Polygenic risk prediction: why and when out-of-sample prediction R2 can exceed SNP-based heritability

Xiaotong Wang, Alicia Walker, Joana Revez, Guiyan Ni, Mark J. Adams, Andrew M. McIntosh, Peter M. Visscher, Naomi R. Wray, Stephan Ripke, Manuel Mattheisen, Maciej Trzaskowski, Enda M. Byrne, Abdel Abdellaoui, Esben Agerbo, Tracy Air, Till F. M. Andlauer, Silviu‐Alin Bacanu, Marie Bækvad‐Hansen, Aartjan T.f. Beekman, Tim B. Bigdeli, Elisabeth B. Binder, Julien Bryois, Henriette N. Buttenschøn, Jonas Bybjerg‐Grauholm, Na Cai, Enrique Castelao, Jane Christensen, Toni‐Kim Clarke, Jonathan R. I. Coleman, Lucía Colodro‐Conde, Baptiste Couvy‐Duchesne, Nick Craddock, Gregory E. Crawford, Gail Davies, Franziska Degenhardt, Eske M. Derks, Neşe Direk, Conor V. Dolan, Erin C. Dunn, Thalia C. Eley, Valentina Escott‐Price, Farnush Farhadi Hassan Kiadeh, Hilary K. Finucane, Jerome C. Foo, Andreas J. Forstner, Josef Frank, Héléna A. Gaspar, Michael Gill, Fernando S. Goes, Scott D. Gordon, Jakob Grove, Lynsey S. Hall, Christine Søholm Hansen, Thomas Folkmann Hansen, Stefan Herms, Ian B. Hickie, Per Hoffmann, Georg Homuth, Carsten Horn, Jouke‐Jan Hottenga, David M. Hougaard, David M. Howard, Marcus Ising, Rick Jansen, Ian Jones, Lisa Jones, Eric Jorgenson, James A. Knowles, Isaac S. Kohane, Julia Kraft, Warren W. Kretzschmar, Zoltán Kutalik, Yihan Li, Penelope A. Lind, Donald J. MacIntyre, Dean F. MacKinnon, Robert Maier, Wolfgang Maier, Jonathan Marchini, Hamdi Mbarek, Patrick J. McGrath, Peter McGuffin, Sarah E. Medland, Divya Mehta, Christel M. Middeldorp, Evelin Mihailov, Yuri Milaneschi, Lili Milani, Francis M. Mondimore, Grant W. Montgomery, Sara Mostafavi, Niamh Mullins, Matthias Nauck, Bernard Ng, Michel G. Nivard, Dale R. Nyholt, Paul F. O’Reilly, Högni ÓskarssonThe American Journal of Human Genetics 2023 · DOI

Abstract

In polygenic score (PGS) analysis, the coefficient of determination (R<sup>2</sup>) is a key statistic to evaluate efficacy. R<sup>2</sup> is the proportion of phenotypic variance explained by the PGS, calculated in a cohort that is independent of the genome-wide association study (GWAS) that provided estimates of allelic effect sizes. The SNP-based heritability (h<sub>SNP</sub><sup>2</sup>, the proportion of total phenotypic variances attributable to all common SNPs) is the theoretical upper limit of the out-of-sample prediction R<sup>2</sup>. However, in real data analyses R<sup>2</sup> has been reported to exceed h<sub>SNP</sub><sup>2</sup>, which occurs in parallel with the observation that h<sub>SNP</sub><sup>2</sup> estimates tend to decline as the number of cohorts being meta-analyzed increases. Here, we quantify why and when these observations are expected. Using theory and simulation, we show that if heterogeneities in cohort-specific h<sub>SNP</sub><sup>2</sup> exist, or if genetic correlations between cohorts are less than one, h<sub>SNP</sub><sup>2</sup> estimates can decrease as the number of cohorts being meta-analyzed increases. We derive conditions when the out-of-sample prediction R<sup>2</sup> will be greater than h<sub>SNP</sub><sup>2</sup> and show the validity of our derivations with real data from a binary trait (major depression) and a continuous trait (educational attainment). Our research calls for a better approach to integrating information from multiple cohorts to address issues of between-cohort heterogeneity.
BibTeX
@article{wang2023polygenic,
  author = {Xiaotong Wang and Alicia Walker and Joana Revez and Guiyan Ni and Mark J. Adams and Andrew M. McIntosh and Peter M. Visscher and Naomi R. Wray and Naomi R. Wray and Stephan Ripke and Manuel Mattheisen and Maciej Trzaskowski and Enda M. Byrne and Abdel Abdellaoui and Mark J. Adams and Esben Agerbo and Tracy Air and Till F. M. Andlauer and Silviu‐Alin Bacanu and Marie Bækvad‐Hansen and Aartjan T.F. Beekman and Tim B. Bigdeli and Elisabeth B. Binder and Julien Bryois and Henriette N. Buttenschøn and Jonas Bybjerg‐Grauholm and Na Cai and Enrique Castelao and Jane Christensen and Toni‐Kim Clarke and Jonathan R. I. Coleman and Lucía Colodro‐Conde and Baptiste Couvy‐Duchesne and Nick Craddock and Gregory E. Crawford and Gail Davies and Franziska Degenhardt and Eske M. Derks and Neşe Direk and Conor V. Dolan and Erin C. Dunn and Thalia C. Eley and Valentina Escott‐Price and Farnush Farhadi Hassan Kiadeh and Hilary K. Finucane and Jerome C. Foo and Andreas J. Forstner and Josef Frank and Héléna A. Gaspar and Michael Gill and Fernando S. Goes and Scott D. Gordon and Jakob Grove and Lynsey S. Hall and Christine Søholm Hansen and Thomas Folkmann Hansen and Stefan Herms and Ian B. Hickie and Per Hoffmann and Georg Homuth and Carsten Horn and Jouke‐Jan Hottenga and David M. Hougaard and David M. Howard and Marcus Ising and Rick Jansen and Ian Jones and Lisa Jones and Eric Jorgenson and James A. Knowles and Isaac S. Kohane and Julia Kraft and Warren W. Kretzschmar and Zoltán Kutalik and Yihan Li and Penelope A. Lind and Donald J. MacIntyre and Dean F. MacKinnon and Robert Maier and Wolfgang Maier and Jonathan Marchini and Hamdi Mbarek and Patrick J. McGrath and Peter McGuffin and Sarah E. Medland and Divya Mehta and Christel M. Middeldorp and Evelin Mihailov and Yuri Milaneschi and Lili Milani and Francis M. Mondimore and Grant W. Montgomery and Sara Mostafavi and Niamh Mullins and Matthias Nauck and Bernard Ng and Michel G. Nivard and Dale R. Nyholt and Paul F. O’Reilly and Högni Óskarsson},
  title = {{Polygenic risk prediction: why and when out-of-sample prediction R2 can exceed SNP-based heritability}},
  journal = {The American Journal of Human Genetics},
  year = {2023},
  doi = {10.1016/j.ajhg.2023.06.006},
}

Loading related content...

Figure