Search

Introduction: Our understanding of the genetic predisposition for age‐at‐onset (AAO) of Alzheimer's disease (AD) is limited. Here, we sought to identify genes modifying AAO and examined whether any have sex‐specific effects. Methods: Genome‐wide association analysis were performed on imputed genetic data of 9219 AD cases and 10,345 controls from 20 cohorts of the Alzheimer's Disease Genetics Consortium. AAO was modeled from cases directly and as a survival outcome. Results: We identified 11 genome‐wide significant loci (P < 5 × 10−8), including six known AD‐risk genes and five novel loci, UMAD1, LUZP2, ARFGEF2, DSCAM, and 4q25, affecting AAO of AD. Additionally, 39 suggestive loci showed strong association. Twelve loci showed sex‐specific effects on AAO including CD300LG and MLX/TUBG2 for females and MIR4445 for males. Discussion: Genes that influence AAO of AD are excellent therapeutic targets for delaying onset of AD. Several loci identified include genes with promising functional implications for AD. [ABSTRACT FROM AUTHOR]

Background: Several studies have attempted to identify genes for age‐at‐onset (AAO) of Alzheimer's disease (AD) through genome‐wide linkage and genome‐wide association studies (GWAS). While these efforts have thus far focused predominantly on non‐Hispanic white (NHW) populations, we present here the first GWAS to identify genetic contributors to AAO in African Americans (AA) using AA genotype and phenotype data from the Alzheimer's Disease Genetics Consortium (ADGC). Method: We combined data from 13 AA prospective and case‐control ADGC datasets, excluding family‐based datasets. Genotype data underwent standard quality control and were imputed to the TOPMed R5 reference panel using the Michigan Imputation Server. Data were merged across cohorts using QCTOOL and GTOOL, filtering out variants with MAF≤0.005). We performed two analyses: case‐only linear mixed‐model regression with AAO as outcome (LMM); and survival analysis using Cox proportional‐hazards frailty modeling (COXPHF) with time‐to‐AD‐onset (AAO in cases and censoring of controls at age‐at‐last‐exam). Both models were adjusted for sex, principal components capturing population substructure, and APOE‐ε4 dosage with a random cohort effect. Result: We examined 2,262 AA AD cases, including 1,561 females (69%) with mean (SD) AAO of 77.7 (8.1) years. The APOE‐adjusted COXPHF analyses identified twelve genome‐wide significant regions (P<5×10−8), the most significant in ABCA7 (P = 3.57×10−10), known to be associated with AD risk in AAs. In the LMM analysis, we observed strong novel associations with variants in CCDC141 on chr2q31.2 (Pmin = 2.93×10−9). In silico analyses of LD structure and functional potential suggest SESTD1, a gene involved in synaptic regulation, as driving the CCDC141 locus signal. Nine of 12 COXPHF loci were driven by African‐specific variants (MAF<0.0002 in gnomAD Non‐Finish Europeans). These significant loci identified in the AA COXPHF or LMM analyses were not significant (P>0.05) in equivalent analyses in NHW ADGC cohorts (9,220 AD cases/10,350 controls; Li et al. Alz Dement 2023), except for the ABCA7 locus (NHW P<0.0002). Conclusion: In a large ADGC African American dataset, we confirmed AAO associations in the APOE region and ABCA7 and observed evidence for potential novel AD AAO loci including SESTD1; many of which might be African‐ancestry specific. [ABSTRACT FROM AUTHOR]

Background: The prevalence of depression in people with Alzheimer's disease is estimated to be nearly 50%, accounting for a significant amount of additional caregiver burden1. With the COVID‐19 crisis, depression in dementia patients, especially those living in long term care facilities, has increased. The Cornell Scale for Depression in Dementia (CSDD) is the gold standard for assessing depression in dementia patients, but is lengthy, taking nearly 20 minutes to administer3. The SLU AMSAD is a previously validated depression tool which consists of 5‐items and takes less than 2 minutes to administer3. The objective of this study was to compare this short tool with the lengthier gold standard, the CSDD, to diagnose depression in cognitively impaired patients. Method: 44 patients, 65years and older were enrolled in the study from an out‐patient clinic setting. A St. Louis University Mental Status Exam (SLUMS) score of less than 25 was used to include the eligible patients. All enrolled subjects were administered the SLUMS, SLU AMSAD and the CSDD. The SLUMS score was used to classify the patients as having mild, moderate and severe dementia Result: The preliminary data set analyzed data for N=44 cases of which 54.5% were female and 88.6% Caucasian with a mean age of 79.9. The average SLUMS score was 14.3 (6.5). For the total sample, area under the curve (AUC), sensitivity, and specificity values exceeded 0.82, with a maximum sensitivity of 0.92 and specificity of 0.84 (for an optimal cut‐off score of 4/5). Conclusion: The SLU AMSAD had a very strong correlation with the CSDD as shown by the above results in this preliminary data sample of 44 patients. We intend to continue with the study and gather a total sample of 105 patients to add stronger statistical power to the study. References: 1. Chi S, et.al. The prevalence of depression in Alzheimer's disease: a systematic review and meta‐analysis. Current Alzheimer research. 2015;12(2):189‐ 98. 2. Alexopoulos, G. S., et.al (1988). Cornell scale for depression in dementia. Biological Psychiatry, 23(3), 271–284. 3. Chakkamparambil B, et.al Development of a brief validated geriatric depression screening tool: the SLU "AM SAD". Am J Geriatr Psychiatry. 2015 Aug;23(8):780‐3. [ABSTRACT FROM AUTHOR]

Background: Alzheimer's disease (AD) is the most common late‐onset neurodegenerative disease. Identifying individuals at increased risk of developing AD is important for early intervention. Risk prediction models are typically based on a limited number of predictors, possibly with sub‐optimal performance. Here, we explore an explainable machine learning (ML) framework, XGBoost and SHapley Additive exPlanations (SHAP) values, for AD risk prediction, which can handle a large number of predictors and output the impact and importance of each predictor. Method: We developed an XGBoost model that aggregates polygenic risk scores (PRSs), which include both PRS for AD risk and PRS for age at onset of AD, baseline individual characteristics (e.g., non‐genetic factors), and information from electronic health records for predicting incident AD. The PRSs were derived using summary statistics from genome‐wide association studies in the Alzheimer's Disease Genetics Consortium (ADGC) dataset (n = 19,918). The model was applied to 457,936 white participants in UK Biobank to predict development of AD within 10 years after the baseline visit (n = 2,177 developed AD). We further used SHAP values to explain the relative information in model predictors. Result: For participants of age 40 and older, the area under the receiver operating characteristic curve (AUC) for AD risk prediction was over 0.880. PRSs ranked second to age (the best predictor) in feature importance. For subjects of age 65 and above, PRSs for AD were the most important features. Our ML model not only identified traditional risk factors for AD, such as age, education, income, body mass index, diabetes, and blood pressure, but also identified predictors from electronic health records that are not typically considered in traditional prediction models, including urinary tract infection, syncope and collapse, chest pain, disorientation and hypercholesterolaemia, for developing AD. Furthermore, SHAP values aided the ranking of feature importance and model explanation. Conclusion: Our ML model improves the accuracy of AD risk prediction by efficiently exploring numerous predictors. PRSs play the most important role in developing AD in individuals of age 65 and older. In application, the model also identified novel feature patterns for AD. [ABSTRACT FROM AUTHOR]

Background: Alzheimer's disease (AD) has high genetic heritability for both disease risk and age‐at‐onset (AAO) of AD. However, our understanding of genetics of AAO of AD lags behind AD risk. Here, we utilized two statistical approaches to identify genes modifying AAO of AD globally or in a sex‐specific manner. Method: Pooled data from 9,219 AD cases and 10,345 controls from 20 cohorts from the Alzheimer Disease Genetic Consortium were analyzed using a linear mixed model (LMM) for AAO from cases only, and a Cox proportional hazard frailty model (CoxFM) from all subjects with AAO observed in cases and censored at age‐at‐exam in controls. Both methods modeled outcome on SNP, sex, APOE‐e4, and 10 principal components (PCs), and incorporated random intercept by cohort. Fisher's method was used to combine results of LMM and CoxFM. Genome‐wide significant variants were determined based on p < 5×10−8 in individual or combined tests. Multiple secondary analyses were performed including sex‐specific association tests for top SNPs (p< 10−6). Two gene‐expression datasets of prefrontal cortex tissues from NCBI/GEO were analyzed by sex to evaluate the sex‐specific AAO genes. Result: We identified 34 genome‐wide significant loci. We confirmed 6 known AD risk genes (APOE, CR1, BIN1, TREM2, PICALM, and FERMT2) also regulating AAO as AD. Twenty‐eight potential novel loci were identified including ARFGEF2, MAPK9, and the MLX region harboring functional related genes for AD. Sex‐specific analyses showed the effects of PICALM, MLX, and CDH2 on AAO were solely from females, and ATP2C1 from males. Particularly, the MLX region harboring 10 genes in strong LD, including variants in COASY and HSD17B1 previously reported to associate with earlier AAO of AD in female AD patients with Down Syndrome. Gene expression further supported higher differential expression of AD over controls for HSD17B1, COASY, PLEKHH3, and MLX in females than males. Conclusion: Using two statistical methods to model AAO of AD, we identified six known AD risk genes and 28 novel loci associated with AAO of AD, and further narrowed down four loci with sex‐specific effects. Notably, HSD17B1, involved in estrogen activation, has strong functional implications warrant for further investigation. [ABSTRACT FROM AUTHOR]

Background: Several studies have attempted to identify genes for age‐at‐onset (AAO) of dementia symptoms in Alzheimer's disease (AD) through genome‐wide linkage (Daw et al. 1999, Li et al. 2002, Choi et al. 2011) and genome‐wide association studies (Kamboh et al. 2012, Naj et al. 2014). Naj et al. conducted a meta‐analysis of 14 datasets from the Alzheimer's Disease Genetics Consortium (ADGC) with genotype imputation to HapMap2. Using an expanded set of ADGC cohorts, we re‐examined AAO using higher‐quality genome‐wide imputation data and pooled‐data analysis instead of meta‐analysis to identify novel genetic contributors to AAO of AD. Methods: We combined data from 20 non‐Hispanic white prospective and case‐control ADGC datasets, excluding family‐based datasets. Genotype data were imputed to the Haplotype Reference Consortium (HRC) r1.1 reference panel using the Michigan Imputation Server and consistent quality control applied to each dataset, after which genotype data were merged across datasets using QCTOOL and GTOOL. We performed case‐only association analyses for log10(AAO) using multivariable linear regression with covariate adjustment for sex, cohort, and principal components capturing population substructure, and then conducted analyses conditioning on dosage of APOE ε4 alleles. Variants with P < 10−5 from conditional analyses were re‐analyzed using linear mixed modeling treating cohort as a random effect. Results: We used 9,228 AD cases, including 5,307 females (57.4%) with mean (SD) AAO of 74.2(7.7) years. In addition to the APOE region, which reached genome‐wide significance (P < 5 × 10−8), we observed strong novel associations with variants in NTM (encoding neurotrimin) on chr11q25 (rs117589002, P = 1.24 × 10−7) and in an intergenic region on chr3p22.3 (Pmin = 1.89 × 10−7). Other AAO loci previously identified in the smaller ADGC dataset (Naj et al 2014) remained nominally significant. After adjusting for APOE ε4, thirty‐two regions showed suggestive association (P < 10−5), with chr3p22.3 and NTM remaining the top loci. Chromosome 11q25 includes a known AD linkage region (Blacker et al. 2003). In addition, NTM variation has been found to be correlated with cognitive function tests (Liu et al. 2007) and intelligence (Pan et al. 2010). Conclusions: In an expanded ADGC dataset, we confirmed AAO associations in the APOE region and observed evidence for potential novel AD AAO loci including NTM. [ABSTRACT FROM AUTHOR]

Background: Alzheimer's disease (AD) is the most common late‐onset neurodegenerative disease. About 5.4 million Americans are living with AD. Unfortunately, there is no cure for AD at present, which makes early prediction crucial. Identifying individuals at increased risk of AD provides a better chance of benefiting from treatments. Risk prediction models are typically based on a limited number of predictors possibly with sub‐optimal performance. Here, we explore a state‐of‐the‐art automated machine learning (AutoML) framework for AD risk prediction, which can handle hundreds of predictors, including non‐traditional variables, with automatic feature engineering and model selection. Method: We developed an AutoML model that aggregates polygenic risk scores (PRSs) and baseline individual characteristics (e.g., non‐genetic factors) for predicting AD. The PRSs were derived using summary statistics of the genome‐wide association studies from the Alzheimer Disease Genetics Consortium (ADGC) dataset (n = 19,918). The model was applied to 455,233 participants in UKBB without AD at baseline to predict development of AD at the final observation (n=1,452 developed AD). Our model was based on the H2O AutoML, an intelligent algorithm that can automatically select hyperparameters, tune ensembles of ML models, and carry out model assessment. Result: The area under the receiver operating characteristic curve (AUC) for AD risk prediction was over 0.86. Polygenic risk scores ranked only second to age in feature importance. Furthermore, our AutoML model identified predictors that are not typically considered in traditional prediction models, such as an individual's overall health rating and usual walking pace. Conclusion: Our AutoML model improves the accuracy of AD risk prediction by efficiently exploring numerous predictors and ensemble models while greatly reducing manual coding hours. Furthermore, AutoML uncovered novel predictors for AD. [ABSTRACT FROM AUTHOR]