Abstract
SummaryStudies of disease incidence have identified thousands of genetic loci associated with complex traits. However, many diseases occur in combinations that can point to systemic dysregulation of underlying processes that affect multiple traits. We have developed a data-driven method for identifying such multimorbidities from routine healthcare data that combines topic modelling through Bayesian binary non-negative matrix factorization with an informative prior derived from the hierarchical ICD10 coding system. Through simulation we show that the method, treeLFA, typically outperforms both Latent Dirichlet Allocation (LDA) and topic modelling with uninformative priors in terms of inference accuracy and generalisation to test data, and is robust to moderate deviation between the prior and reality. By applying treeLFA to data from UK Biobank we identify a range of multimorbidity clusters in the form of disease topics ranging from well-established combinations relating to metabolic syndrome, arthropathies and cancers, to other less well-known ones, and a disease-free topic. Through genetic association analysis of inferred topic weights (topic-GWAS) and single diseases we find that topic-GWAS typically finds a much smaller, but only partially-overlapping, set of variants compared to GWAS of constituent disease codes. We validate the genetic loci (only) associated with topics through a range of approaches. Particularly, with the construction of PRS for topics, we find that compared to LDA, treeLFA achieves better prediction performance on independent test data. Overall, our findings indicate that topic models are well suited to characterising multimorbidity patterns, and different topic models have their own unique strengths. Moreover, genetic analysis of multimorbidity patterns can provide insight into the aetiology of complex traits that cannot be determined from the analysis of constituent traits alone.
Publisher
Cold Spring Harbor Laboratory