Data were obtained from the ADNI database¹ (Okonkwo et al., 2025). The ADNI was launched in 2003 as a public-private partnership, led by Principal Investigator Michael W. Weiner, MD. The original goal of ADNI was to test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other biological markers, and clinical and neuropsychological assessment can be combined to measure the progression of mild cognitive impairment (MCI) and early AD. The current goals include validating biomarkers for clinical trials, improving the generalizability of ADNI data by increasing diversity in the participant cohort, and to provide data concerning the diagnosis and progression of AD to the scientific community.

We downloaded the demographic, various longitudinal clinical assessment and biomarker data from ADNI data portal. Genetic data were obtained from the Alzheimer’s Disease Sequencing Project (Leung et al., 2025) (ADSP).² All genetic data were first converted into plink format using PLINK (Chang et al., 2015) for principal component analysis (PCA) of population structure. The plink genetic file on all the autosome chromosomes was used for PCA. We removed variant sites with missing data > 5%, minor allele frequency <5%, or significantly deviating from Hardy-Weinberg Equilibrium (HWE) (plink –geno 0.05 –maf 0.05 –hwe 1e-6). We further pruned the variants to exclude sites exhibiting high levels of linkage disequilibrium (LD), using the command “plink –indep-pairwise 50 5 0.2.” In addition, variants located in regions known to be under recent selection (Anderson et al., 2010) were removed from the dataset. PCA was performed in PLINK for the first 20 components, and based on the Scree plot, first five PCs were included in the subsequent association analysis.

We considered the alternative allele counts (dosages) of all three SNP sites to account for the diversity of the haplotypes and each variant’s contribution to the targets. The genotypes were extracted from vcf format using bcftools (Danecek et al., 2021) and converted into csv format. Cognitive tests and imaging biomarker data included clinical dementia rating (CDR, all six domains), PET-based amyloid loads (Centiloids), composite PET Tau scores (Braak staging converted from threshold-based tau-positivity in brain regions), and structural T1 MRI measurements from the ADSP Phenotype Harmonization Consortium (ADSP-PHC) (Hohman, 2023) harmonized datasets (release 3). The final dataset consists of 2,168 subjects with the demographic information reported in Table 1. All the data were read into R, where linear mixed effects (LME) models were built using the package “lmerTest”(Kuznetsova et al., 2017), comparing effects of the alternative allele counts (dosages) of the three SNPs (collectively referred to g1, g2, and g3 as shown in Figure 1A) to the longitudinal changes of the target variable with an interaction term for age, while controlling for several covariates such as sex, years of education, APOE4 allele count (the largest known genetic risk for AD), and the first five PCs, with the following Equation 1 in R programming language:

where RID is the unique subject ID.

We validated the results of this analysis in NACC (National Alzheimer’s Coordination Center)³ (Beekly et al., 2007), with longitudinal data from over 42 current and former Alzheimer’s Disease research centers (ADRCs) across the US. Demographics, APOE4 allele count, and longitudinal clinical cognitive assessments (CDR) were obtained from Uniform Data Set (UDS) V3. Genetic data were obtained from the ADRC GWAS Datasets (ADC1-15) profiled by genotyping SNP array and harmonized by Alzheimer’s Disease Genetics Consortium (ADGC) (Naj et al., 2011). Samples closer than second-degree were removed in PLINK (–king-cutoff 0.0884). We only kept subjects with > 60 years of age at the first visit. Variant sites were pruned similarly as ADNI data for PCA. Based on the PCA, first 10 PCs were used in the LME models, where the targets were CDRs in six domains and the global and sum of boxes scores. The final dataset consists of 21,950 subjects with the demographic information reported in Supplementary Table 1.

We performed the eQTL analysis based on the Accelerating Medicines Project for Alzheimer’s Disease (AMP-AD) (Hodes and Buckholtz, 2016) for the subjects within the Religious Orders Study/Memory and Aging Project (ROSMAP) (Bennett et al., 2018). All the demographic, clinical and pathological data of the ROSMAP cohort were obtained from the Rush Alzheimer’s Disease Center Research Resource Sharing Hub,⁴ upon approval of the data-usage agreement. Genetic data were from the variants called on whole genome sequencing (WGS) data from the AMP-AD data portal.⁵ Additional genetic data were obtained from the AMP-AD diverse cohorts study.⁶ The genetic data were merged and filtered similarly to what was done for the ADNI subjects, and PCA was performed. For gene expression data, we downloaded post-mortem gene expression data (RNA-seq) from three brain regions including dorsolateral prefrontal cortex (DLPFC), posterior cingulate cortex (PCC), and head of caudate nucleus (HCN) for the ROSMAP cohort from the AMP-AD data portal⁷ (Bennett et al., 2018). The gene expression data used was the filtered, normalized, and residualized counts to limit the effects of technical artifacts. After combining gene expression and genetic data, the final dataset consists of 947 unique individuals reported in Table 2.

We identified cis-eQTLs by linear regression of the expression of all the genes on chromosome 7 within 1 Mb of the three SNPs (chr7: 140,792,804–142,973,545, 65 genes) with the alternative allele count of the three SNPs for each brain region using the following Equation 2:

Covariates include technical covariates [RNA integrity number (RIN), post-mortem interval (PMI), sequencing batch (batch)], various biological covariates (age, sex, years of education, APOE4 allele count), and the first five PCs based on the PCA of the genetic data. Significant genes (e-genes) with FDR-corrected p< 0.05 were then identified as eQTL of the SNP. The three SNPs were assessed individually for their corresponding e-genes. Semi-quantitative measurements of neuropathologies (Braak staging or CERAD score, Lewy body, TDP-43, and other vascular related neuropathologies) were added as covariates when they were additionally considered.

We carried out additional validations for the identified eQTL in two external cohorts from the AMP-AD study: the Mayo RNA-seq (MAYO) cohort (Allen et al., 2016) and the Mount Sinai Brain Bank (MSBB) cohort (Wang et al., 2018). The data were obtained from the same synapse entry as ROSMAP. The subject demographics are reported in Supplementary Tables 2, 3, respectively. In MSBB, expression data were available for four brain regions: frontal pole (FP), inferior frontal gyrus (IFG), parahippocampus (PHG), and superior frontal gyrus (STG). For MAYO, two regions were profiled: temporal cortex (TCX) and cerebellum (CBE). Each region was assessed, respectively for the eQTL. Neuropathology data for these two cohorts were downloaded from synapse entry syn27000096.

We performed causal inference test (CIT) to identify genes significantly affected by the strongest eQTL rs10246939 (g1) -MGAM using the RNAseq data of 502 AD subjects (defined as niareagansc < 3, i.e., NIA-Reagan diagnosis of AD (Neurobiology Aging, 1997) high or intermediate) from DLPFC of ROSMAP. CIT has been well-described previously (Millstein et al., 2009). In short, it offers a hypothesis test for whether a molecule (in this case MGAM) is potentially mediating a causal association between a DNA locus (g1, i.e., rs10246939), and some other quantitative trait (such as the expression of genes correlated with MGAM and rs10246939). Causal relationships can be inferred from a chain of mathematic conditions, requiring that for a given trio of loci (L i.e., rs10246939), a potential causal mediator (G, i.e., MGAM) and a quantitative trait (T, i.e., some other genes), the following conditions must be satisfied to establish that G is a causal mediator of the association between L and T:

We used the R software package “cit” (Millstein et al., 2016) to perform the causal inference test, calculating a false discovery rate using 1,000 test permutations. Trios with a Q-value (FDR) < 0.05 were classified as significant, and the associated T genes were considered downstream of MGAM. We investigated the significant gene hits for functional annotation and enrichment in knowledge databases such as KEGG (Kanehisa et al., 2023) pathways and Gene Ontology (Gene Ontology Consortium et al., 2023) terms through Metascape (Zhou et al., 2019).

To examine the effects of MGAM inhibiting drugs (acarbose or miglitol, Anatomical Therapeutic Chemical (ATC) code A10BF) on AD risk, we downloaded the comprehensive medical records including medication data from longitudinal assessments at NACC (National Alzheimer’s Coordination Center) (see text footnote 3; Beekly et al., 2007), with data from over 42 current and former Alzheimer’s Disease research centers (ADRCs) across the US. Demographics, APOE4 allele count, longitudinal clinical cognitive assessments, comorbidities, and medication data were obtained from Uniform Data Set (UDS) V3.

The procedures used to create a cohort of 1:1 matched pair of ever users and never users of A10BF are shown in Figure 1C. We chose subjects with type II diabetes mellitus (T2D) based on diabetes diagnosis or use of antidiabetic medications at the first visit. Subjects with any dementia or PD diagnosis at the first visit were excluded. We matched the subjects who took MGAM inhibitors at any time in the follow-up period (ever-users) with another member of the never-user group with propensity score matching, using the R package “matchit”(Ho et al., 2011) by the following command (Equation 3):

We considered the following factors in the matching: age at first visit, sex, race, years of education, APOE4 allele count, follow-up duration, total number of visits, and whether any other antidiabetic drugs taken in the duration of follow-ups, classified by their corresponding ATC subcategories A10A (insulin and analogs), A10BA (biguanides), A10BB (sulfonylureas), A10BG (thiazolidinediones), A10BH [dipeptidyl peptidase 4 (DPP-4) inhibitors], A10BJ [glucagon-like peptide-1 (GLP-1) analogs], A10BK [sodium-glucose co-transporter 2 (SGLT2) inhibitors], and A10BX (other blood glucose lowering drugs). Since it has been reported that A10BF could have joint effects with metformin (A10BA) and pioglitazone (A10BG) in reducing dementia incidences (Tseng, 2020), they were set as exact match together with insulin (A10A), sex, and APOE4 allele count. To reduce overfitting in high-dimensional settings, we chose “glmnet” as the distance method to estimate propensity scores using L1/L2-regularized logistic regression (Friedman et al., 2010). Each individual in the user group was thus matched to the closest individual in the non-user group.

The final dataset contained 76 subjects in total, with 38 subjects taking MGAM inhibitors acarbose or miglitol, matched with non-users by a ratio of 1. Matching balance was evaluated with their subject characteristics reported in Table 3. Standard mean difference (SMD) was calculated for the two groups using the R package “tableone” (Yoshida and Bartel, 2022).

Using the time elapsed at first dementia diagnosis (DEMENT = 1), we performed survival analysis using the R package “survminer” (Kassambara et al., 2024) to examine the group difference in dementia susceptibility between ever-users and never users, and p-value was computed from the χ² statistic in a log-rank test. Kaplan-Meier plots were created to present time to dementia diagnosis. To compare the drug’s effect on cognitive performances between the two groups, we built LME models to assess the group difference in multiple domains of CDR, by the following Equation 4:

For robustness, we repeated the process with matching ratio set to 2 and 3, respectively.

Descriptive statistics for each study were reported for all subject characteristics using means and standard deviations for continuous variables and counts for categorical variables. Continuous variables were compared using a paired t-test between matched groups to check match balance. Categorical variables were compared using the Fisher’s exact test. P-values from other models were obtained from each respective package. All the analyses were performed in R (v4.2.2) statistical language.

From LME modeling of the longitudinal change in cognitive assessments and biomarkers in ADNI (n = 2,168), highly significant interactions (P < 0.001) were found between age and the alternative allele counts of each of the SNPs. Their contributions to the target variables are also significant (P < 0.001) yet in different directions (directed logP values and standardized coefficients reported in Figure 2, and raw effect size reported in Supplementary Table 4). This trend could be seen across different types of target variables, including multiple clinical and imaging biomarkers. The alternative alleles of g1 and g2 are found to have a negative effect on cognitive function in their interaction terms with age, while g3 has a positive effect; thus, with their comparable effect sizes, supertasters who have two copies of the alternative alleles at g1 and g3 (PAV haplotype, allele count combination 202) will have a lower negative effect on the cognitive performance thus slower cognitive decline when compared to non-tasters with only two copies of the g2 alternative allele (AVI haplotype, allele count combination 020). Similar results were also found in imaging biomarkers from structural MRI and PET imaging data, especially for left lateral ventricle, a prominent biomarker for AD (Apostolova et al., 2012; Figure 2A). The results were consistent although less significant in PET imaging for amyloid or tau burden (Supplementary Figure 1 and Supplementary Table 5) due to smaller sample size (n = 1,767 for amyloid, n = 870 for Tau) with limited number of visits (mean = 2.2 for amyloid, 1.7 for Tau).

The results were additionally validated using longitudinal data from NACC. There is more genetic diversity in the NACC study with larger sample size, yet the subjects possess more heterogeneous health conditions causing cognitive impairment, thus the effect size for each of the SNPs is smaller. Nevertheless, the supertaster allele was still found to be strongly associated with slower cognitive decline, surpassing the generally accepted genome-wide association study (GWAS) significance threshold (P < 5e-8) (Supplementary Figure 2).

The e-genes from cis-eQTL for each of the alleles were reported in Table 4, respectively, for all three regions (DLPFC, HCN, and PCC) from ROSMAP cohort. All three alleles are associated with multiple e-genes. In total, five genes were found to be significant across one or more brain regions in addition to DLPFC (WEE2-AS1, MGAM, TAS2R5, CLEC5A, and ENSG00000270157). Among them, WEE2-AS1 and MGAM show the strongest association with the g1 locus (adjusted P < 1E-10). When neuropathologies were added to the model, the expression of MGAM was also found to be significantly correlated with Braak staging (P = 0.028), where higher Braak staging associated with higher MGAM expression (Figure 3A). The association was only significant in DLPFC and not HCN or PCC (Table 4).

We further validated the eQTL relationship as well as associations of MGAM’s expression with neuropathology in different brain regions using the harmonized RNA-seq data from two additional cohorts of AMP-AD. Fewer e-genes were identified due to their smaller sample sizes. In MSBB, among the four brain regions (FP, IFG, STG, and PHG), eQTLs were identified only in IFG and STG, with MGAM as an e-gene in IFG. Interestingly, MGAM expression was also significantly associated with Braak staging in FP, PHG and STG (Supplementary Figure 3 and Supplementary Table 6). In MAYO, g1 and g2 are in complete linkage disequilibrium, so their effects are exactly the same in opposite direction. Their allele count and Braak staging were both significantly associated with MGAM expression in TCX, another AD vulnerable brain region (Supplementary Figure 4 and Supplementary Table 7). Overall, alternative allele of the most significant locus g1 was found to be inversely correlated with MGAM expression, while neurofibril tangles were associated with increased MGAM expression in several AD affected brain regions, suggesting that lower MGAM expression confers lower AD risk for the supertasters and vice versa.

We further investigated the potential mechanistic role of TAS2R38 variants in the pathogenesis of AD through CIT analysis. Since the rs10246939 (g1) locus is most significantly associated with MGAM expression, we focused on the g1-MGAM network for the analysis. We identified 219 significantly affected downstream genes (FDR < 0.05, Supplementary Table 7), many of which encode for multiple units in the mitochondrial protein complex, such as VDAC3, VDAC4, CYC1, and UQCRC1 as well as 26S proteasome units such as PSMC4 and PSMD3. In addition, it was found that the expression of CHRNA7 (neuronal acetylcholine receptor subunit alpha-7), a major player of signal transmission at synapses, had been negatively impacted in the network. Molecular functional enrichments of the eQTL network revealed the disrupted gene modules to be implicated in AD (KEGG: hsa05010), learning or memory (GO: 0007611), and cognition (GO: 0050890) (Figure 3B and Supplementary Table 8). Some impacted genes are also associated with other neurodegenerative diseases (KEGG: hsa05022) such as PD (PSMC4, PSMD3, PRKN, CYC1, VDAC3, and UQCRC1) and Huntington’s disease (CYC1, PSMC4, PSMD3, UQCRC1, and VDAC3), indicating their roles in neurological functions (Figure 3C). The disrupted genes were negatively affected by increased MGAM expression, suggesting that higher MGAM expression drives protein degradation failure and mitochondrial damage, ultimately leading to synaptic loss.

Given MGAM’s role implicated in AD and the availability of FDA-approved inhibitors, we examined whether MGAM inhibiting drugs (A10BF) such as acarbose or miglitol could reduce AD risk or slow cognitive decline in T2D subjects by comparing matched groups with and without drug exposure in the NACC dataset, excluding individuals with Parkinson’s Disease (PD) for accurate propensity score matching.

We matched the individuals in the A10BF drug user group (ever-users) with those in the non-use group (never-users) with 1:1 propensity score matching. An evaluation of the matching balance is reported in Table 3. For the majority of the covariates such as age, sex, race, education, comorbidities, and commonly prescribed medication use, SMD between the two groups was found to be < 0.2, suggesting that the two groups were well matched in these areas. Some of the higher SMDs, such as those for the usage of certain medications, could be attributed to very small sample sizes of the observations in either group. However, the A10BF drug-taking group had a significantly larger (P = 0.026) portion with family history of dementia (either parent reported to have cognitive impairment).

We first compared the dementia-free probability in the two groups by setting the duration of dementia-free as the date from the first visit to the date where the first recorded dementia diagnosis was reported (i.e., first occurrence of DEMENT = 1). In the Kaplan-Meier plot comparing the groups with 1:1 matching, a borderline significant trend (P = 0.059) can be seen, with the non-users having an earlier dementia diagnosis rate or a higher AD susceptibility than the drug user group (Figure 4A). This suggests that the drug could play a role in decreasing dementia risk in T2D patients.

Since dementia is a progressive neurodegenerative condition rather than a binary state, we further examined the drugs’ effects on cognitive performance by comparing longitudinal CDR measurements between the ever-user and never-user groups. In addition to overall CDR deterioration over time (P < 0.001 for follow-up duration), we observed a significant group difference in CDR change (interaction between follow-up duration and A10BF status) across all six domains (P < 0.05 for each domain and P < 0.001 for CDRSUM; Figure 4B and Supplementary Table 9), demonstrating a strong association between drug use and cognitive outcomes. The drug taking group showed a significantly slower cognitive decline, especially for orientation (P < 0.001), community affairs (P < 0.001), personal care (P < 0.01), and home and hobbies (P < 0.01). These observations still held when we increased the matching ratio to 2 and 3, respectively (Supplementary Figure 5 and Supplementary Table 9), strengthening the possibility that A10BF could help ameliorate cognitive decline in T2D subjects.