Abstract
Information generated from longitudinally-sampled microbial data has the potential to illuminate important aspects of development and progression for many human conditions and diseases. Identifying microbial biomarkers and their time-varying effects can not only advance our understanding of pathogenetic mechanisms, but also facilitate early diagnosis and guide optimal timing of interventions. However, longitudinal predictive modeling of highly noisy and dynamic microbial data (e.g., metagenomics) poses analytical challenges. To overcome these challenges, we introduce a robust and interpretable machine-learning-based longitudinal microbiome analysis framework, LP-Micro, that encompasses: (i) longitudinal microbial feature screening via a polynomial group lasso, (ii) disease outcome prediction implemented via machine learning methods (e.g., XGBoost, deep neural networks), and (iii) interpretable association testing between time points, microbial features, and disease outcomes via permutation feature importance. We demonstrate in simulations that LP-Micro can not only identify incident disease-related microbiome taxa but also offers improved prediction accuracy compared to existing approaches. Applications of LP-Micro in two longitudinal microbiome studies with clinical outcomes of childhood dental disease and weight loss following bariatric surgery yield consistently high prediction accuracy. The identified critical early predictive time points are informative and aligned with clinical expectations.