Computationally efficient methods for estimating phenome-wide coheritability of multi-type phenotypes using biobank data.

Biobank data provide a rich source for studying the coheritability of multiple disease phenotypes, which can provide information on shared genetic etiology. However, the large number and heterogeneous types of phenotypes (e.g., continuous, discrete, time-to-event) pose significant statistical and computational challenges for estimating coheritability. In this work, we propose a unified modeling framework with latent random effects distinguishing genetic and family-shared environmental contributions to variation across multi-type phenotypes. To avoid high-dimensional integrals over many phenotypes and family members in joint likelihood approaches, we develop a computationally efficient procedure by first maximizing the marginal likelihood function for each individual phenotype and then estimating the coheritability using only pairs of phenotypes. We apply our method to analyze the heritability and coheritability of 290 phenotypes obtained from the UK Biobank. We find that a substantial number of phenotype pairs present statistically significant genetic coheritability.