Genomics link obesity and type 2 diabetes to Alzheimer’s disease to unveil novel biological insights

Abstract

Body mass index (BMI), type 2 diabetes (T2D) and associated cardiometabolic features modify Alzheimer’s disease (AD) risk, yet shared mechanisms remain poorly understood. Using sex- and age-stratified genotyping data for BMI and T2D, we investigate how these traits converge on shared genetic pathways to AD risk. Employing multi-trait, machine learning and single-cell transcriptomics, we identify sex-specific cardiometabolic liability linked to higher BMI-associated risk in women and T2D-driven risk in men. Variant-level analyses reveal AD risk associates with genetically-driven hypotension and hypoglycaemia. We identify 35 putative effector genes in seven independent loci colocalizing between BMI/T2D and AD, mapping to peripheral immune and metabolic tissues and cell-types. Pathway enrichment identifies druggable targets in calcium and potassium channel signaling. Across 81 approved drugs modulating shared risk genes, levosimendan - a calcium sensitizer for heart failure - inhibits tau oligomerization and emerges as a repurposing candidate. These findings elucidate sex-specific cardiometabolic drivers of AD, identify actionable biological pathways, and reveal drug candidates for AD prevention and treatment.

INTRODUCTION

Late-onset Alzheimer’s disease (AD) is a devastating disorder that affects patients and families alike. AD is the leading form of dementia¹, an umbrella term encompassing multiple diseases that progressively deteriorate cognition severely enough to impair independence. Globally, over 55 million individuals live with dementia, affecting approximately one in nine individuals over age 65 and one in three over age 85². With the global aging population projected to reach 2.4 billion individuals over 65 years by 2100³, from 800 million currently⁴, the global economic burden of AD is projected to escalate from $2.8 trillion in 2019⁵ to $16.9 trillion by 2050⁵.

The sex-specific prevalence in AD is striking, with women accounting for two-thirds of cases⁶ and exhibiting significantly higher age-adjusted odds of AD than men at older ages⁷ independent of their longer life expectancy. This contrasts with other neurodegenerative disorders, such as Parkinson’s disease, where men are at twice the risk⁸. Women also show higher susceptibility to APOE4-associated AD risk^9,10, for which current therapeutic approaches are less effective¹¹. Addressing modifiable risk factors could prevent 45% of dementia cases, with 12% directly attributable to cardiometabolic factors such as obesity, type 2 diabetes (T2D), high LDL cholesterol, and hypertension¹². Obesity and T2D are independent modifiable risk factors for dementia and AD^13–16, with their sex-specific prevalence - obesity being more common in women¹⁷ and T2D in men¹⁸ - possibly influencing the sex-specific risk of AD⁹.

Midlife overweight and obesity elevate dementia risk by 33% and 63% respectively¹⁹. Likewise, diabetes elevates dementia risk by 59%²⁰. Despite strong epidemiological associations, the molecular bases underlying obesity, T2D, and AD are poorly understood. It is unclear whether epidemiological findings stem from non-linear associations such as age-dependent and disease-stage reverse causation or sex-specific differences²¹. This uncertainty is compounded by the prevalent incorporation of AD-proxies - based on undiagnosed parental cases - in genetic association studies^22–24.

Human genetic data enables systematic decoding of shared biology linking modifiable risk factors to the long preclinical and prodromal curves of AD. We mapped the shared genetic architecture linking obesity, T2D, and AD by integrating genetic correlation, colocalization, phenome-wide association, and protein-protein interaction analyses with deep learning and single-cell transcriptomic approaches to identify actionable biology in effector genes and tissues.

RESULTS

Study design

Our multi-trait analyses used European-ancestry genotyping data from: GIANT consortium meta-analysis (BMI)²⁵, T2D genomics Initiative meta-analysis (T2D risk²⁶), and European Alzheimer & Dementia Biobank (EADB) consortium meta-analysis (AD risk)²⁷ (Table 1). The primary AD meta-analysis (AD1) includes parental proxy cases (self-reported by offsprings, undiagnosed), risking signal dilution from potential misclassification (e.g. undiagnosed dementia subtypes or lack of neurodegenerative pathology^27–29). Hence, we additionally analyzed data excluding AD-proxy cases (AD2) and excluding AD cases, age-matched and non-age-matched controls from the UK Biobank (UKB; AD3) to minimize age-related confounding. The data were provided by the EADB consortium and has recently been made publicly available³⁰.

Table 1 |.

Dataset details.

Dataset	Cases	Controls	Total	Source
AD1	39,106 diagnosed 46,828 proxies	401,577	487.511	EADB consortium meta-analysis1
AD2	39,106 diagnosed	401,577	440.683	EADB consortium meta-analysis
AD3	36,659 diagnosed	63,137	99.796
BMI	-	-	681,275	GIANT consortium meta-analysis2
BMI male	-	-	374,756	GIANT consortium3
BMI female	-	-	434,794
BMI male under 50	-	-	50,095
BMI male over 50	-	-	93,021
BMI female under 50	-	-	70,692
BMI female over 50	-	-	106,497
T2D	180,834	1,159,055	1,339,889	T2D genomics Initiative meta-analysis4
T2D male	41,846	383,767	425,613	T2D genomics Initiative5
T2D female	30,053	434,336	464,389

To account for potential sex-specific genetic effects associated with BMI and T2D on AD risk, we analyzed sex-stratified genotyping data from the GIANT consortium (BMI)³¹ and T2D genomics initiative³². We additionally analyzed age-stratified BMI data (under/over age 50)³¹. Age 50 approximates the average age of menopause, a critical period affecting metabolic and hormonal risk factors³¹, and marks onset of asymptomatic AD preclinical pathology³³. Sex-specific AD data is currently not available, which may result in an underestimation of joint sex-specific genetic effects of BMI, T2D, and AD.

Subsequent analyses included: (1) genetic correlations; (2) risk variant identification through multi-trait colocalization; (3) variant-level phenotyping through phenome-wide associations; (4) effector gene mapping through variant-to-gene linkage; (5) effector tissue and cell-type prioritization through multi-omic and deep learning perturbation analyses; (6) mechanistic detailing through enrichment analyses; and (7) genetic-backed drug repurposing identification through gene-drug interaction analyses (Fig. 1).

Fig. 1 |. Schematic overview of the study.

Created in BioRender.

Genetic correlations between BMI and AD, but not T2D and AD, suggest sex and age-specificity

We assessed genetic correlations (rg) between BMI, T2D, and AD using linkage disequilibrium score regression (LDSC; Fig. 2; Table S1; Methods). BMI showed negative genetic correlation with AD risk (r_g= −0.13; p=1x10⁻⁶), linking lower BMI variants with higher AD risk. This correlation was significantly stronger in women than men across all AD datasets (p<4x10⁻⁴).

Fig. 2 |. Cross-phenotype genetic correlation estimates.

AD1: Includes AD proxy cases and non-age-matched controls. AD2: Excludes AD proxy cases, but retains non-age matched controls. AD3: Excludes AD proxy cases and non-age-matched controls. Sex-combined datasets: AD1, AD2, AD3, BMI, T2D. Sex and age-stratified datasets: BMI.Male, BMI.Female, BMI.Male.Under50, BMI.Male.Over50, BMI.Female.Under50, BMI.Female.Over50, T2D.Male, T2D.Female. The genetic correlation matrix was generated using the corrplot library in R (version 4.2.2). P-values are shown next to GC estimates. Statistical significance between genetic correlation estimates are stratified by p<0.05 (*); p<0.01 (**); p<0.001 (***); p<0.0001 (****) (Table S1).

Divergent age patterns (≤50y versus >50y) emerged when excluding the UKB. In men, a stronger genetic correlation was observed for AD and BMI >50y versus ≤50y (p<6x10⁻⁴). In women, a stronger genetic correlation was observed for AD and BMI ≤50y versus >50y (p<6x10⁻⁶). The genetic correlation for BMI ≤50y is significantly stronger for women versus men (p<5x10⁻⁶), but non-significant >50y. These sex and age patterns were absent in AD-proxy- and UKB-inclusive data, indicating sex and age-dependent specificity between BMI and clinically diagnosed AD that disappeared when including AD-proxy cases.

No sex differences emerged in the genetic correlation between T2D and AD despite significant correlation in the overall population (r_g= −0.09; p=6x10⁻³; Table S1). However, excluding AD-proxy cases and UKB abolished this correlation, suggesting that the shared T2D and AD architecture may be biased by AD-proxies or reflect broader dementia susceptibility rather than AD-specific biology.

Overall, BMI showed stronger genetic correlation with AD than T2D (Table S1). Polygenic BMI risk confers stronger genetic susceptibility to AD in women. Marked age-sensitivity, particularly under 50 years in women, suggests non-linear, age- and sex-dependent mechanisms.

Colocalization analysis identifies unique loci between BMI, T2D and AD specific to sex and age

We identified genome-wide significant AD variants (p≤5x10⁻⁸) concurrently associated with BMI and/or T2D (p≤5x10⁻⁴), yielding 32 unique BMI-AD and 16 T2D-AD overlapped loci (Table S2; Methods). Colocalization analyses complemented by locuszoom visual inspection minimized spurious/mediated pleiotropy (Fig. 3A), identifying 14 shared causal variants in seven independent loci (r²<0.01; ±500 kb): four shared loci for BMI and AD (BMI-AD), two shared loci for T2D and AD (T2D-AD) and one shared locus for BMI, T2D and AD (BMI-T2D-AD) (Fig. 3B; Methods). The low percentage of overlapped loci that colocalize reflect a high degree of pleiotropy in the shared genetic architecture of BMI, T2D risk and AD risk.

Fig. 3 |. Genetic colocalization analysis.

(A) Schematic overview of the multi-trait colocalized locus discovery pipeline. (B) Karyotype representation of the colocalized loci for BMI-AD (chr5, chr14, chr19), T2D-AD (chr8, chr17) and BMI-T2D-AD (chr16). (C) Locuszoom plots for the BMI-AD rs17125924 locus (GRCh37). The overlapped variant is circled in green (Table S3) while the colocalized variant is shown in yellow (rs17125924). Variants in linkage disequilibrium (R²) are color-coded. (D) Forest plot for the BMI-AD rs17125924 locus showcasing odds ratio for dichotomous traits (AD and T2D) and effect size (beta) for the continuous trait (BMI). P-values are shown under each odds ratio/beta estimate. Odds ratio and p-values for (E) BMI-AD, (F) T2D-AD and (G) BMI-T2D-AD risk variants across AD datasets (AD1, AD2 and AD3).

We mapped each locus to the closest gene for orientation. COX7C and FERMT2 BMI-AD alleles increased both BMI and AD risk. ATP8B3 and NECTIN2-TOMM40 (APOE) increased AD risk linked to lower BMI. T2D-AD and BMI-T2D-AD alleles increased AD risk associated with lower BMI and T2D risk (Table 2).

Table 2 |.

Novel sex and age-specific loci associated with BMI-AD, T2D-AD and BMI-T2D-AD (Table S3).

Traits	Effect direction	Colocalized SNP	Closest Gene	PPC	Chr	Pos	EA (↑ AD)	NEA	EAF (%)	Beta (SE)	P
BMI-AD
AD2	+	rs2452758	COX7C	0.569	5	86885886	G	A	76.0	0.07 (0.010)	2.2x10−10
BMI_female	+								78.0	0.01 (0.003)	1.6x10−5
AD1	+								9.0	0.09 (0.014)	5.8x10−10
BMI_female	+	rs17125924	FERMT2	0.793	14	52924962	G	A	8.0	0.02 (0.004)	3.0x10−5
BMI_female_over50	+								7.0	0.03 (0.008)	2.2x10−5
AD1	+	rs11668832	ATP8B3	0.853	19	1812683	A	G	4.6	0.05 (0.009)	2.1x10−8
BMI	-								4.6	−0.01 (0.002)	7.5x10−15
AD1	+	rs7254892	TOMM40	0.999	19	44886339	G	A	9.7	0.52 (0.027)	7.7x10−82
BMI	-								9.6	−0.03 (0.005)	4.0x10−9
AD2	+	rs6859	NECTIN2	0.552	19	44878777	A	G	4.5	0.270 (0.009)	1.2x10−173
BMI	-								4.1	−0.009 (0.002)	4.6x10−7
T2D-AD
AD1	+	rs13257021	TP53INP1	0.888	8	94953467	G	A	5.3	0.045 (0.008)	2.7x10−8
T2D	-								5.1	−0.049 (0.006)	3.1x10−14
AD1	+	rs13257021	TP53INP1	0.888	8	94953467	G	A	5.3	0.045 (0.008)	2.7x10−8
T2D_female	-								5.1	−0.050 (0.009)	3.2x10−7
AD1	+	rs896854	TP53INP1	0.870	8	94948283	C	T	5.1	0.044 (0.008)	4.8x10−8
T2D_male	-								5.0	−0.053 (0.009)	3.4x10−10
AD1	+	rs4311	ACE	0.566	17	63483402	C	T	5.2	0.066 (0.008)	6.9x10−16
T2D	-								5.2	−0.028 (0.007)	1.8x10−5
AD1	+	rs4311	ACE	0.566	17	63483402	C	T	5.2	0.066 (0.008)	6.9x10−16
T2D_male	-								5.3	−0.040 (0.009)	3.0x10−6
AD2	+	rs4311	ACE	0.566	17	63483402	C	T	5.1	0.064 (0.009)	2.4x10−11
T2D_male	-								5.3	−0.040 (0.009)	3.0x10−6
BMI-T2D-AD
AD1	+	rs12325539	C16orf92	0.987	16	30022312	T	C	6.2	0.057 (0.008)	1.2x10−11
T2D	-								6.0	−0.041 (0.007)	2.7x10−10
AD1	+								5.7	0.050 (0.008)	1.4x10−9
BMI_female	-	rs11150577	TAOK2	0.321	16	29975204	G	A	5.4	−0.022 (0.002)	2.6x10−21
BMI_female_over50	-								5.2	−0.023 (0.005)	9.0x10−6
T2D_female	-								5.5	−0.035 (0.010)	2.9x10−4
AD1	+	rs12325539	C16orf92	0.962	16	30022312	T	C	6.2	0.057 (0.008)	1.2x10−11
T2D_male	-								6.0	−0.043 (0.009)	9.3x10−7

Effect direction: aligned to the risk increasing allele for AD. Colocalized: SNP identified by HyPrColoc to drive the same genetic signal in both traits. SNP: Single nucleotide polymorphism. PPC: Posterior Probability of Colocalization. Chr: Chromosome. Pos: Position in hg38. EA: Effect allele aligned to the risk increasing allele for AD. NEA: Non-effect allele. LD: Linkage disequilibrium. NEAF: Non-effect allele frequency. Beta (SE): effect size (standard error). P: p-value.

To determine sex-specific effects, we evaluated regional locus structures using locuszoom plots (Fig. 3C) and association metrics (p-value, beta) across traits (Fig. 3D; Fig. S1A–I; Table S3). For instance, the COX7C locus is women and age-specific (Fig. S1A), exhibiting a locus peak and significant GWAS association from BMI in women (p=1.6x10⁻⁵) not found in BMI in men (p=0.13). Conversely, the ACE locus is specific to men, showing no locus peak and weak genetic association from T2D GWAS in women (p=0.02), and significantly higher effect size in T2D in men versus women (p=1.7x10⁻⁷) (Fig. S1G; Table S3).

Effect sizes remained consistent across AD1 (AD-proxy-inclusive), AD2 (AD-proxy-excluded), and AD3 (UKB-excluded; Fig. 3E–G), indicating robustness to case misclassification and improper control age-matching. Overall, stringent colocalization revealed women-predominant BMI-AD associations with age-sensitivity, while ACE T2D-AD effects were specific to men.

Phenome-wide association analyses pinpoint phenotype and sex-specific biology

Phenome-wide association screening (n=1,314 phenotypes) of seven colocalized loci uncovered broad biological links (Table S4; Methods). BMI-AD variants consistently associate with lower blood pressure (Fig. 4; Fig. S2A–J), but show mixed cardiometabolic associations independent of BMI effect direction: COX7C (women-specific) exhibited the standard obesity profile while FERMT2 resembled metabolically healthy obesity (elevated BMI, hematological inflammation markers, but lower blood and pulse pressure)^34,35.

Fig. 4 |. Phenome wide association study (PheWAS) analyses.

(A) Summary of PheWAS associations of inflammatory and cardiometabolic profiles. PheWAS profile of the (B) rs11668832 BMI-AD locus (ATP8B3) in chr19, (C) rs13257021 T2D-AD locus (TP53INP1) in chr8 and (D) rs11150577 BMI-T2D-AD locus (TAOK2) in chr16. PheWAS p-value is set at a Bonferroni corrected threshold of 4.3x10⁻⁵ (orange dotted line). All PheWAS results are aligned to the risk-increasing allele for AD. Position in hg19.

Aligned to the risk increasing allele for AD, T2D-AD and BMI-T2D-AD loci associated with favorable profiles: lower glucose, HbA1c, and GGT (pro-inflammatory liver enzyme), with healthier lipid signatures and blood pressure (Fig. 4; Fig. S2A–J). BMI loci exhibited women-specific associations (SHBG, age at menarche), absent in men. The men-specific TP53INP1 locus (T2D-AD) associated with Y chromosome mosaic alterations, known to accelerate AD progression in men³⁶, corroborating sex-specific variant findings. Overall, BMI influences AD risk primarily through underlying cardiometabolic mechanisms, where genetically-driven hypotension and hypoglycemia associate with elevated AD risk.

Variant-to-gene mapping identifies multi-omic functional links to effector genes

To map the colocalized loci between BMI, T2D risk, and AD risk to effector genes, we applied the combined SNP-to-gene (cS2G) approach that ranks genes based on a hierarchy of gene prioritization strategies (Methods).

We mapped colocalized loci to effector genes using hierarchical cS2G prioritization variants in the same credible set with the colocalized variant (Table S5). Supporting evidence came from nearest transcription start site annotation (closest gene)³⁷, fine-mapped cis-sQTL data (54 tissues)³⁸ and plasma pQTL data³⁹ (Table 3; Table S6).

Table 3 |.

Genes with functional evidence across colocalized shared loci (all variants in the same credible set with the colocalized variant). The closest gene and those with coding evidence (exon or promoter) are shown in bold. Genes in each locus are ranked by mapping confidence from cS2G (Table S5).

Colocalized variant Locus (± 500kb)	Gene	AD	Phenotype pair	Credible set	Coloc. variant
				cS2G Annotation	GTEx-sQTL
BMI-AD
rs2452758 chr5:86385886-87385886	COX7C	Reported	Novel	Closest, ABC	-
rs17125924 chr14:52424962-53424962	FERMT2	Reported	Reported6	Exon, Promoter, Closest, GTEx-eQTL, EpiMap, ABC,	sQTL
rs11668832 chr19:1312683-2312683	ATP8B3	Reported	Novel	Promoter, Closest, GTEx-eQTL, EpiMap, ABC	sQTL
	REXO1	Reported	Novel	Promoter, GTEx-eQTL, EpiMap, ABC	-
	ABHD17A	Reported	Novel	eQTLGen	sQTL
	ADAT3	Novel	Novel	EpiMap	-
		SCAMP4	Novel	Novel	EpiMap	-
rs7254892 chr19:44386339-45386339	NECTIN2 (PVRL2)	Reported	Novel	Exon, Closest, GTEx-eQTL, EpiMap	-
	TOMM40	Reported	Novel	Exon, Closest	-
rs6859 chr19:44378777-45378777	APOC1	Reported	Novel	EpiMap	-
T2D-AD
rs13257021 chr8:94453467-95453467	TP53INP1	Reported	Reported7	Exon, Promoter, Closest, eQTLGen, GTEx-eQTL, Cicero, EpiMap, ABC	-
	INTS8	Novel	Novel	Exon, eQTLGen, EpiMap, ABC	sQTL
rs896854 chr8:94448283-95448283	NDUFAF6	Reported	Reported8	Promoter, GTEx-eQTL, Cicero	sQTL
	CCNE2	Reported	Novel	eQTLGen, GTEx-eQTL, EpiMap, ABC	-
	PLEKHF2	Novel	Novel	Cicero	-
rs4311 chr17:62983402-6398.402	ACE	Reported	Reported8,9	Exon, Promoter, Closest, eQTLGen, GTEx-eQTL, EpiMap, ABC	sQTL
	KCNH6	Novel	Novel	GTEx-eQTL, EpiMap	sQTL
	DCAF7	Novel	Novel	GTEx-eQTL	-
BMI-T2D-AD
rs11150577 chr16:29475204-30475204	TAOK2	Novel	Novel	Exon, Promoter, Closest, EpiMap, ABC,	sQTL
	TMEM219	Reported	Novel	Exon, Promoter, ABC	-
	DOC2A	Reported	Reported T2D-AD8	Exon, Promoter, GTEx-eQTL, ABC	sQTL
	TLCD3B	Novel	Novel	Promoter, GTEx-eQTL, EpiMap, ABC	sQTL
	TBX6	Reported	Novel	Exon	sQTL
	C16orf92	Reported	Novel	Cicero, Closest	-
	INO80E	Reported	Novel	eQTLGen, GTEx-eQTL, ABC	sQTL
	PPP4C	Reported	Novel	eQTLGen, EpiMap, ABC	-
rs12325539 chr16:29522312-30522312	MAPK3	Reported	Novel	eQTLGen, EpiMap	-
	KCTD13	Reported	Novel	eQTLGen, EpiMap,	sQTL
	YPEL3	Reported	Novel	GTEx-eQTL, EpiMap	sQTL
	GDPD3	Reported	Novel	GTEx-eQTL	-
	MVP	Novel	Reported T2D-AD10	EpiMap	sQTL
	SPN	Reported	Novel	EpiMap	-
	MAZ	Reported	Novel	EpiMap	-
	C16orf54	Novel	Novel	EpiMap	-
	HIRIP3	Reported	Novel	ABC	-

We identified 35 unique effector genes across seven loci, of which 10 are novel genes for AD (i.e. not previously reported in the literature) and 29 are novel genes for phenotype pairs. To assess whether effector genes incorporate similar functional networks, we performed a protein-protein interaction (PPI) analysis (Fig. 5; Table S7). Protein-protein interaction (PPI) analysis (Fig. 5; Table S7) revealed that the genes in the BMI-T2D-AD locus formed a strong interaction network, suggesting shared mechanistic pathways. Most other loci lacked inter-gene interactions, indicating locus-specific biology. NECTIN2-TOMM40 (APOE) showed strong intra-locus but no cross-loci interactions, suggesting non-APOE may be independent of APOE effects.

Fig. 5 |. Protein-protein interaction (PPI) analysis.

Genes from Table 3 were included. Analysis performed in STRING (version 12.0), indicating both functional and physical protein associations with sufficient confidence (0.4) including data from text mining, experiments, databases, co-expression, neighborhood, gene fusion and co-occurrence. Line thickness indicates the strength of data support.

Deep-learning and multi-omic analyses implicate peripheral biology in risk loci

We mapped tissue and cell-type associations using AlphaGenome⁴⁰ for lead variant effects across functional genomic data-types (Table S8; Fig. S3; Methods), cS2G, splicing QTL (sQTL) and protein QTL (pQTL) functional links (Table S5–S6), and GTEx tissue enrichment³⁸ via ShinyGO⁴¹ (Fig. S4; Methods). AlphaGenome aligned with cS2G and GTEx sQTLs, mapping risk loci primarily to peripheral immune and cardiometabolic tissues (Fig. 6A) with pervasive mesenchymal effects, particularly endothelial and immune cell-types (Fig. 6B). Notably, the women-specific COX7C locus induces strong expression changes in gonadal and reproductive tissues not predicted in other loci (Fig. 6C).

Fig. 6 |. Immune and metabolic tissue association analyses.

(A) AlphaGenome in silico perturbation prediction of the colocalized risk variant (rs4311) in the T2D-AD chr17 locus. The graph showcases the tissues where histone modifications (predicted from chromatin immunoprecipitation data) are most affected by perturbing the rs4311 variant. (B) AlphaGenome in silico perturbation prediction of the colocalized risk variant (rs13257021) in the T2D-AD chr8 locus. The graph showcases the cell-types where gene expression (predicted from poly(A) tail-containing RNA sequencing data) is most affected by perturbing the rs13257021 variant. (C) AlphaGenome in silico perturbation prediction of the colocalized risk variant (rs2452758) in the BMI-AD chr5 locus. The graph showcases the tissues where gene expression (predicted from total RNA sequencing data) is most affected by perturbing the rs2452758 variant. (D) Individual gene expression and % of expressing cells within the immune compartment of 17 peripheral tissues. (E) GWAS-eQTL colocalization locuszoom (LZ) region plot for BMI and TMEM219 eQTL in adipose tissue. (F) GWAS-eQTL colocalization LZ region plot for T2D and TP53INP1 eQTL in liver. (G) Human adipose tissue single-cell expression of TMEM219. (H) Human liver single-cell expression of TP53INP1. (I) Schematic overview of enhancer-to-gene (E2G) in adipose tissue and liver. (J) Summary associations of prioritized genes by E2G in adipose tissue and liver.

Peripheral immune niches

GTEx enrichment analyses implicated select loci in immune-related tissues (small intestine, colon, blood, spleen; Fig. S4). Single-cell transcriptomic analyses of the immune niches of 17 peripheral tissues⁴² revealed prevalent shared risk gene expression across immune cell-types (Fig. 6D; Fig. S5A–C). While not directly comparable, effector genes showed negligible central nervous system (CNS) microglia expression (Fig. 7B–C), the primary AD-enriched CNS cell-type²⁷, supporting non-CNS immune effects. Despite no overall sex differences in immune-related expression, five genes showed significantly stronger expression in men (FDR 1%): APOC1, NECTIN2, ADATA3, DOC2A, and ACE (Fig. S5D–E; Methods). APOE-associated AD risk in women^10,43 contrasts with the significant APOC1/NECTIN2 enriched expression in men, suggesting that men might benefit from protective immune-associated effects from the APOE locus. In contrast, men show higher baseline ACE activity and greater ACE inhibitor response for hypertension⁴⁴, suggesting that elevated ACE expression may contribute to specific immune signatures in men linking to hypertension-associated AD mechanisms.

Fig. 7 |. Regional and single-cell neural expression of high likelihood effector risk genes across colocalized loci.

(A) Bulk tissue expression from multiple regions across the neurotypical brain (GTEx). (B) Single-cell gene expression in the neurotypical prefrontal cortex (PFC). (C) Single-cell gene expression in the neurotypical hypothalamus. (D) Single-cell gene expression in the middle temporal gyrus (MTG) of individuals with AD. ACTB, FTHI1 and MALAT1 were used as reference genes due to their high and ubiquitous expression levels across brain regions and cell-types¹¹.

Peripheral cardiometabolic associations

GTEx enrichment extended to metabolic tissues (heart, vasculature, liver, adipose, muscle; Fig. S4). COLOCUS analysis evaluated colocalization between metabolic GWAS traits and eQTLs in ±100kb regions around lead variants (Table 4; Fig. S6; Table S9; Methods). Extensive colocalization emerged between metabolic GWAS traits and eQTLs in adipose tissue, liver, muscle, and pancreatic islets. We showcase BMI-TMEM219 colocalization in adipose tissue with broad cell-type expression (Fig. 6E, G) and T2D-TP53INP1 colocalization in liver with hepatocyte-specificity (Fig. 6F, H). The enhancer-to-gene (E2G) analysis⁴⁵ corroborated cS2G associations and validated COLOCUS findings by reporting extensive links to regulatory regions in adipose tissue and hepatocytes using tissue-specific data (Fig. 6I; Table S10). E2G highlights the still uncharacterized LOC100288123 as a novel liver-specific RNA gene (Fig. 6J).

Table 4 |.

Summary table of colocalized GWAS-eQTL pairs (build 37) in metabolic tissues, including adipose, pancreatic islets, liver and muscle.

Colocalized variant Locus (± 100kb)	GWAS trait	eQTL gene	eQTL tissue	PPC
BMI-AD
rs11668832 chr19:1712682-1912682	HC	ATP8B3	adipose	0.96
	HC	ABHD17A	adipose	0.69
	WC	ATP8B3	adipose	0.94
	WC	ABHD17A	adipose	0.73
	HDL	APOC4-APOC2	adipose	0.98
	HDL	APOC2	adipose	1.00
	HDL	APOC4	adipose	0.98
rs7254892 chr19:45289596-45489596	LDL	APOE	adipose	1.00
	LDL	APOC1P1	adipose	1.00
	logTG	APOE	adipose	1.00
rs6859 chr19:45282034-45482034	logTG	APOC1P1	adipose	1.00
	nonHDL	APOE	adipose	1.00
	nonHDL	APOC1P1	adipose	1.00
	TC	APOE	adipose	1.00
	TC	APOC1P1	adipose	0.93
T2D-AD
rs13257021 chr8:95865695-96065695	FG	TP53INP1	liver	0.84
	T2D	TP53INP1	liver	0.88
	FG	NDUFAF6	adipose	0.91
rs896854 chr8:95860511-96060511	T2D	NDUFAF6	adipose	0.67
	T2D	CCNE2	adipose	0.63
rs4311 chr17:61460763-61660763	DBP	ACE	adipose	0.98
	DBP	ACE	adipose	0.98
BMI-T2D-AD
	BMI	INO80E	liver	0.68
	HDL	INO80E	liver	0.84
	T2D	INO80E	liver	0.88
	WC	INO80E	liver	0.92
	WHR	INO80E	liver	0.95
	logTG	INO80E	liver	0.94
	BMI	DOC2A	adipose	0.90
	BMI	TMEM219	adipose	0.96
	BMI	RP11-455F5.3	adipose	0.60
	HC	KCTD13	adipose	0.96
	HDL	DOC2A	adipose	0.90
rs11150577 chr16:29933633-30133633	HDL	HIRIP3	adipose	0.66
	HDL	TMEM219	adipose	0.94
	logTG	TMEM219	adipose	0.75
rs12325539 chr16:29886525-30086525	T2D	PPP4C	adipose	0.75
	T2D	DOC2A	adipose	0.86
	T2D	HIRIP3	adipose	0.63
	T2D	TMEM219	adipose	0.88
	T2D	RP11-455F5.3	adipose	0.88
	WC	KCTD13	adipose	0.81
	WC	DOC2A	adipose	0.82
	WC	TMEM219	adipose	0.92
	WHR	DOC2A	adipose	0.97
	WHR	TMEM219	adipose	0.97
	WHR	HIRIP3	adipose	0.88
	T2D	INO80E	muscle	0.70
	T2D	ALDOA	islet	0.56
	T2D	INO80E	islet	0.72

PPC: Posterior Probability of Colocalization, HC: Hip circumference, WC: Waist circumference, HDL: High-density lipoprotein cholesterol, LDL: Low-density lipoprotein cholesterol, logTG: Log-transformed triglycerides, nonHDL: Non-high-density lipoprotein cholesterol, TC: Total cholesterol, FG: Fasting glucose, T2D: Type 2 diabetes, DBP: Diastolic blood pressure, BMI: Body mass index, WHR: Waist-to-hip ratio

Tissue expression and single-cell profiles reveal CNS regional differences and cell-type specificity

GTEx enrichment implicated brain regions in T2D loci with strong expression and splicing activity (Fig. S4, S7; Table S5). BMI-associated MIR4280 and T2D-associated MIR3150 (Table S6) mapped to regulatory activity in nucleus accumbens, cerebellum, cortex, and hippocampus. To examine CNS effects, we analyzed neurotypical GTEx data (13 brain regions), prefrontal cortex (PFC) and hypothalamus single-cell transcriptomics, and middle temporal gyrus (MTG) single-cell profiles from 5 AD donors (Methods).

Moderate bulk-level gene expression appeared primarily in cerebellum, cerebellar hemisphere and cortical regions (Fig. 7A). PFC excitatory neurons and hypothalamic endothelial cells were the primary expressing cell-types (Fig. 7B), contrasting with AD-genetic CNS microglia enrichment²⁷. Regional gene expression differed substantially. ACE, an AD drug target⁴⁶ and key driver of blood pressure regulation⁴⁷, is primarily expressed in hypothalamic endothelial cells but not the PFC (Fig 7B, C), linking the renin-angiotensin system to hypothalamic glucose homeostasis and AD risk⁴⁸. Expected pancreatic TMEM219 expression contrasted with minimal expression in heart, vascular muscle, or retina (Fig. S8), implicating TMEM219 in CNS-adipose-immune axis regulation of metabolic-immune AD susceptibility. AD MTG neuronal cells expressed select risk genes (COX7C, INTS8, NDUFAF6, DOC2A, KCTD13, YPEL3; Fig. 7D), consistent with PFC neurotypical data (Fig. 7B).

Overall, CNS expression of BMI, T2D, and AD risk effector genes is limited in both neurotypical and AD brain. These results support a systemic metabolic model of AD susceptibility.

Pathway enrichment indicates novel systemic biology

Pathway enrichment (GO, KEGG, Reactome; Fig. 8; Table S11; Methods) and interactome analyses identified overarching immune and RNA pathways. Pervasive mechanisms included cell cycle, adhesion, autophagy, lipid metabolism, MAPK signaling, DNA/mitochondrial biology and cation channels. BMI-AD loci are uniquely involved in viral biology, T2D-AD loci in renin-angiotensin pathways, and BMI-T2D-AD loci in GTPase, glucagon signaling, and cardiac biology (Fig. S9A–C).

Fig. 8 |. Enrichment and proteomic-based drug targeting analyses.

(A) GO cellular components enrichment analysis for all genes with functional annotation in the BMI-AD chr19 locus. (B) Individual gene enrichment analysis for ATP8B3. (C) Individual gene enrichment analysis for ABHD17A. (D) GO cellular components enrichment analysis for genes with coding annotation in the T2D-AD chr17 locus. (E) Individual gene enrichment analysis for KCNH6. (F) Individual gene enrichment analysis for TBX6. (G) Summary of compound categories based on proteomic drug atlas hits for all effector genes. Proteomic drug atlas enrichment analysis for (H) all effector genes (minimum 3 links per drug target); (I) BMI-AD effector genes; (J) T2D-AD effector genes; BMI-T2D-AD effector genes. Enrichment analyses’ false discovery rate (FDR) was set at 0.05.

The ATP8B3 locus (BMI-AD) enriched for synaptic and endosomal biology (Fig. 8A), with ATP8B3 linked to lipid biology and ion channels (Fig. 8B). These findings are of particular interest given that TRPML1, a lysosomal calcium channel, has been shown to revert synaptic decline through lysosomal reacidification⁴⁹. ABHD17A localizes to endosomes⁵⁰ to regulate postsynaptic activity via PSD-95 depalmitoylation⁵¹ (calcium-NMDA-dependent⁵²), modulate RAS-MAPK signaling⁵³ and calcium-activated potassium channels⁵⁴, unifying locus mechanisms. Both calcium channels⁵⁵ and depalmitoylases^56,57 are therapeutically interventionable targets.

The ACE locus (T2D-AD) enriched for cardiac and neuropeptide pathways (Fig. 8D). The same-locus effector gene KCNH6 associates with potassium channels (Fig. 8E), corroborating cation channels as promising druggable targets. In line, TBX6 shows brain-heart developmental links (Fig. 8F), with cardiac pathways represented across multiple loci.

To link systemic and CNS biology, we analyzed secreted protein interactors (Human Protein Atlas; Table S12) to identify angiotensinogen (AGT, ACE interactor) and insulin-like growth factor binding protein 3 (IGFBP3, TMEM219 interactor) as potential mediators. AGT (liver/kidney-predominant) regulates renin-angiotensin-mediated blood pressure and emerges as a cardiometabolic therapeutic target⁵⁸. IGFBP3, primarily produced in the liver, is upregulated in AD brain, interacts with humanin to prevent AD-related neuronal death⁵⁹, and its knockdown reverses senescence^63,64.

Overall, the broad evidence of molecular mechanisms spanning peripheral and central biological signatures support the systemic involvement of pathways mediating BMI and T2D risk in AD risk.

Genetic-backed drug target discovery reveals novel and repurposing candidates

We evaluated whether effector genes underlie ongoing therapeutic programs (Table 5). ACE targets multiple approved drugs and phase 3 trials for cardiometabolic conditions. Additional risk genes are in preclinical development for cancer, liver fibrosis, autoimmune disease, hepatic steatosis, stroke, and mitochondrial disorders.

Table 5 |.

Therapeutic programmes targeting effector genes driving shared BMI, T2D and AD genetics.

Gene	Status	Indication
ACE	Multiple approved drugs12 Phase 313	Hypertension, cardiovascular disease, heart failure, dyslipidemia, kidney disease
MAPK3	Preclinical14	Drug resistance, triple-negative breast cancer, solid tumors
NECTIN2	Preclinical15	Ovarian cancer immunotherapy
TAOK2	Preclinical16,17	Cancer
TMEM219	Preclinical18,19	Liver fibrosis, gastrointestinal autoimmune disorders
ATP8B3	Preclinical20	Colon adenocarcinoma immunotherapy
CCNE2	Preclinical21,22	Breast cancer
ABHD17A	Preclinical23	NRAS-mutant cancers
GDPD3	Preclinical24	Hepatic steatosis
TP53INP1	Preclinical25	Castration-resistant prostate cancer
FERMT2	Preclinical26	Prostate cancer
COX7C	Preclinical27	Stroke, mitochondrial disorders
MVP	Preclinical28	Pancreatic adenocarcinoma
PPP4C	Preclinical29	Ovarian cancer

To identify novel targets of interest, proteomic-based drug targeting using EnrichR-KG (Methods) identified 81 approved drugs modulating effector gene expression (p<0.05), primarily for cardiovascular disorders (heart failure, hypertension), cancer and neurological conditions (schizophrenia, multiple sclerosis) (Fig. 8G; Table S14).

We highlight [1] Guanfacine, a drug approved for ADHD and hypertension (Fig. 8H), [2] Clozapine, a drug approved for atypical antipsychotic and treatment-resistant schizophrenia (Fig. 8I), [3] Levosimendan, a calcium sensitizer⁶⁰ approved for heart failure (Fig. 8I) shown to rescue cognitive decline in mice and inhibit tau oligomerization, a major AD pathology⁶¹, [4] Tafluprost, a prostaglandin analog approved for open angle glaucoma and ocular hypertension (Fig. 8J), [5] INT-747, a farnesoid X receptor agonist with preclinical indications for insulin sensitizing, hypertension, bile acid homeostasis, hepatic and intestinal inflammation, liver fibrosis, and cardiovascular disease^62–64, and [6] A-317491, a P2X2/P2X3 antagonist with preclinical indications for neuropathic pain and chronic inflammation⁶⁵ (Fig. 8K). Of note, Resmetirom (thyroid hormone receptor-β agonist recently approved for NASH/liver fibrosis⁶⁶, Table S14) modulates the novel gene SCAMP4, an endocytic trafficking protein.

These findings demonstrate the genetically-backed repurposing potential of approved and pipeline drugs, identify novel targets, and leverage drug mechanisms to uncover risk biology.

DISCUSSION

Obesity and type 2 diabetes are major modifiable risk factors for AD, and their shared genetic architecture has the potential to unveil actionable biology relevant to the long prodromal curve of AD. We identify loci associated with BMI/T2D and AD, of which some demonstrate sex and age-specific associations. Besides brain involvement, peripheral immune and metabolic tissues and cell-types emerge as primary effectors of shared risk mechanisms. We explore multiple drug targeting approaches to expand the breadth of genetically-backed therapeutic options for potential AD prevention.

Our approach primarily reflects the sex-specific and age-stratified risk from variations in BMI and the pathobiology of T2D in AD development. All BMI-AD loci reported here are novel and exhibit consistent sex- and age-specific effects in women. The APOE locus has been linked to T2D-AD⁶⁷, but not directly for BMI-AD. We report a novel TP53INP1 lead variant (rs12546365) and male-specific ACE effects, expanding prior T2D-AD findings^68,67,69. Further, we identify TAOK2 as the first jointly associated locus with BMI, T2D, and AD.

We find that risk variants linked to lower BMI, and lower risk of T2D are generally linked to higher AD risk, in line with recent clinical evidence that compared people with underweight, normal weight and obesity to the risk of Aβ positivity⁷⁰. Genetics captures lifelong risk and can dissect disease reverse causation, unlike epidemiological studies that often capture the risk profile of comorbidities or risk factors late in adulthood and early in preclinical AD discourse. A critical consideration in our work is the potential for survival bias to confound results at the summary statistics data level. Individuals who develop AD may be depleted of genetic variants that predisposes to higher cardiometabolic risk. These individuals are selectively removed from the population before the age of AD onset, thus enriching AD GWAS for protective cardiometabolic alleles²⁴. This could obscure causal relationships, as apparent protective associations could reflect selective survival rather than true biological antagonism. Age-stratified BMI data partially addresses the potential bias, though the lack of necessary information to adjust for these biases at the aggregated summary statistics level could be further mitigated using individual-level data.

AD is often associated with weight loss⁷¹, BMI decrease in older adults is linked to higher risk of AD^72,73 and a common missense variant in APOE (rs429358) has been associated with longitudinal BMI decrease and weight loss⁷⁴. Nonetheless, we find that BMI-AD loci effects are independent of BMI direction but driven by underlying cardiometabolic associations. Genetically-driven low blood pressure and low glucose levels link to elevated AD risk, directly opposing the hypothesis motivating trials of GLP1R agonists for AD, aligning with observed lack of clinical efficacy in recent Phase III trials despite key biomarker engagement⁷⁵. Cross-sectional and longitudinal studies substantiate an increased prevalence of AD and dementia in older individuals with low blood pressure^76–78, but with higher blood pressure in younger individuals⁷⁹. Antihypertensive clinical trials are inconclusive⁸⁰ and only recently genetic evidence was linking hypertension to AD in multivariate Mendelian randomization studies²⁴ whereas multi-trait GWAS failed to identify strong links^,81. Likewise, our findings that low glucose levels link to elevated AD risk align with AD patients showing strong glucose hypometabolism and insulin resistance in the brain^82,83. As increased fasting plasma glucose levels directly associate with higher brain glucose concentrations⁸⁴, our results suggest that lower peripheral glucose levels may potentiate long term cerebral energy deficits.

Cardiometabolic risk research in AD has primarily focused on T2D and AD, but our findings show that BMI-AD biology is more specific to AD while T2D-AD is more ubiquitous, in line with recent observations^15,85. We find consistent evidence that women are at higher risk from cardiometabolic outcomes associated with variations in BMI and men from cardiometabolic outcomes associated with T2D, in line with recent longitudinal findings⁸⁶. These observations align with higher prevalence of obesity in women¹⁷ and higher susceptibility to AD⁶. Our findings further indicate that women suffer more severe consequences associated with BMI variations driven by APOE, yet current approved drugs are less effective in women and APOE4 carriers⁹. Consistently, T2D and Parkinson’s disease (PD) share a stronger correlation compared to T2D and AD⁸⁷, with PD being substantially more common in men⁸. For example, women show superior insulin sensitivity compared to men, but insulin resistance is not a relevant phenotype in our loci. On the other hand, clinical trials targeting the male-specific ACE locus (T2D-AD) show that men have significantly higher baseline plasma ACE activity than women and benefit more when taking ACE inhibitors for hypertension⁴⁴. These considerations should prompt the field to examine sex-stratified data for clinical trials targeting genes or pathways evidencing sex-dysmorphic profiles.

Deciphering the genetic architecture and underlying biology of AD has been limited to the brain and its constituent cell-types. While T2D-related loci show brain associations, peripheral immune and metabolic tissues, not brain or microglia, primarily drive BMI and T2D-associated AD risk. This aligns with recent findings that APOE4 drives immune-related proteomic signatures enriched in circulating immune cells, directly linking AD’s primary genetic risk factor to peripheral immunity⁸⁸.

Human genetic evidence boosts the likelihood of drugs’ clinical success by 2.6-fold⁸⁹, with GWAS variants pointing to highly successful drug targets⁹⁰. We identify multiple genetically-backed drug targets with vast repurposing and novel development potential. Even drug targets unrelated to the input cardiometabolic phenotypes are experiencing a substantial repurposing push for neurological disorders, exemplified by immunomodulatory cancer drugs like Lenalidomide currently in clinical trial for AD⁹¹.

Limitations of our work include the current unavailability of sex-specific AD data that may underestimate joint sex-specific genetic effects of BMI, T2D, and AD. The upcoming AD meta-analysis will provide large-scale sex-stratified data, though preliminary results appear underpowered⁹². Moreover, HyPrColoc stringency ensures reliability but may miss other true shared genetic signals. Also, assessing genome-wide significance alone may miss key insights for AD, exemplified by genetic variation linked to GLP1R driving obesity therapeutics but reaching BMI significance only in upcoming GWAS (5M+, unpublished).

Our findings enable informed therapeutics targeting the shared genetic blueprint of BMI, T2D, and cardiometabolic-driven AD risk. Blood pressure interventions warrant more refined sex, age-range and disease stage eligibility criteria. Given prior focus on T2D-AD biology, BMI-related traits may yield outcomes more targeted to AD. The identification of peripheral tissues, not brain, as primary effectors of cardiometabolic risk in AD opens novel research directions for AD risk mitigation via genetically-backed candidates.

METHODS

GWAS summary statistics curation

All BMI and T2D summary statistics (sex-combined and sex-stratified) were originally sourced in the human reference genome GRCh37 (HG19) while the AD1, AD2 and AD3 summary statistics were made available in the human reference genome GRCh38 (HG38). Given that conversion rates are higher from HG19 to HG38⁹³, we proceeded to liftover the BMI and T2D GWAS data from HG19 to HG38 using the “rtracklayer” R statistical software⁹⁴ (version 4.2.2) package from the Bioconductor project⁹⁵. BMI, T2D, AD2 and AD3 summary statistics required reference SNP cluster ID (rsID) annotation. Altogether, a total of ≈52.5 million chromosome positions were annotated using the “BSgenome” R statistical software (version 4.2.2) package from the Bioconductor project⁹⁶. The APOE locus was retained in all analyses, reflecting a methodological consideration specific to our study design. The three AD meta-analyses used here (AD1, AD2, AD3) differ systematically in case ascertainment and control matching strategies. Regional APOE exclusion would result in unpredictable interactions with these pre-existing differences. Further, we aimed to evaluate sex-specific differences in cardiometabolic risk associated with the APOE region given the differential drug responses for women and APOE4 carriers⁹. To address potential APOE dominance concerns, we report: [1] full transparency of APOE findings (NECTIN2-TOMM40 locus), including novel BMI-AD associations with female-specific effects; and [2] protein-protein interaction analyses to assess non-APOE loci independence from APOE biology.

Heritability and genetic correlation

We leveraged LD score regression⁹⁷ (LDSC) to estimate narrow-sense heritability (additive genetic effects of common variants) and genetic correlation (GC) between AD, BMI and T2D datasets. We leveraged precalculated LD scores calculated in individuals of European ancestry from the 1000 Genomes Project (eur_w_ld_chr). Standard GC data curation and quality control procedures were followed^97,98. We tested the differences between trait-pair genetic correlation estimates using

$$z=\frac{r_{g1}-r_{g2}}{\sqrt{S{E}_1^2+S{E}_2^2-2r\cdot S{E}_1\cdot S{E}_2}}$$

where rg is the genetic correlation for each trait-pair, SE are the standard errors, and r is the genetic correlation coefficient estimate between the different traits in both trait-pairs¹⁰¹. The genetic correlation coefficient r between non-AD traits accounts for their underlying genetic similarity. This parameter is key for controlling whether observed differences in genetic correlation estimates (rg) reflect true significant biological differences rather than shared genetic architecture of the traits themselves. For example, when comparing rg estimate differences between BMI.Male-AD1 and BMI.Female-AD1, r represents the genetic correlation between BMI.Male and BMI.Female, effectively quantifying their genetic similarity and correcting for it in statistical testing. We assumed a Bonferroni corrected p-value threshold of (0.05/27=1.85x10⁻³).

Variant discovery

We conducted genetic overlap analyses using Python⁹⁹ (version 3.9.13) between AD, BMI and T2D datasets in HG38. We defined the significance thresholds (p-value) of 5x10⁻⁸ for AD datasets and 5x10⁻⁴ for all BMI and T2D datasets. The use of asymmetric significance thresholds reflects the distinct roles these traits play in the etiology of AD. The biological expectation that variants affecting intermediate or risk-factor phenotypes (BMI, T2D) through additive cascade mechanisms may exert smaller individual effects than those affecting disease endpoints (AD), and therefore may not reach genome-wide significance in individual cross-section GWAS but remain relevant in shared genetic architecture. The GLP1R locus exemplifies this principle: variants linked to GLP1R reach genome-wide significance for T2D¹⁰⁰, but more modest associations in BMI²⁵. Yet, therapeutic GLP1R agonists produce substantial body weight reduction¹⁰¹, demonstrating that variants below genome-wide significance for intermediate or risk-factor phenotypes can have substantial biological and therapeutic relevance. For AD, we applied the standard genome-wide significance threshold to minimize the propagation of systematic biases inherent to the use of AD-proxy phenotypes in all contemporary GWAS^24,102. We then performed phenotype overlap on chromosome position, including all pair-traits and triangulations. After variant overlap, we performed colocalization analysis was performed HyPrColoc, a Bayesian method for identifying shared genetic associations between traits within genomic regions. HyPrColoc analysis was conducted using GWAS summary statistics, including regression coefficients and standard errors for each variant and trait. The method tests whether multiple traits share causal variants in a genomic region and, when evidence for full colocalization is insufficient, employs a clustering algorithm to identify subsets of traits that colocalize at distinct causal variants. Default parameters were used for the analysis. We adopted a hypothesis-generating approach, without pre-specifying a posterior probability threshold. Instead, we [1] performed HyPrColoc across the full posterior probability distribution, [2] performed sensitivity analyses to assess the robustness of identified colocalization clusters and [3] validated candidate colocalization signals through visual inspection of local Locuszoom plots to confirm signal, LD structure and regional architecture. This approach prioritizes identification of robust candidate signals for hypothesis generation while minimizing the risk of threshold-driven false negatives. Candidate signals identified at lower posterior probabilities but persistent across sensitivity analyses and confirmed via Locuszoom inspection were deemed biologically plausible for further investigation. All analyses were performed using the HyPrColoc R package¹⁰³. The HLA region (chr6:28,510,120-33,480,577) was removed post-analysis. Locuszoom plots were generated using a package that creates regional Manhattan plots leveraging an european linkage disequilibrium reference panel (https://github.com/Geeketics/LocusZooms). We tested the differences between women-specific and men-specific beta-estimates (b_women, b_men) using:

$$t=\frac{b_{men}-b_{women}}{\sqrt{S{E}_{men}^2+S{E}_{women}^2-2r·S{E}_{men}·S{E}_{women}}}$$

where SE are the standard errors and r is Spearman rank correlation coefficient across all SNPs across BMI and T2D strata. We assumed a Bonferroni corrected p-value threshold of (0.05/12=4.2x10⁻³).

Phenome-wide association analysis

Phenome-wide association studies (PheWAS) were conducted using the Association to Function Knowledge Portal (A2F; a2fkp.org) to systematically test genetic variants for associations across a broad spectrum of phenotypes. We analyzed a total of 1,314 distinct phenotypes, encompassing diseases, disorders, and clinical traits derived from electronic health records and biobank data. We screened PheWAS associations for the seven colocalized loci aligned to the risk increasing allele for AD. To account for multiple testing, a statistical significance was determined using a Bonferroni-corrected p-value threshold of 4.3x10⁻⁵ (0.05/1,314 phenotypes). Results meeting the corrected significance threshold were considered evidence for phenome-wide significant associations.

Variant-to-gene mapping

For establishing connections between genetic variants and their corresponding target genes, we utilized the cS2G methodology¹⁰⁴. The analysis was conducted using the GRCh37/hg37 genome assembly. The cS2G approach integrates various SNP-to-gene association methods by incorporating functional annotations derived from seven component strategies, listed according to their assigned weights: Exon regions, Promoter sequences, fine-mapped cis-eQTL data from eQTLGen blood studies, fine-mapped cis-eQTL information from GTEx across 54 human tissue types, enhancer-gene associations from EpiMap spanning 833 cellular types, Activity-By-Contact (ABC) methodology covering 167 cellular types, and Cicero analysis for blood/basal conditions. The cS2G framework computes an integrated association score by applying weights to individual strategy scores, thereby identifying the most probable effector gene(s) for each single nucleotide polymorphism (SNP).

AlphaGenome

We performed an in-silico perturbation screen using batch variant scoring from AlphaGenome⁴⁰ to predict the functional consequences of colocalized BMI/T2D-AD variants across multiple functional domains. The effect of each colocalized variants was predicted within a 100kb genome window in humans. We used “UBERON” and “CL” ontologies to predict variant effects on up to 116 tissues and 106 cell-types, respectively. The colocalized variants within the APOE region were not considered in this analysis given their overwhelming predicted effects across all data-types. Importantly, neither data-type in AlphaGenome included central nervous system microglia. For each variant, we report the raw scores representing the predicted regulatory effect for each data-type. Raw scores are specific to individual data-types, thus not comparable between different them. We present variant effects by ordering raw scores on magnitude of effect from highest to lowest, thus identifying the tissues and cell-types where predicted variant impacts are strongest. The raw scores are shown as absolute values due to current limitations of deep neural network models in accurately predicting the direction of variant effects¹⁰⁵. To contextualize the magnitude of variant effects, we report quantile scores which provide a percentile-based ranking of variant effects relative to the background genome (Table S8). The quantile score is computed by establishing an empirical background distribution using scores from common variants (minor allele frequency > 0.01 in any GnomAD v3 population) for each data-type. For example, a variant with a quantile score of 0.99 indicates that its predicted regulatory effects in a given tissue or cell-type ranks at the 99^th percentile among common variants. Variant effect predictions are evaluated across the following functional scorers:

Data Type	Description	Biosamples	Resolution	Category
Total RNA-seq	Measures RNA expression across all RNA types (polyadenylated, non-polyadenylated, and non-coding RNAs).	285	Single bp	Transcriptomic
PolyA+ RNA-seq	Measures RNA expression of mature polyadenylated transcripts enriched for polyA+ tails.	285	Single bp	Transcriptomic
CAGE	Measures RNA expression at transcription start sites (TSS) using 5’ cap capture.	264	Single bp	Transcriptomic
PROCAP	Measures nascent RNA expression at TSS via run-on sequencing with capping.	6	Single bp	Transcriptomic
DNase-seq	Measures chromatin accessibility through DNase I hypersensitive site mapping.	305	Single bp	Epigenomic
ATAC-seq	Measures chromatin accessibility through transposase insertion patterns.	167	Single bp	Epigenomic
ChIP-seq (Histone Modifications)	Measures relative abundance of histone modification marks (24 histone markers).	219	128 bp	Epigenomic
ChIP-seq (Transcription Factors)	Measures DNA-bound transcription factors (43 proteins) via chromatin immunoprecipitation.	163	128 bp	Epigenomic
Splice Sites	Predicts location of donor and acceptor splice sites on both DNA strands with probability scores.	-	Single bp	Splicing
Splice Junctions	Predicts spliced read counts at exonexon junctions from RNA-seq data.	282	Single bp	Splicing
Splice Usage	Predicts fraction of transcripts using a given splice site from RNA-seq data.	282	Single bp	Splicing

GWAS-eQTL colocalization

We colocalized the lead cross-phenotype colocalized variants (from HyPrColoc) plus its credible sets (± 100kb) with eQTLs in metabolic tissues, including adipose, pancreatic islets, liver and muscle (build 37) using COLOCUS, a platform from HuGeAMP. COLOCUS implements the COLOC Bayesian algorithm, utilizing conditionally independent or fine-mapped signals identified through statistical fine-mapping approaches, with each signal represented by a lead variant and corresponding credible set.

Immune, adipose and liver analyses

We evaluated the individual gene expression and % of cells in the immune compartment of 17 peripheral tissues using the Single Cell Portal¹⁰⁶, including ileum, lung, thoracic lymph node, mesenteric lymph node, bone marrow, skeletal muscle tissue, liver, spleen, omentum, caecum, thymus, duodenum, blood, transverse colon, sigmoid colon, lamina propria and jejunal epithelium⁴². Statistical sex-strata differences were evaluated using a Wilcoxon matched-pairs signed rank test (nonparametric) following the evaluation of data normality using the Anderson-Darling, D’Agostino-Pearson omnibus, Shapiro-Wilk and Kolmogorov-Smirnov normality tests. Outliers in Fig. S5D–E were identified using the Robust Regression and Outlier removal Using Nonlinear TukeyBiweight (ROUT) method (Q = 1%) implemented in GraphPad Prism (version 9.0.0). This method applies robust regression with a false discovery rate (FDR) of 1% to identify data points as outliers. We checked the single-cell expression in the human adipose tissue using the adipose tissues knowledge portal¹⁰⁷ and the single-cell expression in the human liver¹⁰⁸ using the liver cell atlas.

Enhancer-to-gene

We adapted the E2G pipeline from Perrin et al.⁴⁵ to map shared risk variants falling within regulatory genomic regions of human adipose tissue, liver, adipocytes and hepatocytes. The analysis was conducted in hg37. For the adipose tissue and adipocytes, we leveraged promoter capture HiC data from SGBS cells (both pre- and mature adipocytes)¹⁰⁹, RNA-seq, ATAC-seq and ABC-STARE data from SGBS cells (days 0, 4 and 14)⁴⁵, human adipose tissue ATAC-seq data from METSIM (10,197 Finnish men)¹¹⁰ and eQTL data from AdipoExpress (2344 human subcutaneous adipose tissue samples)¹¹¹. We further supplemented E2G with peak annotations from ChIP using the ChIPseeker R package¹¹², Cis-regulatory element annotations from single-cell epigenetic datasets using the Cis-element Atlas (CATlas)¹¹³, and Roadmap epigenomics data from samples E023 (mesenchymal stem cell (MSC) derived adipocyte cultured cells) and E063 (adipose nuclei)¹¹⁴. E023 derives from in vitro differentiated MSCs into adipocytes and models adipocyte development. E063 consists of nuclei isolated directly from adult adipose tissue and reflects the native epigenetic state of in vivo mature adipocytes. For the liver and hepatocytes, we leveraged promoter capture HiC data from the liver¹¹⁵, RNA-seq and ATAC-seq data from primary human hepatocytes¹¹⁶ and eQTL data from the most recent liver eQTL meta-analysis¹¹⁷. We further implemented ATAC peak annotations¹¹² using the annotatr and annotatePeak R packages and Roadmap epigenomics data from liver samples. We complemented E2G with e/sQTL data from GTEx¹¹⁸ (version 10).

Brain analysis

We performed single-cell analyses using the Single Cell Portal¹⁰⁵ in the prefrontal cortex¹¹⁸ (PFC, key AD region¹¹⁹) and hypothalamus¹¹⁹ (key region in obesity, T2D and metabolism^48,120). We used the Allen Brain Atlas¹²⁰ to perform exploratory analyses of the single-cell diversity of the middle temporal gyrus of individuals with AD (MTG 10x SEA-AD; vulnerable AD brain region¹²¹) (166,868 total nuclei from 5 post-mortem human AD brain specimens, median age of 80)¹²¹.

Enrichment analysis

We performed enrichment analyses in GO Biological Processes, GO Cellular Components, GO Molecular Functions, KEGG and GTEx tissues³⁸ using ShinyGO⁴¹ (v0.85) and an FRD cutoff of 0.05. The GTEx tissue gene expression profiles dataset contains gene expression data from 17,382 RNA-seq samples of 54 different human tissues (plus 2 cell lines) from 838 postmortem donors. We performed enrichment analyses based on the “Co.Expression GTEx.Tissue Expression.Down,”, “Co.Expression GTEx.Tissue Expression.Up.” and “Co-expression.GTEx.Tissues.V8.2023” categories.

This analysis evaluates whether shared risk genes are enriched within genes showing elevated or decreased expression in specific tissues. Additionally, we performed an over-representation analysis based on GO terms, KEGG, Reactome and Proteomic Drug Atlas using EnrichR-KG¹²². Of note, the proteomic drug targeting was originally performed on human immortalized cell lines¹²³.

EADB ACKNOWLEDGMENTS

The EADB consortium groups together 20,464 Alzheimer’s disease (AD) cases and 22,244 controls after quality controls from 16 European countries (Austria, Belgium, Bulgaria, Czech Republic, Denmark, Finland, France, Germany, Greece, Italy, Portugal, Spain, Sweden, Switzerland, The Netherlands and the UK). These samples were genotyped in three independent centers (France, Germany and the Netherlands) leading to define three nodes: EADB-France, EADB-Germany and EADB-Netherlands. In addition, EADB also included Australian partners. We thank the numerous participants, researchers, and staff from many studies who collected and contributed to the data. We thank the high performance computing service of the University of Lille. We thank all the CEA-CNRGH staff who contributed to sample preparation and genotyping for their excellent technical assistance. This work was supported by a grant (European Alzheimer DNA BioBank, EADB) from the EU Joint Programme – Neurodegenerative Disease Research (JPND). Inserm UMR1167 is also funded by Inserm, Institut Pasteur de Lille, the Lille Métropole Communauté Urbaine, the French government’s LABEX DISTALZ program (development of innovative strategies for a transdisciplinary approach to Alzheimer’s disease). The Bulgarian cohort was partially supported by the European Union-NextGenerationEU, through the National Recovery and Resilience Plan of the Republic of Bulgaria, project BG-RRP-2.004-0004-C01, Strategic research and innovation program for development of Medical university - Sofia.