Deep phenotyping can enhance the power of genetic analysis, including genome-wide association studies (GWAS), but the occurrence of missing phenotypes compromises the potential of such resources. Although many phenotypic imputation methods have been developed, the accurate imputation of millions of individuals remains challenging. In the present study, we have developed a multi-phenotype imputation method based on mixed fast random forest (PIXANT) by leveraging efficient machine learning (ML)-based algorithms. We demonstrate by extensive simulations that PIXANT is reliable, robust and highly resource-efficient. We then apply PIXANT to the UKB data of 277,301 unrelated White British citizens and 425 traits, and GWAS is subsequently performed on the imputed phenotypes, 18.4% more GWAS loci are identified than before imputation (8710 vs 7355). The increased statistical power of GWAS identified some additional candidate genes affecting heart rate, such as RNF220, SCN10A, and RGS6, suggesting that the use of imputed phenotype data from a large cohort may lead to the discovery of additional candidate genes for complex traits.
