An Ensemble Penalized Regression Method for Multi-ancestry Polygenic Risk Prediction

Abstract

Great efforts are being made to develop advanced polygenic risk scores (PRS) to improve the prediction of complex traits and diseases. However, most existing PRS are primarily trained on European ancestry populations, limiting their transferability to non-European populations. In this article, we propose a novel method for generating multi-ancestry Polygenic Risk scOres based on enSemble of PEnalized Regression models (PROSPER). PROSPER integrates genome-wide association studies (GWAS) summary statistics from diverse populations to develop ancestry-specific PRS with improved predictive power for minority populations. The method uses a combination of L1(lasso) and L2(ridge) penalty functions, a parsimonious specification of the penalty parameters across populations, and an ensemble step to combine PRS generated across different penalty parameters. We evaluate the performance of PROSPER and other existing methods on large-scale simulated and real datasets, including those from 23andMe Inc., the Global Lipids Genetics Consortium, and All of Us. Results show that PROSPER can substantially improve multi-ancestry polygenic prediction compared to alternative methods across a wide variety of genetic architectures. In real data analyses, for example, PROSPER increased out-of-sample prediction R2for continuous traits by an average of 70% compared to a state-of-the-art Bayesian method (PRS-CSx) in the African ancestry population. Further, PROSPER is computationally highly scalable for the analysis of large SNP contents and many diverse populations.

Ensembled best subset selection using summary statistics for polygenic risk prediction

“…While a large portion of GWAS data currently comes from individuals of European ancestry, it is well-known that polygenic risk scores do not transfer well between individuals of different ancestries, which can impact their utility for patients of non-European ancestry [64][65][66][67][68][69] . Many methods that utilize summary data from multiple populations have already been proposed and demonstrate improved prediction in under-represented populations 25,29,30,[70][71][72][73][74] . As genetics research moves towards greater diversity, ALL-Sum serves as a valuable foundation for extension to incorporating data from multiple ancestries, as well as admixed individuals [75][76][77] .…”

“…In high-dimensional problems such as PRS which involve a large number of SNPs, the best tuning parameters may be in the gaps or outside the bounds of the fixed grid being considered 23 . Recent works, in both theory and PRS applications, have demonstrated that ensembling multiple predictors can yield better prediction accuracy than choosing a single best predictor from grid search [24][25][26][27][28][29][30] .…”

JointPRS: A Data-Adaptive Framework for Multi-Population Genetic Risk Prediction Incorporating Genetic Correlation

“…Similarly, SDPRX linear 2 and XPASS linear 2 linearly combined the standardized PRS of two populations to obtain the final PRS. Conversely, MUSSEL super max and PROSPER super max consider a super learning step that substitutes the linear combination by the non-linear machine learning techniques, and use the ensemble learning instead of selecting the optimally-tuned parameters, following their original work [18,19].…”

“…To address this need, there has been an increase in the number of GWAS focused on non-European populations [8][9][10][11][12][13][14][15], complemented by the developments of various models tailored for multi-population PRS predictions. These models employ different strategies to leverage multiple GWAS results, including utilizing multiple populations, assuming sparse distributions for genetic risk variants across populations, and accounting for genetic correlations among populations [16][17][18][19][20][21][22]. However, to our knowledge, there is no method that integrates all these components under a coherent framework in the absence of individual-level validation data -a common situation in real data applications.…”

“…We illustrate the benefits of integrating genetic correlation structures and jointly using multiple populations over exclusive reliance on the target and the auxiliary European groups. In addition, we compare JointPRS with other methods, including PRS-CSx [17], MUSSEL [18], PROSPER [19], SDPRX [20], XPASS [21], and BridgePRS [22], with and without a validation dataset. Our results illustrate the distinct contributions of joint modelling and genetic correlations to prediction accuracy, and suggest JointPRS as a promising method for predicting complex traits across populations with the auto-tuned JointPRS achieving comparable performances as top-performing alternative approaches that require heavy tuning on individual-level validation sets.…”