Decomposition of Shared Latent Factors Using Bayesian Multi-morbidity Dependency Maps

Abstract

The use of multiple diseases and complex phenotypic descriptors is a new trend of genetic association analysis, motivated by the pathway diseases and network medicine paradigms. Comorbidity information is an important resource in this exploration of shared molecular background. To extend the current pairwise, correlation based methods, we investigate a systems-based approach for the use of separated large-scale multi-morbidity data to explore common latent factors of related diseases. We constructed a multi-morbidity dataset from the UK Biobank by filtering rare diseases. In the first phase of our method, we use a Markov Chain Monte Carlo method over Bayesian networks to construct a Bayesian dependency map, which is confounded with many known factors. In the second phase, the method could incorporate prior causal information between the diseases and information about the known confounding by demographic, medical, genetic, environmental factors. The difference between the known causal and confounding relations and the observed dependencies is used to bind the extent of further latent factors. This reconstruction of the shared latent factors happens hierarchically in a top-down fashion, terminating with the identification of latent factors for pair of diseases. We compare our method with other comorbidity methods and systems-based network approaches in the field of psychiatry, focusing on depression and anxiety. We demonstrate the use of molecular, symptomatic and environmental knowledge bases to interpret the reconstructed latent factors. This research has been conducted using the UK Biobank Resource.