Alzheimer’s disease heterogeneity explained by polygenic risk scores based on brain transcriptomic profiles

Abstract

AbstractBackgroundConsidering the complex etiology of Alzheimer’s disease (AD), classifying AD patients into clinically‐relevant subtypes is important for precision medicine. We proposed a novel way to decompose AD heterogeneity using AD polygenic risk scores (PRSs) from biologically coherent gene‐sets, which were developed based on coexpressed genes in AD brains.MethodWe conducted network analysis using RNA‐sequencing data from 64 autopsied AD brains in the Framingham Heart Study and Boston University Alzheimer’s Disease Research Center and validated preservation of modules in independent AD brain RNA‐Seq datasets from Religious Orders Study and Memory and Aging Project (n=363) and Mayo Clinic Study of Aging (n=82). Next we identified AD‐associated modules by enrichment analyses using a false discovery rate (FDR)<0.05 with gene‐sets informed by GWAS for neuritic plaque and neurofibrillary tangles. Based on the genes ascribed to each module, we computed and evaluated associations of module‐based PRS (mbPRS) for NP and NFT mbPRSs with i) domain‐specific neuropsychological performances, ii) cognitively‐defined subgroups including memory (n=196), executive functioning (n=16), language (n=52), visuospatial functioning (n=91), multiple domain (n=28), and no domain (n=289) (Mukherjee, 2018), and iii) cortical thickness across whole brains in the 672 AD patients of Alzheimer’s Disease Neuroimaging Initiative (ADNI).ResultAmong the 143 co‐expressed gene‐sets (modules) identified in the discovery, 14 modules were preserved in both validation datasets with significant enrichment FDR<0.05 for AD associated genes. PRSs from three modules were significantly associated with performance on a particular cognitive test (logical memory test, Boston Naming test, and Clock copy score) each representing a distinct cognitive domain (memory, language, and visuospatial functioning, respectively). Importantly, mbPRS was significantly associated cortical atrophy in module‐specific regions. Further analysis showed that PRS for each of these three modules significantly differentiated subgroups of AD patients by memory (P=0.02), language (P=0.01) and visuospatial (P=0.04) dysfunction, respectively.ConclusionOur findings indicate that an approach integrating genetics and transcriptomic data can differentiate subtle variations in neuropsychological performance among AD patients and clinically‐defined AD subgroups (especially, memory, language, and visuospatial domains), which can be utilized for genetically‐informed AD patient classification.