Dynamic Factor Analysis for Sparse and Irregular Longitudinal Data: An Application to Metabolite Measurements in a...

Area of Science:

• Biostatistics
• Systems Biology
• Computational Biology

Background:

• Factor analysis (FA) models biological pathways as latent factors driving observed biomarker activity.
• Classical FA assumes factor independence, which is often violated in complex biological systems with interacting pathways.
• Longitudinal biological data, such as metabolite measurements, are frequently sparse and irregularly collected.

Purpose of the Study:

• To develop a dynamic factor analysis (DFA) model that accounts for cross-correlations between interacting biological pathways.
• To address challenges posed by sparse longitudinal data and mitigate overfitting.
• To provide a computationally efficient algorithm for parameter estimation.

Main Methods:

• Proposed a dynamic factor analysis model incorporating a multi-output Gaussian process (MOGP) prior for factor trajectories to capture pathway interactions.
• Introduced a roughness penalty on MOGP hyperparameters and allowed non-zero mean functions to handle data sparsity and prevent overfitting.
• Developed a scalable stochastic expectation maximization (StEM) algorithm for efficient and stable estimation of MOGP hyperparameters.

Main Results:

• The StEM algorithm demonstrated superior performance, being 20 times faster and yielding more accurate and stable MOGP hyperparameter estimates compared to existing methods in simulations.
• The methodology successfully identified a kynurenine pathway significantly associated with clinical severity in COVID-19 patients.
• The study uncovered a novel role for the biomarker taurine in the context of COVID-19 disease progression.

Conclusions:

• The proposed dynamic factor analysis model with MOGP priors effectively captures interacting biological pathways from sparse longitudinal data.
• The StEM algorithm offers a scalable and efficient solution for parameter estimation in such models.
• The findings provide new insights into the biological mechanisms underlying COVID-19 severity, highlighting the kynurenine pathway and taurine.