An alternative method of SNP inclusion to develop a generalized polygenic risk score analysis across Alzheimer's disease cohorts

Abstract

IntroductionPolygenic risk scores (PRSs) have great clinical potential for detecting late-onset diseases such as Alzheimer's disease (AD), allowing the identification of those most at risk years before the symptoms present. Although many studies use various and complicated machine learning algorithms to determine the best discriminatory values for PRSs, few studies look at the commonality of the Single Nucleotide Polymorphisms (SNPs) utilized in these models.MethodsThis investigation focussed on identifying SNPs that tag blocks of linkage disequilibrium across the genome, allowing for a generalized PRS model across cohorts and genotyping panels. PRS modeling was conducted on five AD development cohorts, with the best discriminatory models exploring for a commonality of linkage disequilibrium clumps. Clumps that contributed to the discrimination of cases from controls that occurred in multiple cohorts were used to create a generalized model of PRS, which was then tested in the five development cohorts and three further AD cohorts.ResultsThe model developed provided a discriminability accuracy average of over 70% in multiple AD cohorts and included variants of several well-known AD risk genes.DiscussionA key element of devising a polygenic risk score that can be used in the clinical setting is one that has consistency in the SNPs that are used to calculate the score; this study demonstrates that using a model based on commonality of association findings rather than meta-analyses may prove useful.

Evaluating the Classification Accuracy of Expression Quantitative Trait Loci Calculated Polygenic Risk Scores in Alzheimer’s Disease

“…In contrast, it is interesting to note that the effect sizes of the APOE isoform SNPs differ greatly between the summary statistics from the IGAP [12] and Jansen [13] datasets (rs429358: 1.35 vs. 0.162; rs7412: −0.387 vs. −0.885, respectively-Supplemental Tables S3 and S4). However, the accuracy of the classification is very similar, indicating that the allele frequency difference in the target dataset is the main influence on the classification accuracy, which has also been shown in a previous study [24].…”