that therapies targeting tangle formation/degradation could potentially affect LOAD. Finally, our fine-mapping analyses of HLA and genetic correlation results point to LOAD’s shared genetic architecture with many immunemediated and cognitive traits, suggesting that research and interventions that elucidate mechanisms behind these relationships could also yield fruitful therapeutic strategies for LOAD.

Methods

Samples.

All Stage 1 meta-analysis samples are from four consortia: ADGC, CHARGE, EADI, and GERAD/PERADES. Summary demographics of all 46 case-control studies from the four consortia are described in Supplementary Tables 1 and 2. Written informed consent was obtained from study participants or, for those with substantial cognitive impairment, from a caregiver, legal guardian, or other proxy. Study protocols for all cohorts were reviewed and approved by the appropriate institutional review boards. Further details of all cohorts can be found in the Supplementary Note.

Pre-imputation genotype chip quality control.

Standard quality control was performed on all datasets individually, including exclusion of individuals with low call rate, individuals with a high degree of relatedness, and variants with low call rate. Individuals with non-European ancestry according to principal components analysis of ancestry-informative markers were excluded from the further analysis.

Imputation and pre-analysis quality control.

Following genotype chip quality control, each dataset was phased and imputed to the 1,000 Genomes Project (phase 1 integrated release 3, March 2012)¹³² using SHAPEIT/IMPUTE2^133,134 or MaCH/Minimac^135,136 software (Supplementary Table 3). All reference population haplotypes were used for the imputation, as this method improves accuracy of imputation for low-frequency variants¹³⁷. Common variants (MAF ≥ 0.01%) with an r² or an information measure <0.40 from MaCH and IMPUTE2 were excluded from further analyses. Rare variants (MAF < 0.01%) with a ‘global’ weighted imputation quality score of <0.70 were also excluded from analyses. This score was calculated by weighting each variant’s MACH/IMPUTE2 imputation quality score by study sample size and combining these weighted scores for use as a post-analysis filter. We also required the presence of each variant in 30% of cases and 30% of controls across all datasets.

Stage 1 association analysis and meta-analysis.

Stage 1 single variant-based association analysis employed an additive genotype model adjusting for age (defined as age-at-onset for cases and age-at-last exam for controls), sex, and population substructure using principal components¹³⁸. The score test was implemented on all case-control datasets. This test is optimal for meta-analysis of rare variants due to its balance between power and control of type 1 error¹³⁹. Family datasets were tested using GWAF¹⁴⁰, with generalized estimating equations (GEE) implemented for common variants (MAF ≥ 0.01), and a general linear mixed effects model (GLMM) implemented for rare variants (MAF < 0.01), per our preliminary data showing that the behavior of the test statistics for GEE was fine for common variants but inflated for rare variants, while GLMM controlled this rare-variant inflation. Variants with regression coefficient |β| > 5 or P value equal to 0 or 1 were excluded from further analysis.

Within-study results for Stage 1 were meta-analyzed in METAL¹⁴¹ using an inverse-variance-based model with genomic control. The meta-analysis was split into two separate analyses according to the study sample size, with all studies being included in the analysis of common variants (MAF ≥ 0.01), and only studies with a total sample size of 400 or greater being included in the rare-variant (MAF < 0.01) analysis. See the Supplementary Note for further details of the meta-analyses methods.

Stage 1 summary statistics quality control and analysis.

Genomic inflation was calculated for lambda (λ) in the GenABEL package¹⁴². In addition, we performed LDSC regression via LD Hub v.1.9.0 (refs. ^12,13) to calculate the LD score regression intercept and derive a heritability estimate for the inverse-variance weighted meta-analysis summary statistics. The APOE region (Chr19:45,116,911–46,318,605) was removed to calculate the intercept. Removal of the APOE region reduced the heritability estimate slightly from 0.071 (s.e.m. = 0.011) to 0.0637 (s.e.m. = 0.009).

LDSC was also employed via the LD Hub web server to obtain genetic correlation estimates (rg)¹¹⁶ between LOAD and a wide range of other disorders, diseases and human traits, including 518 UK BioBank traits¹⁴³. UK BioBank is a large long-term study (~500,000 volunteers aged 40 to 69) begun in 2006 in the United Kingdom, which is investigating the contributions of genetic predisposition and environmental exposure (that is, nutrition, lifestyle, and medications) to the development of disease. While volunteers in the study are generally healthier than the overall United Kingdom population¹⁴⁴, its large size and comprehensive data collection make the study an invaluable resource for researchers looking to interrogate the combined effect of genetics and environmental factors on disease. Before analyses in LD Hub, we removed all SNPs with extremely large effect sizes including the MHC (Chr6:26,000,000–34,000,000) and APOE region, as outliers can overly influence the regression analyses. A total of 1,180,989 variants were used in the correlation analyses. Statistical significance of the genetic correlations was estimated using 5% Benjamini−Hochberg FDR corrected P values.

GCTA COJO²⁷ was used to conduct conditional analysis of the Stage 1 summary statistics, with 28,730 unrelated individuals from the ADGC as a reference panel for calculation of LD. See URLs for methods for creation of the ‘ADGC reference dataset’.

Stage 2 and 3 genotyping, quality control, and analysis.

Stage 2 genotypes were determined for 8,362 cases and 10,483 controls (Supplementary Table 4). 1,633 variants from the I-select chip were located in the 24 genome-wide loci (defined by the LD blocks of the sentinel variants; excluding the APOE region), with an average of 68 variants per locus. The most well-covered loci were HLA-DRB1, M24A2, and PICALM (763, 202, and 156 variants available, respectively); the least were MAF, ADAMTS1, and INPP5D (0, 4, and 5 variants, respectively).

Stage 3A was conducted for variants selected as novel loci from meta-analyses of Stages 1 and 2 with P < 5 × 10⁻⁷ (9 variants) and variants that were previously significant (P < 5 × 10⁻⁸) that were not genome-wide significant after Stages 1 and 2 (2 variants) (4,930 cases and 6,736 controls) (Supplementary Table 5). Variants were genotyped using Taqman.

Stage 3B, which combined samples from Stage 2 and 3A, included variants with MAF < 0.05 and P < 1 × 10⁻⁵ or variants with MAF ≥ 0.05 and P < 5 × 10⁻⁶ in novel loci not covered in the 2013 iSelect genotyping¹ (13,292 cases and 17,219 controls) (Supplementary Table 7). See the Supplementary Note for details on selection of variants for Stage 3B follow-up genotyping. For Stages 1, 2, and 3, samples did not overlap.

Per-sample quality checks for genetic sex and relatedness were performed in PLINK. Sex mismatches or individuals showing a high degree of relatedness (identical-by-descent value of 0.98 or greater) were removed from the analysis. A panel of ancestry-informative markers was used to perform principal component analysis with SMARTPCA from EIGENSOFT 4.2 software¹⁴⁵, and individuals with non-European ancestry were excluded. Variant quality control was also performed separately in each country including removal of variants missing in more than 10% of individuals, having a Hardy−Weinberg P value in controls lower than 1 × 10⁻⁶ or a P value for missingness between cases and controls lower than 1 × 10⁻⁶.

Per-study analysis for Stage 2 and Stage 3 followed the same analysis procedures described for Stage 1, except for covariate adjustments per cohort, where all analyses were adjusted on sex and age apart from the Italian, Swedish, and Gr@ACE cohorts, which were also adjusted for principal components. Within-study results were meta-analyzed in METAL¹⁴¹ using an inverse-variance-based model.

Characterization of gene(s) and non-coding features in associated loci.

We determined the base-pair boundaries of the search space for potential gene(s) and non-coding features in each of the 24 associated loci (excluding APOE) using the ‘proxy search’ mechanism in LDLink¹⁴⁶. LDLink uses 1,000 genomes genotypes to calculate LD for a selected population; in our case all five European populations were selected (population codes CEU, TSI, FIN, GBR, and IBS). The boundaries for all variants in LD (r² ≥ 0.5) with the top associated variant from the Stage 2 meta-analysis for each region ±500 kb of the ends of the LD blocks (as eQTL controlled genes are typically less than 500 kb from their controlling variant¹⁴⁷) were input into the UCSC genome browser’s ‘Table Browser’ for RefSeq¹⁴⁸ and GENCODEv24lift37¹⁴⁹ genes at each associated locus. The average size of the LD blocks was 123 kb.

Identification of potentially causal coding or splicing variants.

To identify deleterious coding or splicing variants that may represent causal variants for our genome-wide loci, we first used SNIPA¹⁵⁰ to identify variants in high LD (defined as r² > 0.7) with the sentinel variants of the 24 genome-wide loci (excluding APOE) (n = 1,073). The sentinel variants were defined as the variants with the lowest P in each genome-wide locus. We then used Ensembl VEP¹⁵¹ for annotation of the set of sentinel variants and their proxies. We used BLOSUM62 (ref. ¹⁵²), SIFT¹⁵³, Polyphen-2 (ref. ¹⁵⁴), CADD¹⁵⁵, Condel¹⁵⁶, MPC¹⁵⁷ and Eigen¹⁵⁸ to predict the pathogenicity of protein-altering exonic variants and MaxEntScan to predict the splicing potential of variants. Splicing variants with high splicing potential according to MaxEntScan¹⁵⁹ and protein-coding variants predicted to be deleterious by two or more programs were considered to be potentially causal variants for a locus. It should be noted that while we do include rare variants from imputation in our analyses, we may be missing many rare causal variants in this study.

Identification of genes with rare-variant burden via gene-based testing.

We used the summary statistics results of a large whole-exome sequencing (WES) study of LOAD, the Alzheimer’s Disease Sequencing Project (ADSP) case-control study (n = 5,740 LOAD cases and 5,096 cognitively normal controls of NHW ancestry) to identify genes within our genome-wide loci that may contribute to the association signal through rare deleterious coding, splicing or LOF variants. The individuals in the ADSP study largely overlap with individuals in the ADGC and CHARGE cohorts included in our Stage 1 meta-analysis. All 400 protein-coding genes within our LD-defined genome-wide loci were annotated with the gene-based results from this study, and the results were corrected using a 1% FDR P as a cutoff for significance. Complete details of the analysis can be found in Bis et al.⁴⁹ and the Supplementary Note.

Regulatory variant and eQTL analysis.

To identify potential functional risk variants and genes at each associated locus, we first annotated a list of prioritized variants from the 24 associated loci (excluding APOE) (n = 1,873). This variant list combined variants in LD with the sentinel variants (r² ≥ 0.5) using INFERNO¹⁶⁰ LD expansion (n = 1,339) and variants with suggestive significance (P < 10⁻⁵) and LD (r² ≥ 0.5) with the sentinel variants for the 24 associated loci (excluding APOE) (n = 1,421 variants). We then identified variants with regulatory potential in this set of variants using four programs that incorporate various annotations to identify likely regulatory variants: RegulomeDB⁵⁶, HaploReg v.4.1 (refs. ^57,161), GWAS4D⁵⁹, and the Ensembl Regulatory Build⁵⁸. We used the ChromHMM (core 15-state model) as ‘source epigenomes’ for the HaploReg analyses. We used immune (Monocytes-CD14⁺, GM12878 lymphoblastoid, HSMM myoblast) and brain (NH-A astroctyes) for the Ensembl Regulatory Build analyses. We then used the list of 1,873 prioritized variants to search for genes functionally linked via eQTLs in LOAD relevant tissues including various brain and blood tissue types, including all immune-related cell types, most specifically myeloid cells (macrophages and monocytes) and B-lymphoid cells, which are cell types implicated in LOAD and neurodegeneration by a number of recent studies^{14,45,162,163}. While their specificity may be lower for identifying Alzheimer’s disease risk eQTLs, we included whole blood cell studies in our Alzheimer’s disease–relevant tissue class due to their high correlation of eQTLs with Alzheimer’s disease–relevant tissues (70% with brain¹⁶⁴; 51–70% for monocytes and lymphoblastoid cell lines¹⁶⁵) and their large sample sizes that allow for increased discovery power. See the Supplementary Note for details on the eQTL databases and studies searched, and Supplementary Table 13 for sample sizes of each database/study.

Formal co-localization testing of our summary Stage 1 results was conducted using (1) COLOC¹⁶⁶ via INFERNO and (2) Summary Mendelian Randomization (SMR)-Heidi analysis¹⁶⁷. The approximate Bayes factor (ABF), which was used to assess significance in the INFERNO COLOC analysis, is a summary measure that provides an alternative to the P value for the identification of associations as significant. SMR-Heidi analysis, which employs a heterogeneity test (HEIDI test) to distinguish pleiotropy or causality (a single genetic variant affecting both gene expression and the trait) from linkage (two distinct genetic variants in LD, one affecting gene expression and one affecting trait), was also employed for co-localization analysis. Genes located less than 1 Mb from the GWAS sentinel variants that pass a 5% Benjamini–Hochberg FDR-corrected SMR P-value significance threshold and a HEIDI P-value > 0.05 threshold were considered significant. The Westra eQTL¹⁶⁸ summary data and Consortium for the Architecture of Gene Expression (CAGE) eQTL summary data were used for analysis. These datasets, conducted in whole blood, are large eQTL studies (Westra: discovery phase n = 5,311, replication phase n = 2,775; CAGE: n = 2,765), and while there is some overlap in samples between the two datasets, CAGE provides finer coverage. The ADGC reference panel dataset referenced above for GCTA COJO analysis was used for LD calculations.

Human brain gene expression analyses.

We also evaluated gene expression of all candidate genes in the associated loci (see Supplementary Table 8 for a complete list of genes searched), using differential Alzheimer’s disease gene expression results from AlzBase³¹, brain tissue expression from the Brain RNA-seq Database^32,33 (see URLs), and the HuMi_Aged gene set³⁴, a set of genes preferentially expressed in aged human microglia established through RNA-seq expression analysis of aged human microglial cells from ten post-mortem brains. AlzBase includes transcription data from brain and blood from aging, non-dementia, mild cognitive impairment, early-stage Alzheimer’s disease, and late-stage Alzheimer’s disease. See ALZBase (see URLs) for a complete list of studies included in the search. Correlation values for the BRAAK stage expression were taken from the Zhang et al.³⁰ study of 1,647 post-mortem brain tissues from LOAD patients and non-demented subjects.

Pathway analysis.

Pathway analyses were performed with MAGMA⁶¹, which performs SNP-wise gene analysis of summary statistics with correction for LD between variants and genes to test whether sets of genes are jointly associated with a phenotype (that is, LOAD), compared to other genes across the genome. Adaptive permutation was used to produce an empirical P value and an FDR-corrected q value. Gene sets used in the analyses were from GO^169,170, KEGG^171,172, REACTOME^173,174, BIOCARTA, and MGI¹⁷⁵ pathways. Analyses were restricted to gene sets containing between 10 and 500 genes, a total of 10,861 sets. Variants were restricted to common variants (MAF ≥ 0.01) and rare variants (MAF < 0.01) only for each analysis, and separate analyses for each model included and excluded the APOE region. Analyses were also performed after removal of all genome-wide-significant genes. Primary analyses used a 35-kb upstream/10-kb downstream window around each gene in order to capture potential regulatory variants for each gene, while secondary analyses were run using a 0-kb window¹⁷⁶. To test for significant correlation between common and rare-variant gene results, we performed a gene property analysis in MAGMA, regressing the gene-wide association statistics from rare variants on the corresponding statistics from common variants, correcting for LD between variants and genes using the ADGC reference panel. The Aβ-centered network pathway analysis used a curated list of 32 Aβ-related gene sets and all 335 genes combined (see Campion et al.⁶⁴ for details). The combined dataset of 28,730 unrelated individuals from the ADGC referenced in the GCTA COJO analysis was used as a reference set for LD calculations in these analyses.

Validation of prioritization method.

Evaluation of the prioritization of the risk genes in genome-wide loci was done using STRING¹⁷⁷, and Jensen Diseases¹⁷⁸, Jensen Tissues¹⁷⁹, dbGAP gene sets, and the ARCHS4¹⁸⁰ resource via the EnrichR¹⁸¹ tool. We evaluated both the 400 genes set list and a list of 53 genes with priority score ≥ 5 (adding in APOE to both lists as the top gene in the APOE locus) using the standard settings for both STRING and EnrichR. We used the q value, which is the adjusted P value using the Benjamini–Hochberg FDR method with a 5% cutoff for correction for multiple hypotheses testing. We also performed ‘differentially expressed gene (DEG)’ sets analysis via FUMA¹⁸². These analyses were performed in order to assess whether our 53 prioritized genes were significantly differentially expressed in certain GTEx v.7 (ref. ¹⁰²; 30 general tissues and 53 specific tissues) or BrainSpan tissues (11 tissue developmental periods with distinct DEG sets ranging from early prenatal to middle adulthood)¹⁰³. FUMA defines DEG sets by calculating a two-sided t-test per tissue versus all remaining tissue types or developmental periods. Genes with a Bonferonni-corrected P < 0.05 and absolute log(fold change) ≥ 0.58 were considered DEGs. Input genes were tested against each of the DEG sets using the hypergeometric test. Significant enrichment was defined by Bonferonni-corrected P ≤ 0.05.

HLA region analysis.

Non-familial datasets from the ADGC, EADI and GERAD consortiums were used for HLA analysis. After imputation quality control, a total of 14,776 cases and 23,047 controls were available for analysis (Supplementary Table 27). Within ADGC, GenADA, ROSMAP, TARC1, TGEN2, and a subset of the UMCWRMSSM datasets were not imputed as Affymetrix genotyping arrays are not supported by the imputation software.

Imputation of HLA alleles.

Two-field resolution HLA alleles were imputed using the R package HIBAG v.1.4 (ref. ¹⁸³) and the NHW-specific training set. This software uses specific combinations of variants to predict HLA alleles. Alleles with an imputation posterior probability lower than 0.5 were considered as undetermined as recommended by HIBAG developers. HLA-A, HLA-B, HLA-C class I genes, and HLA-DPB1, HLA-DQA1, HLA-DQB1, and HLA-DRB1 class II genes were imputed. Individuals with more than two undetermined HLA alleles were excluded.

Statistical analysis.

All analyses were performed in R¹⁸⁴. Associations of HLA alleles with disease were tested using logistic regressions, adjusting for age, sex, and principal components as specified above for single variant association analysis. Only HLA alleles with a frequency higher than 1% were analyzed. Haplotype estimations and association analyses with disease were performed using the ‘haplo.glm’ function from the haplo.stats R package¹⁸⁵ with age, sex, and principal components as covariates. Analysis was performed on two-loci and three-loci haplotypes of HLA-DQA1, HLA-DQB1, and HLA-DRB1 genes. Haplotypes with a frequency below 1% were excluded from the analysis. Considering the high LD in the MHC region, only haplotypes predicted with posterior probabilities higher than 0.2 were considered for analysis. Meta-analysis P values were computed using an inverse-variance-based model as implemented in METAL software¹⁴¹. For haplotypes analysis, only individuals with no undetermined HLA alleles and only datasets with more than 100 cases or controls were included. Adjustments on HLA significant variants and HLA alleles were performed by introducing the variant or alleles as covariates in the regression models. Adjusted P values were computed using the FDR method and the R ‘p.adjust’ function, and applied to the meta-analysis P values. The FDR threshold was set to 10%.

Reporting Summary.

Further information on research design is available in the Nature Research Reporting Summary linked to this article.