Statistical modelling of non-linear effects in genetic prediction models

Hannah Klinkhammer, Phlipps-University Marburg

Co-authors: Christian Staerk, TU Dortmund; Carlo Maj, Philipps-University Marburg; Qiong Wu, Philipps-University Marburg; Andreas Mayr, Philipps-University Marburg

Abstract: Polygenic prediction models are based on a large number of genetic variants that have small to medium effect sizes. Typically, in genome-wide association studies (GWAS) as well as in most polygenic score (PGS) approaches, only linear effects of genetic variants are considered. However, recent studies suggest that variants with a recessive or dominant effect pattern might therefore be missed. In previous work, we have introduced the statistical boosting framework snpboost which enables the construction of multivariable PGS directly from individual-level genotype data instead of relying on GWAS summary statistics. So far, snpboost has been limited to linear variant effects. In this work, we incorporate P-splines and thus enable for the first time fitting of smooth variant effects. The approach is illustrated on data from the UK Biobank.