The XPASS package implements the XPASS approach for constructing PRS in an under-representated target population by leveraging Biobank-scale GWAS data in European populations.
#install.packages("devtools")
devtools::install_github("YangLabHKUST/XPASS")
We illustrate the usage of XPASS using the GWAS summary statistics of BMI from UKB and BBJ. For demonstration, we use the easily accessible genotypes from The 1000 Genomes Project as the reference panel. However, because this reference panel only contains 1.3 million SNPs from 377 EAS amples and 417 EUR samples, the result of XPASS is slightly less accurate than what we reported in the AJHG paper. The performance of XPASS can be better when a GWAS dataset with a larger sample size (e.g.,n>2000) and more SNPs (e.g., 3M) is used as the reference panel. We strongly suggest that users use their own reference panels with sufficiently large sample sizes if available.
The datasets involved in the following example can be downloaded from here.
Input files of XPASS include:
- summay statistics file of the target population
- summay statistics file of the auxiliary population
- reference panel of the target population in plink 1 format
- reference panel of the auxiliary population in plink 1 format
- covariates file associated with the target population reference panel. The covariates files can include sex, age and population information of the individuals from the reference panel.
- covariates file associated with the auxiliary population reference panel
Different from LDSC that only utilizes LD from local SNPs, XPASS uses the LD information from the entire chromosomes to estimate heritability and coheritability. This approach yields smaller standard error, but requires the population structures in the reference pannel to be properly corrected by including the covariates (e.g., principal components). Otherwise, the estimated heritability and coheritability can be biased by the population structures (see here for an example). Therefore, we strongly suggest that users include covariates (e.g., principal components) when using XPASS although they are optional in the software.
The XPASS format GWAS summary statistics file has 5 fields:
- SNP: SNP rsid
- N: sample size
- Z: Z-scores
- A1: effect allele
- A2: other allele.
Here, we use the BMI GWAS from BBJ male as the target training set and BMI GWAS from UKB as the auxiliary training set
$ head BMI_bbj_male_3M_format_3MImp.txt
SNP N Z A1 A2
rs117086422 159095 -1.20413423957762 T C
rs28612348 159095 -1.25827042089233 T C
rs4475691 159095 -1.19842287777303 T C
rs950122 159095 -1.2014434974188 C G
rs3905286 159095 -1.27046106136441 T C
rs28407778 159095 -1.26746342605063 A G
rs4246505 159095 -1.24706211285128 A G
rs4626817 159095 -1.26366297625074 A G
rs11507767 159095 -1.28611566069053 G A$ head height_ukb_3M_format.txt
SNP N Z A1 A2
rs117086422 429312 1.42436004338939 T C
rs28612348 429312 1.48706291417224 T C
rs4475691 429312 1.53977372135067 T C
rs950122 429312 1.37958155329171 C G
rs3905286 429312 1.77045946243262 T C
rs28407778 429312 1.9908370573435 A G
rs4246505 429312 1.90922505355565 A G
rs4626817 429312 1.53216668392479 A G
rs11507767 429312 1.55873328059033 G AWe keep the BMI GWAS from BBJ female as the external validation dataset:
$ head BMI_bbj_male_3M_format_3MImp.txt
SNP N Z A1 A2
rs117086422 72390 0.897252679561414 T C
rs28612348 72390 0.904738461538462 T C
rs4475691 72390 0.89177374486687 T C
rs950122 72390 0.891523178807947 C G
rs3905286 72390 0.827441738675046 T C
rs28407778 72390 0.816801884323476 A G
rs4246505 72390 0.812148186935463 A G
rs4626817 72390 0.811624558188245 A G
rs11507767 72390 0.811100929441026 G AThe covariates files should not include row names and column names. The rows should be exactly corresponding to the individuals in the .fam file of reference genotypes. Each column corresponds to one covariate. A column of one should not be included in the file.
$ head 1000G.EAS.QC.hm3.ind.pc5.txt
-0.0242863 0.0206888 -0.0028171 0.0343263 0.0211044
-0.025051 0.0275325 -0.0332156 -0.0233166 0.0588989
-0.0198603 0.0286747 0.008464 -0.0215478 0.0189802
-0.0117008 0.0251493 -0.0228276 -0.0670019 0.0186454
-0.0246233 0.0281415 -0.0418795 0.0059947 0.0196607
-0.0270411 0.0127586 -0.0376545 0.0182809 0.0344744
-0.0193744 0.0367263 -0.0176168 -0.0307868 0.0254125
-0.0212299 0.022658 0.0225698 -0.0249273 0.0144324
-0.0159054 0.00558952 -0.00609582 -0.033497 0.0518336
-0.0258843 0.0476758 0.0073353 0.0164056 0.0072118$ head 1000G.EUR.QC.hm3.ind.pc20.txt
0.0290072 0.0717627 -0.0314029 0.0317316 0.0618357 0.0385132 0.122857 -0.0289581 -0.0114267 -0.0205926 -0.0466983 0.0836711 0.00690379 0.0345008 -0.0179313 0.0109661 -0.0214763 0.0014544 0.0182944 0.0399625
0.0378568 0.0499758 -0.00811504 0.0363021 -0.0579984 0.0422509 0.141279 -0.0167868 -0.0181999 0.0165593 -0.0304088 0.0423324 0.0226001 0.00843853 0.0212477 -0.0666462 -0.0787379 0.0136196 0.108933 0.0801246
0.0417318 0.0732938 -0.0409508 0.0176872 0.0801957 0.0124742 0.0252689 -0.0444921 -0.00305238 0.0035535 -0.0070929 0.0240194 -0.00543616 0.0272464 -0.0048309 -0.0207223 0.0415044 0.0494025 0.0213837 -0.0021093
0.0348071 0.0715395 -0.0266058 0.00280025 -0.0166164 0.0440144 0.135709 0.017364 0.0276564 -0.00286321 0.0314583 0.00299185 0.0792055 0.012042 -0.0337269 -0.00999033 0.0186435 -0.0699027 0.00791191 -0.0131168
0.0444763 0.0713348 0.00162451 -0.0107805 0.0868909 0.0218014 0.0314216 -0.0429928 -0.0137937 0.00913544 -0.062828 0.0555199 0.0378234 -0.0162297 0.000344947 -0.0164497 0.0523967 -0.0861731 -0.038893 0.028166
0.0300713 0.0522082 -0.0116997 -0.029994 0.0539977 0.0162067 -0.0522507 0.0022036 0.0255471 0.0129012 0.0371803 0.0701241 0.0314957 0.00870374 0.00566795 0.116672 0.0267188 0.0451948 0.00288655 -0.0427304
0.0339424 0.0718272 0.000614329 -0.0183555 0.0162768 -0.0600599 0.00343218 0.0149793 -0.0236301 -0.0267658 0.0387814 -0.00624387 -0.0364751 0.00486515 -0.0341221 0.0415286 -0.0274807 -0.013188 0.0695243 0.0495376
0.0423378 0.0656126 -0.0331807 -0.0361484 -0.0155739 0.0557459 0.00428138 0.0840953 -0.034451 0.0753096 -0.0180153 0.0595412 -0.0367107 -0.0285888 -0.0986386 0.00845412 -0.00388558 -0.0641134 -0.05815 0.0242433
0.0429698 0.0431213 -0.0357636 -0.00270477 -0.0567958 0.0892002 0.0980711 0.0468321 -0.0359592 0.011195 -0.000235849 -0.0192522 -0.00271491 0.0381155 -0.00845755 -0.0171629 -0.026532 -0.0415778 0.0274635 0.122063
0.0290721 0.0541744 -0.0238317 0.0254426 0.0986334 0.0706142 0.0977585 0.00427919 -0.0381976 0.0020029 -0.0161052 -0.016666 -0.00627125 -0.00490556 -0.0410802 0.0125096 -0.0175252 0.0320359 0.00866061 0.0736791Once the imput files are formatted, XPASS will automatically process the datasets, including SNPs overlapping and allele matching. Run XPASS with the following comand:
# library(devtools)
# install_github("https://github.com/YangLabHKUST/XPASS")
library(XPASS)
library(data.table)
library(RhpcBLASctl)
blas_set_num_threads(30)
# reference genotypes for EAS (prefix of plink file bim/bed/fam)
ref_EAS <- "1000G.EAS.QC.hm3.ind"
# covariates of EAS reference genotypes
cov_EAS <- "1000G.EAS.QC.hm3.ind.pc5.txt"
# reference genotypes for EUR (prefix of plink file bim/bed/fam)
ref_EUR <- "1000G.EUR.QC.hm3.ind"
# covariates of EUR reference genotypes
cov_EUR <- "1000G.EUR.QC.hm3.ind.pc20.txt"
# genotype file of test data (plink prefix).
# Note: for demonstration, we assume that the genotypes of prediction target are used as the reference panel of target population.
# In practice, one can also use genotypes from other sources as reference panel.
BMI_test <- "1000G.EAS.QC.hm3.ind"
# sumstats of height
BMI_bbj_male <- "BMI_bbj_male_3M_format_3MImp.txt" # target
BMI_ukb <- "BMI_ukb_sumstat_format_all.txt" # auxiliary
BMI_bbj_female <- "BMI_bbj_female_3M_format_3MImp.txt" # external validation
fit_bbj <-XPASS(file_z1 = BMI_bbj_male,file_z2 = BMI_ukb,file_ref1 = ref_EAS,
file_ref2 = ref_EUR,
file_cov1 = cov_EAS,file_cov2 = cov_EUR,
file_predGeno = BMI_test,
compPRS=T,
pop = "EAS",sd_method="LD_block",compPosMean = T,
file_out = "BMI_bbj_ukb_ref_TGP")
Summary statistics file 1: BMI_bbj_male_3M_format_3MImp.txt
Summary statistics file 2: BMI_ukb_sumstat_format_all.txt
Reference file 1: 1000G.EAS.QC.hm3.ind
Reference file 2: 1000G.EUR.QC.hm3.ind
Covariates file 1: 1000G.EAS.QC.hm3.ind.pc5.txt
Covariates file 2: 1000G.EUR.QC.hm3.ind.pc20.txt
Reading data from summary statisitcs...
3506148 and 3777871 SNPs found in summary statistics files 1 and 2.
Reading SNP info from reference panels...
1209411 and 1313833 SNPs found in reference panel 1 and 2.
746454 SNPs are matched in all files.
0 SNPs are removed because of ambiguity; 746454 SNPs remained.
Calculating kinship matrix from the both reference panels...
127749 SNPs in the second reference panel are alligned for alleles according to the first.
14337 SNPs have different minor alleles in population 1, z-scores are corrected according to reference panel.
14332 SNPs have different minor alleles in population 2, z-scores are corrected according to reference panel.
Assigning SNPs to LD Blocks...
Calculate PVE...
h1 h2 h12 rho
[1,] 0.165790395 0.247603545 0.129399058 0.63866483
[2,] 0.007631346 0.008542654 0.006856057 0.02002658
...
Predicting PRS from test genotypes...
Done.
XPASS returns a list of results, some key outputs are:
- H: a table of estimated heritabilities, co-heritability and genetic correlation (first row) and their corresponding standard erros (second row).
> fit_bbj$H
h1 h2 h12 rho
[1,] 0.165790395 0.247603545 0.129399058 0.63866483
[2,] 0.007631346 0.008542654 0.006856057 0.02002658
- mu: a data frame storing the posterior means computed by LDpred-inf using only the target dataset (mu1) and only the auxiliary dataset (mu2), and the posterior means of the target population and the auxiliary population computed by XPASS (mu_XPASS1 and mu_XPASS2). In eff_type1 and eff_type2 columns, "RE" indicates the SNP effect is random effect and "FE" indicates the effect is treated as population-specific fixed effect in the corresponding populations. SNPs information is also returned: A1 is the effect allele, A2 is the other allele.
> head(fit_bbj$mu)
CHR SNP POS A1 A2 mu1 mu2 mu_XPASS1
1 1 rs4475691 846808 T C -1.271443e-04 -0.0002770532 -0.0003248579
2 1 rs7537756 854250 G A -4.778206e-05 -0.0003401613 -0.0002979870
3 1 rs3748592 880238 A G -3.201406e-04 -0.0012640981 -0.0007477506
4 1 rs2340582 882803 A G -3.396992e-04 -0.0012885694 -0.0008076512
5 1 rs4246503 884815 A G -3.318260e-04 -0.0013238182 -0.0008148704
6 1 rs3748597 888659 T C -3.327131e-04 -0.0013031399 -0.0008113772
mu_XPASS2 eff_type1 eff_type2
1 -0.0003319743 RE RE
2 -0.0003509346 RE RE
3 -0.0014312273 RE RE
4 -0.0014550808 RE RE
5 -0.0014828292 RE RE
6 -0.0014623186 RE RE
- PRS (if file_predGeno provided and compPRS=T): a data frame storing the PRS generated using mu1, mu2, mu_XPASS1, and mu_XPASS2, respectively.
> head(fit_bbj$PRS)
FID IID PRS1 PRS2 PRS_XPASS1 PRS_XPASS2
1 HG00403 HG00403 0.14310344 1.1430451 0.55960910 1.1638272
2 HG00404 HG00404 -0.19478304 0.7149414 0.03773900 0.6465811
3 HG00406 HG00406 -0.09440555 0.8028816 0.14173517 0.7110257
4 HG00407 HG00407 0.05533467 0.1135734 0.05091083 0.1354040
5 HG00409 HG00409 0.25063980 1.9968242 0.83527097 1.9851891
6 HG00410 HG00410 0.13454737 0.2543978 0.10214320 0.2608015
# One can also compute PRS after fitting the model:
> PRS <- predict_XPASS(fit_bbj$mu,ref_EAS)
> head(PRS)
FID IID PRS1 PRS2 PRS_XPASS1 PRS_XPASS2
1 HG00403 HG00403 0.14310344 1.1430451 0.55960910 1.1638272
2 HG00404 HG00404 -0.19478304 0.7149414 0.03773900 0.6465811
3 HG00406 HG00406 -0.09440555 0.8028816 0.14173517 0.7110257
4 HG00407 HG00407 0.05533467 0.1135734 0.05091083 0.1354040
5 HG00409 HG00409 0.25063980 1.9968242 0.83527097 1.9851891
6 HG00410 HG00410 0.13454737 0.2543978 0.10214320 0.2608015
XPASS will also write above outputs into the files with file_out prefix, if provided.
We use the GWAS of female BMI from BBJ as the external validation dataset to approximate the prediction R2. Specifically we use the following equation:
where z is the z-score of external summsry statistics, n is its sample size, is the posterior mean of effect size at the standardized genotype scale,
is the LD reference panel.
The evalR2_XPASS function takes posterior means from XPASS output and the external validation summary statistics as its input to evaluate the approximated predictive R2. The ouput is a vector of prdictive R2 evaluated using mu1, mu2, mu_XPASS1, and mu_XPASS2, respectively.
> R2 <- evalR2_XPASS(fit_bbj$mu,BMI_bbj_female,ref_EAS)
> R2
PRS1 PRS2 PRS_XPASS1 PRS_XPASS2
0.02235596 0.01638359 0.02905665 0.01949609
While the reference panels have only limmited samples, XPASS still achieves 30% relative improvement compared to LDpred-inf in terms of R2.
XPASS+ allows the population-specific effects to be utilized in PRS construction. To fit XPASS+, we first need to apply the P+T procedure to construct a set of pre-selected SNPs that are population-specific. Here, we use the ieugwasr package.
library(ieugwasr)
# P+T procedure for bbj
z_bbj <- fread(BMI_bbj_male) # read summary statistics file
pval <- data.frame(rsid=z_bbj$SNP,pval=2*pnorm(abs(z_bbj$Z),lower.tail=F))
clp_bbj <- ld_clump(pval, clump_kb=1000, clump_r2=0.1, clump_p=1e-6,
bfile=ref_EAS,
plink_bin="/import/home/mcaiad/plink/plink")
snps_l <- clp_bbj$rsid
The rs id of pre-selected SNPs are stored in snps_l, which is then passed to the snps_fe1 argument in XPASS function.
fit_bbj <-XPASS(file_z1 = BMI_bbj_male,file_z2 = BMI_ukb,file_ref1 = ref_EAS,
file_ref2 = ref_EUR,
file_cov1 = cov_EAS,file_cov2 = cov_EUR,
file_predGeno = BMI_test,
snps_fe1 = snps_l,
compPRS=T,
pop = "EAS",sd_method="LD_block",compPosMean = T,
file_out = "BMI_bbj_ukb_plus_ref_TGP")
R2 <- evalR2_XPASS(fit_bbj$mu,BMI_bbj_female,ref_EAS)
> R2
PRS1 PRS2 PRS_XPASS1 PRS_XPASS2
0.02562267 0.01638359 0.03149972 0.01963232
By including the population-specific effects, XPASS+ achieves 3.15% predictive R^2, offering 8% improvement compared to XPASS.
While in the above example we only include population-specific effects in the target population, our implementation of XPASS+ allows the inclusion of such effects for both populations. To include fixed effects in both populations, the pre-selected SNPs for corresponding populations should be passed to snps_fe1 and snps_fe2 arguments, respectively.
# P+T procedure for ukb
z_ukb <- fread(BMI_ukb) # read summary statistics file
pval <- data.frame(rsid=z_ukb$SNP,pval=2*pnorm(abs(z_ukb$Z),lower.tail=F))
clp_ukb <- ld_clump(pval, clump_kb=1000, clump_r2=0.1, clump_p=1e-10,
bfile=ref_EUR,
plink_bin="/import/home/mcaiad/plink/plink")
snps_ukb <- clp_ukb$rsid
fit_both <-XPASS(file_z1 = BMI_bbj_male,file_z2 = BMI_ukb,file_ref1 = ref_EAS,
file_ref2 = ref_EUR,
file_cov1 = cov_EAS,file_cov2 = cov_EUR,
file_predGeno = BMI_test,
snps_fe1 = snps_bbj,
snps_fe2 = snps_ukb,
compPRS=T,
pop = "EAS",sd_method="LD_block",compPosMean = T,
file_out = "BMI_bbj_ukb_plus_ref_TGP")
R2 <- evalR2_XPASS(fit_both$mu,BMI_bbj_female,ref_EAS)
> R2
PRS1 PRS2 PRS_XPASS1 PRS_XPASS2
0.02353173 0.01192856 0.03294502 0.01419943
This yields similar results when the population-specific effects are included only in the target population.
An additional example for constructing PRS of Type 2 Diabetes can be found in this PDF.
Common issues are discussed in FAQ
The XPASS package is developed by Mingxuan Cai (mcaiad@ust.hk).
Please contact Mingxuan Cai (mcaiad@ust.hk) or Prof. Can Yang (macyang@ust.hk) if any enquiry.
Mingxuan Cai, Jiashun Xiao, Shunkang Zhang, Xiang Wan, Hongyu Zhao, Gang Chen, Can Yang. A unified framework for cross-population trait prediction by leveraging the genetic correlation of polygenic traits. The American Journal of Human Genetics. 108, 632-655, April 2021.