A Bayesian functional principal component analysis framework for longitudinal genome-wide association studies

Introduction: When interrogated along with genetic data, large clinical longitudinal datasets have potential to yield new scientific insights on the genetic contribution to temporal patterns of disease progression and response to treatments. However, existing approaches for longitudinal genome-wide association studies (GWAS) are not suited to unlock the full power of large datasets, as they typically rely on strong distributional assumptions on the longitudinal outcome.
 
Methods: We propose to reframe the longitudinal GWAS problem as a joint latent variable estimation and regression problem in which genetic variants influence longitudinal trajectories via effects on functional latent variables. We leverage existing Bayesian inference frameworks for functional principal component analysis (FPCA) and sparse spike-and-slab regression to develop a two-stage variational inference scheme for this model that conveys uncertainty from the principal component estimation into the second-stage regression.
 
Results: Using simulations, we show that our approach is both scalable and accurate, and that our modelling approach can recover SNPs that influence latent dynamics underlying longitudinal trajectories. We further demonstrate the usefulness and applicability of our framework in a study of genetic effects on longitudinal outcomes in the UK biobank.
 
Conclusion: We present a modelling scheme that has the computational scalability and flexibility to take advantage of large datasets for longitudinal modelling, while rigorously handling estimation uncertainty.

Daniel Temko, Tui Nolan, Sylvia Richardson, Hélène Ruffieux