transPGS：Polygenic prediction for underrepresented populations through transfer learning by utilizing shared genetic similarity shared with European populations

Introduction

Current GWASs have been predominantly conducted in individuals of European (EUR) ancestry, with 94.6% in the EUR population and only 3.7% in the East Asian (EAS) population and 0.2% in the African (AFR) population. Due to this underrepresentation, the performance of PGS behaves poorly in non-EUR populations, particularly in populations of AFR ancestry. For example, the PGS accuracy reduced approximately 78% across multiple traits in individuals of AFR ancestry relative to those of EUR ancestry; similarly, the accuracy of PGS across traits was on average 40% lower in individuals of South Asian ancestry and 5% lower in individuals of EAS ancestry compared to that in those of EUR ancestry. The poor transferability of PGS derived from EUR ancestry data to non-EUR populations leads to great concern in health disparities . Therefore, there is an urgent need to develop novel PGS methods which can exploit data across diverse populations to better perform genetic risk prediction.

Increasing sample sizes in non-EUR GWASs for the understanding of genetic architecture underlying complex phenotypes is a necessary road for understudied populations such as EAS and AFR; but this requires plenty of expense and time. Alternatively, integrating existing knowledge available from EURs into non-EURs by novel approaches may be another promising strategy to improve the portability of PGS. Actually, there is a deal of evidence that significant genetic similarity exists between the EUR and non-EUR populations at both SNP and gene levels. Such genetic similarity provides theoretical and biological support for trans-ethnic leveraging of EUR information into non-EUR studies. Currently, there are a range of trans-ethnic statistical methods that help enhance the transferability of PGS across distinct ancestral groups; however, how to optimally integrate EUR information into non-EUR genetic research remains unknown.

In the machine learning filed, transfer learning is recognized as a novel technique that enables the utilization of knowledge acquired from auxiliary samples to enhance learning capability in target samples, which are distinct yet related to the former (Bastani, 2021; Li, et al., 2023; Lu, et al., 2023; Pan and Yang, 2009; Tian, et al., 2022; Zhao, et al., 2022). By borrowing the idea of transfer learning,we propose transPGS to boost the genetic prediction accuracy in non-EUR populations by leveraging trans-ethnic genetic similarity shared with the EUR population. Theoretically, transPGS helps capture genetic information across diverse ancestral populations and renders the prediction more efficiently and accurately.

Example

library(data.table)
library(SKAT)
library(glmnet)
library(MASS)
library(Rcpp)
library(RcppArmadillo)
library(doParallel)
sourceCpp("lmm_PXEM.cpp")
source("Individual_level_transPGS.R")
source("Summary_level_transPGS.R")

################ Individual-level transPGS #######################

###### data1 is the phenotypic data for the target population, including outcome and covariates.
###### data2 is the phenotypic data for the auxiliary population, including outcome and covariates.
###### G1 is the target population genotype data (matched to 1000 Genomes Project).
###### G2 is the auxiliary population genotype data (matched to 1000 Genomes Project).

data1 <- data.frame(fread("data1.txt"))
data2 <- data.frame(fread("data2.txt"))
G1 <- data.frame(fread("geno1.txt"))
G2 <- data.frame(fread("geno2.txt"))
a1 <- Individual_level_transPGS(data1, data2,G1,G2)
head(a1)
original_beta  tl_beta
1      2.106475 2.082117
2      1.961041 1.909088
3      1.160025 1.239434
4      2.407902 2.368119
5      1.800083 1.832781
6      2.318421 2.314352

#### original_beta is the effect before transfer learning
#### tl_beta is the effect after transfer learning  

################ Summary_level transPGS #######################

###### T is the GWAS summary statistics for the target and auxiliary populations, including marginal effects as well as standard errors.
###### G1 is the target population genotype data (matched to 1000 Genomes Project).
###### G2 is the auxiliary population genotype data (matched to 1000 Genomes Project).

T <- data.frame(fread("data.txt"))
G1 <- data.frame(fread("target_geno.txt"))
G2 <- data.frame(fread("auxiliary_geno.txt"))
a2 <- Summary_level_transPGS(T,G1,G2)
head(a2)
original_beta      tl_beta
1   -0.05005137  0.015091193
2    0.09789231  0.021784294
3   -0.08230863 -0.009494420
4   -0.05116747 -0.006389153
5    0.26905642  0.007526012
6    0.02256624 -0.008591966

#### original_beta is the joint effect before transfer learning
#### tl_beta is the joint effect after transfer learning      

Cite

Yiyang Zhu^$, Wenying Chen^$ , Kexuan Zhu and Ping Zeng^# (2024). Polygenic prediction for underrepresented populations through transfer learning by utilizing shared genetic similarity shared with European populations.

Contact

We are very grateful to any questions, comments, or bugs reports; and please contact Ping Zeng via zpstat@xzhmu.edu.cn.