Improving polygenic prediction from summary data by learning patterns of effect sharing across multiple phenotypes

Polygenic prediction of complex trait phenotypes has become important in human genetics, especially in the context of precision medicine. Recently, Morgante et al . introduced mr.mash , a flexible and computationally efficient method that models multiple phenotypes jointly and leverages sharing of effects across such phenotypes to improve prediction accuracy. However, a drawback of mr.mash is that it requires individual-level data, which are often not publicly available. In this work, we introduce mr.mash-rss , an extension of the mr.mash model that requires only summary statistics from Genome-Wide Association Studies (GWAS) and linkage disequilibrium (LD) estimates from a reference panel. By using summary data, we achieve the twin goal of increasing the applicability of the mr.mash model to data sets that are not publicly available and making it scalable to biobank-size data. Through simulations, we show that mr.mash-rss is competitive with, and often outperforms, current state-of-the-art methods for single- and multi-phenotype polygenic prediction in a variety of scenarios that differ in the pattern of effect sharing across phenotypes, the number of phenotypes, the number of causal variants, and the genomic heritability. We also present a real data analysis of 16 blood cell phenotypes in UK Biobank, showing that mr.mash-rss achieves higher prediction accuracy than competing methods for the majority of traits, especially when the data has smaller sample size. Author summary Polygenic prediction refers to the use of an individual’s genetic information ( i.e ., genotypes) to predict traits ( i.e ., phenotypes), which are often of medical relevance. It is known that some phenotypes are related and are affected by the same genotypes. When this is the case, it is possible to improve the accuracy of predictions by using methods that model multiple phenotypes jointly and account for shared effects. mr.mash is a recently developed multi-phenotype method that can learn which effects are shared and has been shown to improve prediction. However, mr.mash requires large data sets of genetic and phenotypic information collected at the individual level. Such data are often unavailable due to privacy concerns, or are difficult to work with due to the computational resources needed to analyze data of this size. Our work extends mr.mash to require only summary statistics from Genome-Wide Association Studies instead of individual-level data, which are usually publicly available. In addition, the computations using summary statistics do not depend on sample size, making the newly developed mr.mash-rss scalable to extremely large data sets. Using simulations and real data analysis, we show that our method is competitive with other methods for polygenic prediction.