Genotype imputation and variability in polygenic risk score estimation

Background Polygenic risk scores (PRSs) are a summarization of an individual's genetic risk for a disease or trait. These scores are being generated in research and commercial settings to study how they may be used to guide healthcare decisions. PRSs should be updated as genetic knowledgebases improve; however, no guidelines exist for their generation or updating. Methods Here, we characterize the variability introduced in PRS calculation by a common computational process used in their generation-genotype imputation. We evaluated PRS variability when performing genotype imputation using 3 different pre-phasing tools (Beagle, Eagle, SHAPEIT) and 2 different imputation tools (Beagle, Minimac4), relative to a WGS-based gold standard. Fourteen different PRSs spanning different disease architectures and PRS generation approaches were evaluated. Results We find that genotype imputation can introduce variability in calculated PRSs at the individual level without any change to the underlying genetic model. The degree of variability introduced by genotype imputation differs across algorithms, where pre-phasing algorithms with stochastic elements introduce the greatest degree of score variability. In most cases, PRS variability due to imputation is minor (< 5 percentile rank change) and does not influence the interpretation of the score. PRS percentile fluctuations are also reduced in the more informative tails of the PRS distribution. However, in rare instances, PRS instability at the individual level can result in singular PRS calculations that differ substantially from a whole genome sequence-based gold standard score. Conclusions Our study highlights some challenges in applying population genetics tools to individual-level genetic analysis including return of results. Rare individual-level variability events are masked by a high degree of overall score reproducibility at the population level. In order to avoid PRS result fluctuations during updates, we suggest that deterministic imputation processes or the average of multiple iterations of stochastic imputation processes be used to generate and deliver PRS results.