Abstract: FR-PO670
Improving Polygenic Risk Scores for Steroid-Sensitive Nephrotic Syndrome by Integrating Relevant Traits
Session Information
- Pediatric Nephrology - II
November 03, 2023 | Location: Exhibit Hall, Pennsylvania Convention Center
Abstract Time: 10:00 AM - 12:00 PM
Category: Pediatric Nephrology
- 1900 Pediatric Nephrology
Authors
- Nagano, China, Boston Children’s Hospital, Boston, United States
- Mcnulty, Michelle, Boston Children’s Hospital, Boston, Massachusetts, United States
- Sampson, Matt G., Boston Children’s Hospital, Boston, Massachusetts, United States
- Reynolds, Kaylia M., Boston Children’s Hospital, Boston, Massachusetts, United States
Background
Polygenic risk scores (PRS) have been widely applied in research studies to identify associations of a person’s overall genetic profile with diseases/traits. We are in the early stages of developing PRS for steroid sensitive nephrotic syndrome (SSNS). PRSmix and PRSmix+ have been created to incorporate additional PRS for both the same disease and correlated traits to improve PRS prediction accuracy. Here, we tested the hypothesis that the prediction accuracy of an SSNS PRS would be increased using these strategies.
Methods
Using our previously created SSNS PRS (Barry et al.), we applied PRSmix and PRSmix+ in a European SSNS cohort. Candidate PRSs were obtained from the Polygenic Score Catalog. Individuals in our European SSNS cohort were randomly assigned to a training/testing cohort. Three models were created to assess the additive benefit of combining multiple PRS and correlated traits: the “best PRS” model only includes our SSNS PRS as a predictor, the “PRSmix” model incorporates all SSNS PRS, and the “PRSmix+” model adds high-power PRS from correlated traits to the PRSmix model. The Likelihood-Ratio test (LRT) was used to compare the goodness of when adding a PRS to the model. The Matthews Correlation Coefficient (MCC) was used to evaluate binary classifications; it ranges from −1 to 1, with extreme values 1 reached in case of perfect classification. Comparative predictive accuracy in the testing dataset was assessed using a two-sample z test.
Results
In a 100-fold cross validation study, adding another SSNS PRS (PRSmix) always significantly improved model fit (LRT<0.05) compared to the best PRS. Furthermore, adding high-powered traits (PRSmix+) always significantly improved model fit (LRT<0.05) compared to the PRSmix. The average case prediction accuracy (as measured by MCC) was slightly improved for both the PRSmix (MCC=0.499) over bestPRS (MCC=0.477; p=0.38) and the PRSmix+ (MCC=0.502) over best PRSmix (p=0.36). For PRSmix+, the PRS traits with highest inclusion across 100 models were eosinophil (100%), asthma (79%), and white blood cell counts (72%).
Conclusion
We improved the prediction capability of SSNS PRS using PRSmix and PRS mix+. PRS of other diseases/traits can increase the accuracy of predictive models, may ultimately improve clinical inference, and could provide a better understanding of SSNS disease etiology.