NG00186 – Genetic Component of the Mapping the Human Connectome During Typical Aging (HCP-A) & Vulnerability and Resiliency in the Aging Adult Brain Connectome (AABC)

Objectives: This project identifies ancestry- and sex-specific APOE ε4 modifier genes—variants that amplify or attenuate APOE ε4’s effect on AD risk and ARIA susceptibility from anti-amyloid immunotherapy. Aim 1: Trans-ancestry sex-stratified GWIS to construct an APOE-Wide Epistasis Map. Aim 2: Mechanistic validation via eQTL/pQTL colocalization and epistasis network. Aim 3: Explainable AI (XAI) integrating modifier SNPs, multi-omics subtypes, and ARIA proxy outcomes to stratify pre-treatment ARIA risk. Study Design: Multi-cohort secondary analysis using NIAGADS-controlled ADSP data exclusively. Individual-level data from all 15 ADC cohorts (NG00022–NG00151) and multi-ancestry ADSP WGS (NG00067, NG00166) span European, African American, Hispanic/Latino, and South/East Asian ancestries. Functional datasets (eQTL/pQTL: NG00102, NG00118, NG00120, NG00130) support Aim 2; imaging and neuropathology datasets (NG00103, NG00147, NG00175) enable Aim 3 ARIA proxy development. No prospective recruitment. Multi-dataset rationale: GWIS requires 4–8× more samples than standard GWAS (Gauderman 2002); no single cohort is independently powered—all 15 ADC cohorts must be pooled. Trans-ancestry GWIS requires ancestry-matched datasets (NG00100/African, NG00106/South Asian, NG00141/Hispanic) because population-specific LD cannot be imputed from summary statistics. Functional datasets (eQTL, pQTL, methylation) are non-redundant—each covers a distinct regulatory layer for Aim 2. All datasets are AD-specific; non-AD neurodegeneration data are excluded. Analysis Plan: Phenotypes: AD case/control (primary); APOE ε4 × SNP interaction; lobar microbleed count (ARIA-H proxy); SVD score (WMH, lacunar infarcts, perivascular spaces); longitudinal cognitive decline. Covariates: age, sex, top 20 ancestry PCs, stratum. Methods: logistic GWIS; trans-ancestry meta-analysis (METAL/MR-MEGA); sex-stratified/X-chromosome analyses; eQTL/pQTL colocalization (COLOC2/SMR); XGBoost XAI with 5-fold CV and SHAP.

Alzheimer’s disease affects tens of millions worldwide. Lecanemab, approved in 2024, slows Alzheimer’s progression by removing amyloid plaques—but causes dangerous brain side effects (ARIA: Amyloid-Related Imaging Abnormalities) especially in APOE ε4 carriers, who also most need treatment. Currently, doctors cannot predict which APOE ε4 carriers will benefit versus be harmed. Our research identifies modifier genes controlling how dangerous APOE ε4 is. We leverage the ADSP’s diverse dataset spanning 15+ cohorts across European, African American, Hispanic/Latino, and Asian ancestries—a scale statistically necessary because detecting gene–gene interactions requires 4–8× more samples than standard genetic studies. Population-specific patterns allow high-confidence modifier identification. MRI-based brain bleeds and vascular markers serve as validated ARIA surrogates available at scale. The result is an explainable AI tool that predicts—before treatment begins—which APOE ε4 patients face high ARIA risk and which will benefit from lecanemab, enabling precision Alzheimer’s therapy.

We have been developing more powerful statistical methods to detect common variant (CV)- or rare variant (RV)-complex trait associations and/or putative causal relationships for GWAS and DNA sequencing data. Here we propose applying our new methods, along with other suitable existing methods, to the existing ADSP sequencing data and other AD GWAS data provided by NIA, hence requesting approval for accessing the ADSP sequencing and other related GWAS/genetic data. We have the following two specific Aims: Aim1. Association testing under genetic heterogeneity: For complex traits, genetic heterogeneity, especially of RVs, is ubiquitous as well acknowledged in the literature, however there is barely any existing methodology to explicitly account for genetic heterogeneity in association analysis of RVs based on a single sample/cohort. We propose using secondary and other omic data, such as transcriptomic or metabolomic data, to stratify the given sample, then apply a weighted test to the resulting strata, explicitly accounting for genetic heterogeneity that causal RVs may be different (with varying effect sizes) across unknown and hidden subpopulations. Some preliminary analyses have confirmed power gains of the proposed approach over the standard analysis. Aim 2. Meta analysis of RV tests: Although it has been well appreciated that it is necessary to account for varying association effect sizes and directions in meta analysis of RVs for multi-ethnic cohorts, existing tests are not highly adaptive to varying association patterns across the cohorts and across the RVs, leading to power loss. We propose a highly adaptive test based on a family of SPU tests, which cover many existing meta-analysis tests as special cases. Our preliminary results demonstrated possibly substantial power gains. Aim 3. Inferring (putative) causal genes/proteins/metabolites for AD. We will apply TWA/PWAS/MWAS/xWAS methods to ADSP, AD GWAS and other omic data to identify (putative) causal genes, proteins, metabolites and other molecular traits for AD. These methods may be based on standard linear models or emerging ML/AI methods.

We propose applying our newly developed statistical and computational methods, along with other suitable existing methods, to the existing ADSP sequencing data, other AD GWAS data and other omic data to detect common or rare genetic variants and other molecular traits, such as genes/proteins/metabolites, associated with and/pr (putative) causal to Alzheimer’s disease (AD). The novelty and power of our new methods are in three aspects: first, we consider and account for possible genetic heterogeneity with several subcategories of AD; second, we apply powerful meta-analysis methods to combine the association analyses across multiple subcategories of AD; third, we will develop and apply standard linear model- and emerging ML/AI-based causal inference methods to infer causal genes, proteins, metabolites and other traits for AD. In addition, our proposed analyses of the existing large amount of ADSP sequencing data and other AD GWAS data with our developed new methods are novel, powerful and cost-effective.

We will use the data to build an AD-related omics database both by using state-of-the-art and by developing advanced deep learning-based methods for harmonization and imputation of multi-omics data, facilitating system biology studies to gain deep understanding of AD. We will use the data to identify the genetic biomarkers with causal effect on behavioral deficits in Alzheimer’s study and use these biomarkers to help predict Alzheimer's disease (AD). SNP data will be screened first using the GWAS and significant SNPs will be selected as genetic biomarkers according to their p-values. Then causal effects will be estimated to evaluate the contribution of genetic biomarkers. For AD prediction, enhanced statistical, machine learning, and deep learning approaches will be explored and compared, which may include but not limited to: the PCA decomposition, ridge regression/elastic net algorithm, boosting algorithms such as XGBoost/lightGBM, deep learning models such as the deep factorization machine. In far future, we want to develop a structured, literature and expert-knowledge based knowledge graph for better prediction, diagnosis and treatment of AD.

Leveraging our newly developed causal inference method, we aim to identify genetic causal pathways for the Alzheimer's disease (AD) from genomic data collected from multiple populations. The identified features will be used to predict cognitive and behavior scores among patients with AD or mild cognitive impairment (MCI). Finally, we want to develop a knowledge graph for better prediction, diagnosis and treatment of AD.