Integrating common and rare variants into a genetic risk score for Alzheimer's disease risk predicti

Erica H Suh1, Garam Lee1, Sang-Hyuk Jung2, Zixuan Wen1, Jingxuan Bao1, Li Shen1,3, Dokyoon Kim1,3* 1Department of Biostatistics, Epidemiology & Informatics, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA 19104, USA 2Department of Digital Health, SAIHST, Sungkyunkwan University, Samsung Medical Center, Seoul, Republic of Korea 3Institute for Biomedical Informatics, University of Pennsylvania, Philadelphia, PA 19104, USA

Background: Polygenic risk scores (PRS) are commonly used to estimate an individual's risk for disease, recently including Alzheimer's disease (AD). However, PRS fails to demonstrate sufficient specificity and sensitivity in risk prediction and does not incorporate rare variants, which also influence AD risk. We generated a novel genetic risk score for Alzheimer's disease which combines a smaller set of more informative common and rare variants using known neuroimaging biomarkers to better predict those with high or low risk of disease, as well as mild cognitive impairment (MCI) conversion to AD. Method: Whole genome sequencing data of 1,704 patients were collected from the Alzheimer's Disease Neuroimaging Initiative (ADNI). Common variants of each patient were aggregated into a gene-based burden to enable smooth model estimation and gene-level interpretation. We then used SAIGE-GENE+ to conduct a gene burden rare variant analysis for each participant (minor allele frequency ≤ 0.01). The risk score was estimated based on common gene burdens exhibiting high correlations with neuroimaging biomarkers, FDG- and AV45-PET, and significantly associated rare loci from the corresponding rare variant analysis. PRS were generated with clumping and thresholding using PLINK. We compared the performances between our risk score and PRS on two tasks: cases vs controls classification and MCI conversion prediction. Result: Our risk score demonstrated better prediction performance and improved distinguishability between cases and controls over PRS. Our score also showed lower risk for controls and MCI non-converter patients and, conversely, higher risk for AD and MCI converter patients compared to PRS. Additionally, our score is well-calibrated compared to PRS-the number of predicted AD and MCI converter patients increased smoothly as the risk score increased. Conclusion: We proposed a novel approach to calculate an AD-specific genetic risk score that integrates both common and rare variants. The main advantage of this method is that it uses AD-relevant loci for calculation, yielding a score that is a better predictor of AD risk. Experiments demonstrated that this score enhances performance of prediction accuracy, calibration, and interpretability over PRS.