Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations

NHLBI’s Trans-Omics in Precision Medicine (TOPMed) Consortium

doi:10.1038/s42003-022-03812-z

Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations

NHLBI’s Trans-Omics in Precision Medicine (TOPMed) Consortium

Preventive Medicine

Research output: Contribution to journal › Article › peer-review

39 Scopus citations

Abstract

Polygenic risk scores (PRS) are commonly used to quantify the inherited susceptibility for a trait, yet they fail to account for non-linear and interaction effects between single nucleotide polymorphisms (SNPs). We address this via a machine learning approach, validated in nine complex phenotypes in a multi-ancestry population. We use an ensemble method of SNP selection followed by gradient boosted trees (XGBoost) to allow for non-linearities and interaction effects. We compare our results to the standard, linear PRS model developed using PRSice, LDpred2, and lassosum2. Combining a PRS as a feature in an XGBoost model results in a relative increase in the percentage variance explained compared to the standard linear PRS model by 22% for height, 27% for HDL cholesterol, 43% for body mass index, 50% for sleep duration, 58% for systolic blood pressure, 64% for total cholesterol, 66% for triglycerides, 77% for LDL cholesterol, and 100% for diastolic blood pressure. Multi-ancestry trained models perform similarly to specific racial/ethnic group trained models and are consistently superior to the standard linear PRS models. This work demonstrates an effective method to account for non-linearities and interaction effects in genetics-based prediction models.

Original language	English (US)
Pages (from-to)	856
Number of pages	1
Journal	Communications Biology
Volume	5
Issue number	1
DOIs	https://doi.org/10.1038/s42003-022-03812-z
State	Published - Aug 22 2022

Funding

Molecular data for the Trans-Omics in Precision Medicine (TOPMed) program was supported by the National Heart, Lung and Blood Institute (NHLBI). Genome sequencing for “NHLBI TOPMed: Whole Genome Sequencing and Related Phenotypes in the Framingham Heart Study” (phs000974.v4.p3) was performed at the Broad Institute Genomics Platform (3R01HL092577-06S1, 3U54HG003067-12S2). Genome sequencing for “NHLBI TOPMed: The Jackson Heart Study” (phs000964.v1.p1) was performed at the Northwest Genomics Center (HHSN268201100037C). Genome sequencing for the “NHLBI TOPMed: The Atherosclerosis Risk in Communities Study” (phs001211.v3.p2) was performed at the Broad Institute Genomics Platform (3R01HL092577-06S1) and the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201500015C, 3U54HG003273-12S2). Genome sequencing for “NHLBI TOPMed: Coronary Artery Risk Development in Young Adults Study” (phs001612.v1.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Cleveland Family Study” (phs000954.v3.p2) was performed at the Northwest Genomics Center (3R01HL098433-05S1, HHSN268201600032I). Genomics sequencing for “NHLBI TOPMed: Cardiovascular Health Study” (phs001368.v2.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (3U54HG003273-12S2, HHSN268201500015C, HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Hispanic Community Health Study/Study of Latinos” (phs001395.v1.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Multi-Ethnic Study of Atherosclerosis” (phs001416.v2.p1) was performed at Broad Institute Genomics Platform (HHSN268201500014C, 3U54HG003067-13S1). Core support including centralized genomic read mapping and genotype calling, along with variant quality metrics and filtering were provided by the TOPMed Informatics Research Center (3R01HL-117626-02S1; contract HHSN268201800002I). Core support including phenotype harmonization, data management, sample-identity QC, and general program coordination were provided by the TOPMed Data Coordinating Center (R01HL-120393; U01HL-120393; contract HHSN268201800001I). We gratefully acknowledge the studies and participants who provided biological samples and data for TOPMed. The Genome Sequencing Program (GSP) was funded by the National Human Genome Research Institute (NHGRI), the National Heart, Lung, and Blood Institute (NHLBI), and the National Eye Institute (NEI). The GSP Coordinating Center (U24 HG008956) contributed to cross-program scientific initiatives and provided logistical and general study coordination. The Centers for Common Disease Genomics (CCDG) program was supported by NHGRI and NHLBI, and whole genome-sequencing was performed at the Baylor College of Medicine Human Genome Sequencing Center (UM1 HG008898 and R01HL059367). M.E., T.S., and S.R. are supported by NHLBI grants R35HL135818 to SR. G.M.P. is supported by R01HL142711 from the NHLBI. The project described was supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through grant KL2TR002490 to L.M.R. P.d.V. was supported by American Heart Association grant number 18CDA34110116 and NHLBI grant number R01HL146860. The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services. Molecular data for the Trans-Omics in Precision Medicine (TOPMed) program was supported by the National Heart, Lung and Blood Institute (NHLBI). Genome sequencing for “NHLBI TOPMed: Whole Genome Sequencing and Related Phenotypes in the Framingham Heart Study” (phs000974.v4.p3) was performed at the Broad Institute Genomics Platform (3R01HL092577-06S1, 3U54HG003067-12S2). Genome sequencing for “NHLBI TOPMed: The Jackson Heart Study” (phs000964.v1.p1) was performed at the Northwest Genomics Center (HHSN268201100037C). Genome sequencing for the “NHLBI TOPMed: The Atherosclerosis Risk in Communities Study” (phs001211.v3.p2) was performed at the Broad Institute Genomics Platform (3R01HL092577-06S1) and the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201500015C, 3U54HG003273-12S2). Genome sequencing for “NHLBI TOPMed: Coronary Artery Risk Development in Young Adults Study” (phs001612.v1.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Cleveland Family Study” (phs000954.v3.p2) was performed at the Northwest Genomics Center (3R01HL098433-05S1, HHSN268201600032I). Genomics sequencing for “NHLBI TOPMed: Cardiovascular Health Study” (phs001368.v2.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (3U54HG003273-12S2, HHSN268201500015C, HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Hispanic Community Health Study/Study of Latinos” (phs001395.v1.p1) was performed at the Baylor College of Medicine Human Genome Sequencing Center (HHSN268201600033I). Genome sequencing for “NHLBI TOPMed: Multi-Ethnic Study of Atherosclerosis” (phs001416.v2.p1) was performed at Broad Institute Genomics Platform (HHSN268201500014C, 3U54HG003067-13S1). Core support including centralized genomic read mapping and genotype calling, along with variant quality metrics and filtering were provided by the TOPMed Informatics Research Center (3R01HL-117626-02S1; contract HHSN268201800002I). Core support including phenotype harmonization, data management, sample-identity QC, and general program coordination were provided by the TOPMed Data Coordinating Center (R01HL-120393; U01HL-120393; contract HHSN268201800001I). We gratefully acknowledge the studies and participants who provided biological samples and data for TOPMed. The Genome Sequencing Program (GSP) was funded by the National Human Genome Research Institute (NHGRI), the National Heart, Lung, and Blood Institute (NHLBI), and the National Eye Institute (NEI). The GSP Coordinating Center (U24 HG008956) contributed to cross-program scientific initiatives and provided logistical and general study coordination. The Centers for Common Disease Genomics (CCDG) program was supported by NHGRI and NHLBI, and whole genome-sequencing was performed at the Baylor College of Medicine Human Genome Sequencing Center (UM1 HG008898 and R01HL059367). M.E., T.S., and S.R. are supported by NHLBI grants R35HL135818 to SR. G.M.P. is supported by R01HL142711 from the NHLBI. The project described was supported by the National Center for Advancing Translational Sciences, National Institutes of Health, through grant KL2TR002490 to L.M.R. P.d.V. was supported by American Heart Association grant number 18CDA34110116 and NHLBI grant number R01HL146860. The views expressed in this manuscript are those of the authors and do not necessarily represent the views of the National Heart, Lung, and Blood Institute; the National Institutes of Health; or the U.S. Department of Health and Human Services.

ASJC Scopus subject areas

Medicine (miscellaneous)
General Biochemistry, Genetics and Molecular Biology
General Agricultural and Biological Sciences

Access to Document

10.1038/s42003-022-03812-z

Cite this

@article{eca648b79f6748a18b469cba7d273a83,

title = "Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations",

abstract = "Polygenic risk scores (PRS) are commonly used to quantify the inherited susceptibility for a trait, yet they fail to account for non-linear and interaction effects between single nucleotide polymorphisms (SNPs). We address this via a machine learning approach, validated in nine complex phenotypes in a multi-ancestry population. We use an ensemble method of SNP selection followed by gradient boosted trees (XGBoost) to allow for non-linearities and interaction effects. We compare our results to the standard, linear PRS model developed using PRSice, LDpred2, and lassosum2. Combining a PRS as a feature in an XGBoost model results in a relative increase in the percentage variance explained compared to the standard linear PRS model by 22% for height, 27% for HDL cholesterol, 43% for body mass index, 50% for sleep duration, 58% for systolic blood pressure, 64% for total cholesterol, 66% for triglycerides, 77% for LDL cholesterol, and 100% for diastolic blood pressure. Multi-ancestry trained models perform similarly to specific racial/ethnic group trained models and are consistently superior to the standard linear PRS models. This work demonstrates an effective method to account for non-linearities and interaction effects in genetics-based prediction models.",

author = "{NHLBI{\textquoteright}s Trans-Omics in Precision Medicine (TOPMed) Consortium} and Michael Elgart and Genevieve Lyons and Santiago Romero-Brufau and Nuzulul Kurniansyah and Brody, {Jennifer A.} and Xiuqing Guo and Lin, {Henry J.} and Laura Raffield and Yan Gao and Han Chen and {de Vries}, Paul and Lloyd-Jones, {Donald M.} and Lange, {Leslie A.} and Peloso, {Gina M.} and Myriam Fornage and Rotter, {Jerome I.} and Rich, {Stephen S.} and Morrison, {Alanna C.} and Psaty, {Bruce M.} and Daniel Levy and Susan Redline and Tamar Sofer",

note = "Publisher Copyright: {\textcopyright} 2022. The Author(s).",

year = "2022",

month = aug,

day = "22",

doi = "10.1038/s42003-022-03812-z",

language = "English (US)",

volume = "5",

pages = "856",

journal = "Communications Biology",

issn = "2399-3642",

publisher = "Springer Nature",

number = "1",

}

TY - JOUR

T1 - Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations

AU - NHLBI’s Trans-Omics in Precision Medicine (TOPMed) Consortium

AU - Elgart, Michael

AU - Lyons, Genevieve

AU - Romero-Brufau, Santiago

AU - Kurniansyah, Nuzulul

AU - Brody, Jennifer A.

AU - Guo, Xiuqing

AU - Lin, Henry J.

AU - Raffield, Laura

AU - Gao, Yan

AU - Chen, Han

AU - de Vries, Paul

AU - Lloyd-Jones, Donald M.

AU - Lange, Leslie A.

AU - Peloso, Gina M.

AU - Fornage, Myriam

AU - Rotter, Jerome I.

AU - Rich, Stephen S.

AU - Morrison, Alanna C.

AU - Psaty, Bruce M.

AU - Levy, Daniel

AU - Redline, Susan

AU - Sofer, Tamar

PY - 2022/8/22

Y1 - 2022/8/22

N2 - Polygenic risk scores (PRS) are commonly used to quantify the inherited susceptibility for a trait, yet they fail to account for non-linear and interaction effects between single nucleotide polymorphisms (SNPs). We address this via a machine learning approach, validated in nine complex phenotypes in a multi-ancestry population. We use an ensemble method of SNP selection followed by gradient boosted trees (XGBoost) to allow for non-linearities and interaction effects. We compare our results to the standard, linear PRS model developed using PRSice, LDpred2, and lassosum2. Combining a PRS as a feature in an XGBoost model results in a relative increase in the percentage variance explained compared to the standard linear PRS model by 22% for height, 27% for HDL cholesterol, 43% for body mass index, 50% for sleep duration, 58% for systolic blood pressure, 64% for total cholesterol, 66% for triglycerides, 77% for LDL cholesterol, and 100% for diastolic blood pressure. Multi-ancestry trained models perform similarly to specific racial/ethnic group trained models and are consistently superior to the standard linear PRS models. This work demonstrates an effective method to account for non-linearities and interaction effects in genetics-based prediction models.

AB - Polygenic risk scores (PRS) are commonly used to quantify the inherited susceptibility for a trait, yet they fail to account for non-linear and interaction effects between single nucleotide polymorphisms (SNPs). We address this via a machine learning approach, validated in nine complex phenotypes in a multi-ancestry population. We use an ensemble method of SNP selection followed by gradient boosted trees (XGBoost) to allow for non-linearities and interaction effects. We compare our results to the standard, linear PRS model developed using PRSice, LDpred2, and lassosum2. Combining a PRS as a feature in an XGBoost model results in a relative increase in the percentage variance explained compared to the standard linear PRS model by 22% for height, 27% for HDL cholesterol, 43% for body mass index, 50% for sleep duration, 58% for systolic blood pressure, 64% for total cholesterol, 66% for triglycerides, 77% for LDL cholesterol, and 100% for diastolic blood pressure. Multi-ancestry trained models perform similarly to specific racial/ethnic group trained models and are consistently superior to the standard linear PRS models. This work demonstrates an effective method to account for non-linearities and interaction effects in genetics-based prediction models.

UR - http://www.scopus.com/inward/record.url?scp=85136172218&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85136172218&partnerID=8YFLogxK

U2 - 10.1038/s42003-022-03812-z

DO - 10.1038/s42003-022-03812-z

M3 - Article

C2 - 35995843

AN - SCOPUS:85136172218

SN - 2399-3642

VL - 5

SP - 856

JO - Communications Biology

JF - Communications Biology

IS - 1

ER -

Non-linear machine learning models incorporating SNPs and PRS improve polygenic prediction in diverse human populations

Abstract

Funding

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this