Power of Inclusion: enhancing polygenic prediction with admixed individuals

Tanigawa and Kellis. Am J Hum Genet. (2023).

Phenotype: Eosinophil count

Link to the source data field in UK Biobank
Estimated h² in white British population in UKB: 0.139 (95% CI:[0.122, 0.156]).
Link to PGS catalog (Study ID: PGP000502).
GWAS Catalog for mapped trait(s): eosinophil count (EFO_0004842)

Our FAQ page shows the description of the file format and how you may use iPGS coefficients in your research.

iPGS prediction in the held-out test set individuals

We compared the polygenic prediction from our iPGS model and the phenotype values using the held-out test set individuals in UK Biobank. Note the difference in the number of individuals in the five population groups.

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.WB.PGS_vs_phe.png

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.NBW.PGS_vs_phe.png

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.SA.PGS_vs_phe.png

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.Afr.PGS_vs_phe.png

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.others.PGS_vs_phe.png

Predictive performance

Population	Model	Metric	Predictive Performance	95% CI	P-value
Population	Model	Metric	Predictive Performance	95% CI	P-value

white British	Covariate-only model	R²	0.009	[0.007, 0.010]	9.8x10^-129
white British	Genotype-only model	R²	0.093	[0.089, 0.097]	<1.0x10^-300
white British	Full model (covariates and genotypes)	R²	0.101	[0.097, 0.106]	<1.0x10^-300
Non-British white	Covariate-only model	R²	0.013	[0.005, 0.021]	2.4x10^-09
Non-British white	Genotype-only model	R²	0.072	[0.054, 0.090]	1.8x10^-47
Non-British white	Full model (covariates and genotypes)	R²	0.084	[0.065, 0.104]	1.4x10^-55
South Asian	Covariate-only model	R²	0.005	[-0.002, 0.013]	5.8x10^-03
South Asian	Genotype-only model	R²	0.048	[0.027, 0.069]	5.6x10^-17
South Asian	Full model (covariates and genotypes)	R²	0.053	[0.031, 0.075]	1.2x10^-18
African	Covariate-only model	R²	0.010	[-0.001, 0.021]	7.1x10^-04
African	Genotype-only model	R²	0.007	[-0.002, 0.016]	6.1x10^-03
African	Full model (covariates and genotypes)	R²	0.013	[0.000, 0.026]	9.9x10^-05
Others	Covariate-only model	R²	0.028	[0.021, 0.035]	3.5x10^-50
Others	Genotype-only model	R²	0.069	[0.058, 0.080]	3.7x10^-122
Others	Full model (covariates and genotypes)	R²	0.093	[0.081, 0.105]	2.0x10^-165

The predictive performance (R²), its 95% confidence interval (CI), and statistical significance (P-value) are shown for each population in UK Biobank in the held-out test set. The "model" column indicates whether the predictive performance is from the covariate-terms alone (covariate-only model), PGS terms alone (Genotype-only model), or the full model containing both PGS and covariate terms. We used the following sets of covariates in our analysis: age, sex, age², age*sex, Townsend deprivation index, and genotype PCs (PC1-PC18). Please refer to our publication for a more detailed description of the methods.

Coefficients (BETA) of PGS models

/static/data/tanigawakellis2023/per_trait/INI30150/INI30150.BETAs.png

We show the coefficients (BETA) of PGS models. Our iPGS model selected 16859 variants with non-zero coefficients. The genetic variants with the large absolute values of coefficients are annotated in the plot. There is no guarantee that our iPGS model selects causal variants. We use the GRCh37/hg19 reference genome.

The top 100 genetic variants with the largest absolute value of coefficients

CHROM	POS	Variant	Variant ID	Effect allele	Consequence	Gene symbol	Beta
CHROM	POS	Variant	Variant ID	Effect allele	Consequence	Gene symbol	Beta

17	4535314	17:4535314:G:A	rs34210653	A	PAVs	ALOX15	-0.016
9	6255967	9:6255967:G:C	rs146597587	C	PTVs	IL33	-0.011
3	128316435	3:128316435:A:G	rs4328821	G	Others		-0.011
12	111884608	12:111884608:T:C	rs3184504	C	PAVs	SH2B3	-0.010
9	6197392	9:6197392:T:C	rs1888909	C	Others	RP11-575C20.1	-0.008
21	36387806	21:36387806:C:T	rs2242886	T	Intronic	RUNX1	-0.008
10	64431206	10:64431206:T:C	rs12413946	C	UTR	ZNF365	0.006
7	92408370	7:92408370:C:T	rs445	T	Intronic	CDK6	-0.005
18	60920854	18:60920854:C:T	rs17758695	T	Intronic	BCL2	-0.005
7	75442723	7:75442723:G:A	rs11465293	A	PAVs	CCL24	-0.005
2	213824045	2:213824045:A:G	rs12619285	G	Intronic	AC079610.1	-0.005
6	31106501	6:31106501:C:CC	Affx-89026413	CC	PTVs	PSORS1C1	-0.004
2	242690675	2:242690675:G:A	rs1106639	A	PAVs	D2HGDH	-0.004
10	64397538	10:64397538:T:C	rs16917546	C	Intronic	ZNF365	-0.004
19	16426657	19:16426657:A:G	rs386869	G	Intronic	CTD-2562J15.6	-0.004
11	309127	11:309127:A:G	rs1059091	G	PAVs	IFITM2	0.004
14	94417531	14:94417531:T:C	rs11555542	C	PAVs	ASB2	0.004
2	102966549	2:102966549:A:G	rs10197862	G	Intronic	IL1RL1, IL18R1	-0.004
19	16423979	19:16423979:G:A	rs74366434	A	Intronic	CTD-2562J15.6	0.004
3	3157051	3:3157051:C:T	rs1153462	T	Intronic	IL5RA	-0.003
19	1079959	19:1079959:G:A	rs36084354	A	PAVs	HMHA1	-0.003
5	148206473	5:148206473:G:C	rs1042714	C	PAVs	ADRB2	0.003
2	102950822	2:102950822:A:C	rs13019081	C	Intronic	IL1RL1, IL18R1	0.003
21	36437379	21:36437379:A:G	rs8133412	G	PAVs	RUNX1	0.003
11	327334	11:327334:C:T	rs11246068	T	UTR	IFITM3	-0.003
14	75978121	14:75978121:T:C	rs175707	C	Others		0.003
2	102968209	2:102968209:TC:T	Affx-52334280	T	PTVs	IL1RL1	-0.003
6	32629755	6:32629755:G:A	rs1130399	A	PAVs	HLA-DQB1	0.003
5	131952222	5:131952222:T:C	rs6596086	C	Intronic	RAD50	0.003
5	110401872	5:110401872:T:C	rs1837253	C	Others	TSLP	0.003
5	141523000	5:141523000:C:T	rs1062158	T	Others	NDFIP1	0.003
6	135431318	6:135431318:T:C	rs6920211	C	Others		-0.003
11	76299194	11:76299194:G:T	rs2155219	T	Others	RP11-672A2.7	0.003
7	75442730	7:75442730:T:G	rs2302006	G	PAVs	CCL24	-0.003
7	20506245	7:20506245:T:C	rs11974870	C	Others		-0.003
10	28793239	10:28793239:A:G	rs2998285	G	Others		0.003
6	107420516	6:107420516:A:C	rs4946811	C	PAVs	BEND3	-0.003
10	9053132	10:9053132:C:T	rs962993	T	Others		-0.003
19	40219562	19:40219562:T:G	rs439881	G	Others	CLC	0.003
17	7742601	17:7742601:G:A	rs74480102	A	Others	KDM6B	-0.003
16	67680806	16:67680806:G:A	rs117556162	A	PAVs	RLTPR	0.003
5	131676320	5:131676320:C:T	rs1050152	T	PAVs	SLC22A4	-0.003
3	42906116	3:42906116:T:C	rs2228467	C	PAVs	ACKR2	0.003
21	36715761	21:36715761:C:T	rs9979383	T	Intronic	RUNX1	0.003
17	41926126	17:41926126:C:T	rs72836561	T	PAVs	CD300LG	-0.003
9	94121330	9:94121330:G:A	rs7020893	A	Intronic	AUH	-0.003
1	198598663	1:198598663:G:A	rs7555082	A	Others		-0.003
16	30103160	16:30103160:C:A	rs3809627	A	UTR	TBX6	0.003
9	135869255	9:135869255:T:C	rs77267537	C	Others	GFI1B	0.003
17	44067400	17:44067400:T:C	rs10445337	C	PAVs	MAPT	-0.003
2	111989372	2:111989372:T:G	rs6720394	G	Intronic	MIR4435-1HG	0.003
17	55465771	17:55465771:C:T	rs17834140	T	Intronic	MSI2	-0.002
1	87750853	1:87750853:A:G	rs11161968	G	Others		0.002
1	9713500	1:9713500:G:A	rs4240896	A	UTR	C1orf200	-0.002
17	37931220	17:37931220:A:G	rs7503018	G	Intronic	IKZF3	-0.002
2	213897308	2:213897308:C:T	rs10932459	T	Intronic	IKZF2	-0.002
10	990143	10:990143:C:T	rs61833265	T	Others	RP11-363N22.2	0.002
5	131901225	5:131901225:A:G	rs2244012	G	Intronic	RAD50	0.002
6	32411035	6:32411035:A:C	rs8084	C	PTVs	HLA-DRA	0.002
2	242371101	2:242371101:C:T	rs757978	T	PAVs	FARP2	-0.002
11	57146225	11:57146225:A:G	rs34108746	G	PAVs	PRG3	-0.002
3	188442480	3:188442480:T:C	rs9290877	C	Intronic	LPP	0.002
19	42728836	19:42728836:T:C	rs3810151	C	PAVs	ZNF526	0.002
4	38663660	4:38663660:T:C	rs12507197	C	Intronic	AC021860.1, RP11-617D20.1	0.002
3	3139957	3:3139957:T:C	rs2290610	C	PAVs	IL5RA	-0.002
8	129007207	8:129007207:A:G	rs10956403	G	Intronic	PVT1	-0.002
11	117076972	11:117076972:C:A	rs45574931	A	PAVs	PCSK7	-0.002
12	56435412	12:56435412:G:A	rs705704	A	Others	RP11-603J24.4	0.002
1	65420190	1:65420190:T:C	rs7553101	C	Intronic	JAK1	-0.002
5	35874575	5:35874575:C:T	rs6897932	T	PAVs	IL7R	-0.002
14	23589057	14:23589057:G:A	rs2239633	A	Others	CEBPE	0.002
9	113643100	9:113643100:G:A	rs491749	A	Intronic	LPAR1	0.002
5	169690854	5:169690854:T:C	rs452833	C	Intronic	LCP2	-0.002
17	4531882	17:4531882:T:C	rs17764723	C	Others	ALOX15	-0.002
6	32565331	6:32565331:A:C	rs9270623	C	Others		-0.002
7	37400469	7:37400469:T:C	rs16879645	C	Intronic	ELMO1	0.002
15	67442596	15:67442596:C:T	rs17293632	T	Intronic	SMAD3	0.002
5	133451683	5:133451683:C:A	rs5742913	A	PAVs	TCF7	0.002
6	245466	6:245466:A:G	rs6899887	G	Others		0.002
4	83564536	4:83564536:G:A	rs6812648	A	Intronic	SCD5	-0.002
20	42660286	20:42660286:C:A	rs2179593	A	Intronic	TOX2	-0.002
4	38645460	4:38645460:T:C	rs11937257	C	Intronic	AC021860.1, RP11-617D20.1	0.002
5	131797547	5:131797547:A:G	rs6894249	G	UTR	C5orf56	0.002
11	128186882	11:128186882:T:G	rs10893845	G	Others		0.002
17	4574751	17:4574751:T:A	rs9436	A	PAVs	PELP1	-0.002
17	47402807	17:47402807:C:T	rs12940887	T	Intronic	ZNF652	0.002
12	54685880	12:54685880:C:T	rs35979828	T	Intronic	RP11-968A15.8	-0.002
2	28686332	2:28686332:G:A	rs1581035	A	Intronic	PLB1	-0.002
5	150458146	5:150458146:C:T	rs10036748	T	Intronic	TNIP1	-0.002
10	26727454	10:26727454:G:A	rs2992333	A	Intronic	APBB1IP	-0.002
10	94485211	10:94485211:A:C	rs2497306	C	Others		-0.002
15	61069988	15:61069988:C:T	rs11071559	T	Intronic	RORA	-0.002
16	88573347	16:88573347:G:T	rs17700789	T	Intronic	ZFPM1	0.002
5	131411138	5:131411138:C:T	rs25881	T	Intronic	CSF2	0.002
1	198544313	1:198544313:A:C	rs10494780	C	Others		-0.002
16	11241906	16:11241906:A:G	rs17673553	G	Intronic	CLEC16A	-0.002
10	8107572	10:8107572:A:G	rs10905279	G	Intronic	GATA3	-0.002
3	128306418	3:128306418:T:C	rs6772849	C	Others		-0.002
7	22788274	7:22788274:A:G	rs7801406	G	Others	AC073072.6	-0.002
4	123071608	4:123071608:G:A	rs7691220	A	Others	KIAA1109	0.002

There is no guarantee that our iPGS model selects causal variants. We show the top 100 variants with the largest effect size (BETA). To see 16859 variants included in our iPGS model, please download the iPGS coefficients by clicking the download button. We use the GRCh37/hg19 reference genome.

Follow-up analysis

There are several ways to use the resource in your research. First, you may use our iPGS coefficients and compute individual-level polygenic scores for your cohort. Second, you may also investigate the genetic variants with non-zero coefficients and their annotated genes to learn more about biology by taking advantage of the sparsity of our iPGS models. For your convenience, here we suggest several resources as an example of follow-up analysis. We do not intend to cover all the relevant follow-up analyses.

Using iPGS coefficients

By clicking the download button above, you may download the iPGS coefficients. Our FAQ page shows the description of file format and how you may use iPGS coefficients in your research.

HaploReg

HaploReg is a tool for exploring annotations of the non-coding genome at variants on haplotype blocks. The button above submits the top 100 genetic variants with the largest absolute value of coefficients as a query to HaploReg using the default parameters in HaploReg v4.2 (LD threshold r² >= 1, ChromHMM 15-state model, SiPhy-omega, and GENCODE genes). HaploReg's ability to browse haplotypes is useful here as there is no guarantee that our iPGS model selects causal variants. The 'top 100 variant' cutoff is an arbitrary threshold; we aim to demonstrate how one may investigate the selected variants. Please check Ward and Kellis. Nucleic Acids Res. 2012 and Ward and Kellis. Nucleic Acids Res. 2016 for more information on HaploReg.

GREAT

GREAT: Genomic Regions Enrichment of Annotations Tool evaluates enrichment of pathway and ontology terms. The ability of GREAT to map non-coding genetic variants to their downstream target genes would be suitable for investigating pathway and ontology enrichment of genetic variants selected in our sparse iPGS model. The button above submits the top 1000 genetic variants with the largest absolute value of coefficients as a query to GREAT using the default parameters in GREAT v4.0.4. The 'top 1000 variant' cutoff is an arbitrary threshold; we aim to demonstrate how one may investigate the selected variants. Please check McLean et al. Nat Biotechnol. 2010 and Tanigawa*, Dyer*, and Bejerano. PLoS Comput Biol. 2022 for more information on GREAT.

References

Y. Tanigawa and M. Kellis. Power of inclusion: Enhancing polygenic prediction with admixed individuals. Am J Hum Genet. (2023).
- We introduce our inclusive PGS approach in this publication.
- Supplementary Data Files at figshare (doi: 10.6084/m9.figshare.22905368)
- MIT News article: Making genetic prediction models more inclusive
- Author interview: Inside AJHG: A Chat with Yosuke Tanigawa
J. Qian, Y. Tanigawa, W. Du, M. Aguirre, C. Chang, R. Tibshirani, M. A. Rivas, T. Hastie. A fast and scalable framework for large-scale and ultrahigh-dimensional sparse regression with application to the UK Biobank. PLoS Genet. 16, e1009141 (2020).
- In iPGS, we fit penalized multivariate regression on individual-level data using the BASIL algorithm implemented in the R snpnet package. This publication describes the BASIL algorithm and the R snpnet package.