# Joint modeling of effect sizes for two correlated traits: Characterizing trait properties to enhance polygenic risk prediction

## TL;DR

A new method called PleioSDPR improves genetic risk prediction by combining information from related traits, enhancing accuracy for traits like bipolar disorder and hip circumference.

## Contribution

PleioSDPR is a novel method that jointly models effect sizes of two correlated traits to improve polygenic risk prediction.

## Key findings

- PleioSDPR improves prediction accuracy for bipolar disorder and hip circumference by 14.5% and 14.6%, respectively.
- Traits with stronger genetic correlation, higher heritability, and limited sample overlap contribute most to prediction gains.
- PleioSDPR outperforms existing univariate and multivariate PGS methods, especially when no validation dataset is available.

## Abstract

Recent years have witnessed a surge in the development of innovative polygenic score (PGS) methods, driving their extensive application in disease prevention, monitoring, and treatment. However, the accuracy of genetic risk prediction remains moderate for most traits. Currently, most PGSs were built based on the summary statistics from the target trait, while many traits exhibit varied degrees of shared genetic architecture or pleiotropy. Appropriate leveraging of pleiotropy from correlated traits can potentially improve the performance of PGS of the target trait. In this study, we present PleioSDPR, a novel method that jointly models the genetic effects of complex traits and identifies conditions under which leveraging pleiotropy can improve polygenic risk prediction. PleioSDPR models the joint distribution of effect sizes across traits, allowing SNPs to be null for both traits, causal for only one trait, or causal for both traits, and it flexibly captures region-specific genetic correlations and unequal heritability across traits. Through extensive simulations and real data applications, we demonstrate that PleioSDPR improves prediction performance compared with several univariate and multivariate PGS methods, especially when there is no validation dataset. For example, by incorporating information from schizophrenia or leg fat-free mass, PleioSDPR effectively improves the prediction accuracy of bipolar disorder (14.5% accuracy gain) and hip circumference (14.6% accuracy gain), respectively. Moreover, our results indicate that traits with stronger genetic correlations to the target trait, greater heritability, and limited sample overlap contribute more substantially to enhancing prediction accuracy for the target trait. Overall, our study highlights the potential of PleioSDPR to enhance the accuracy of genetic risk prediction by effectively leveraging pleiotropy across traits and diseases. These findings contribute to a broader understanding of polygenic risk prediction and underscore the importance of incorporating pleiotropic information to improve the use of these predictions in disease prevention and treatment strategies.

Most complex human traits and diseases are influenced by many genetic variants across the genome. Although polygenic scores (PGS) have emerged as a powerful tool for predicting an individual’s genetic risk, their accuracy remains limited for many traits. Much research has found that complex traits are genetically correlated, meaning they share a portion of their underlying biological basis. Existing PGS methods often use only the target trait’s data and therefore miss valuable information contained in related traits.

In this study, we introduce PleioSDPR, a statistical method that improves genetic risk prediction by jointly modeling two genetically correlated traits. PleioSDPR separates genetic variants into those that affect both traits, one trait, or neither, and allows their shared genetic effects to vary across genomic regions. It also accounts for sample overlap between studies, which can otherwise bias results. Through simulations and real data analysis, we show that PleioSDPR achieves more accurate or comparable predictions than existing methods, especially when the auxiliary trait has higher heritability, stronger genetic correlation with the target trait, and includes samples not shared with the target trait. Our work demonstrates that incorporating pleiotropic information can substantially enhance the accuracy and utility of polygenic scores.