2125-LB: Leveraging Rare Variant Association Analyses to Develop Polygenic Risk Scores (PRSs) for Type 2 Diabetes (T2D)

Introduction and Objective: T2D has a strong genetic basis, making PRSs useful for risk prediction. Most PRSs exclude rare variants (minor allele frequency [MAF] <0.01) due to data limitations in genome-wide association studies (GWAS). We used the current largest rare variant GWAS meta-analysis to develop a rare variant PRS (rvPRS) to improve T2D risk prediction. Methods: We developed rvPRSs using multi-ancestry GWAS data from 51,256 cases and 370,487 controls with rare variant imputation or whole-genome sequencing (MAF ≥ 5x10-5) and applied them to 7,382 cases and 8,885 controls of Latin American ancestry. We generated 84 PRSs with varying linkage disequilibrium clumping parameters (R2 and window size) and p-value thresholds. Ensemble learning weighted each PRS by predictive performance and applied these weights to variant effects. Model performance was assessed using adjusted Area Under the Receiver Operating Characteristic Curves (AAUC), adjusting for principal components, sex, age, and body mass index (BMI). We compared the rvPRS to a common variant PRS (cvPRS) derived from a European-based GWAS of 201,262 cases and 1,285,672 controls. Results: Our rvPRS (MAF ≥ 5x10-5) captured a higher proportion of rare variants (82% MAF <0.01 in Latin American ancestry vs. 38% in the cvPRS). Combining rvPRS and cvPRS (rv_cvPRS) improved prediction accuracy (AAUC = 0.68 vs. 0.66 for cvPRS alone), particularly at extreme PRS scores. Individuals at the 95th percentile of the rv_cvPRS had an odds ratio (OR) of 5.7 (4.3-7.4, p = 3.7x10-36) compared to 3.7 (2.9-4.7, p = 1.2x10-26) for the cvPRS. The rvPRS enhanced prediction for known rare variant carriers. Carriers of the HNF4A variant (p.Arg114Trp, OR = 7.9) had an average rvPRS corresponding OR of 3.1 compared to an average cvPRS corresponding OR of 0.82. Conclusion: Incorporating rare variants to PRSs substantially improved T2D prediction, particularly identifying individuals at high genetic risk driven by large-effect rare variants. Disclosure K. Taylor: None. A. Huerta: None. X. Wang: None. M. Vora: None. J. Kim: None. M. Ng: None. H. Zhang: None. J. Mercader: None. Funding American Diabetes Association (11-22-ICTSPM-16); National Human Genome Research Institute (U01HG011723); National Institute of Diabetes and Digestive and Kidney Diseases (R01DK137993 and U01 DK140757); Medical University of Bialystok (MUB) grant from the Ministry of Science and Higher Education (Poland)