Monday 28th of April 2025, 13:30-14:30 in HG03.085
He Li
Dynamic Principal Components Modelling of Longitudinal Omics Data
For a longitudinally measured omics dataset, the interest might be identifying a set of variables representing the dynamic structure. PCA approaches enable identifying sets of omics variables representing the cross-sectional structure. On the other hand, univariate analysis with fixed and random effects provides insights into whether a variable changes over time, but ignores the joint distribution. We propose a novel multivariate dynamic probabilistic PCA-approach (DPPCA) which models the scores over time using a mixed model.
For estimation of the parameters, we maximize the log-likelihood using the EM algorithm. Via an extensive simulation study, we evaluate the performance of DPPCA for varying numbers of omics variables, dynamic components and time points. Finally, we apply DPPCA to a longitudinal metabolomics dataset from the TwinsUK study.
The simulation results show that the parameter estimators for the time effect and random intercept variance related to the first component are unbiased. Concerning the data, scree plots are utilized to obtain the dynamic components. Preliminary results show that the components change and correlated over time. In the first component, most metabolites with high weights belong to lipoprotein subclasses.
Key words: Latent variable models; Linear mixed models; Dimension reduction; Metabolomics