Time-Informed Dimensionality Reduction for Longitudinal Microbiome Studies

Abstract

AbstractComplex dynamics of microbial communities underlie their essential roles in health and disease, but our understanding of these dynamics remains incomplete. To bridge this gap, longitudinal microbiome data are being rapidly generated, yet their power is limited by technical challenges in design and analysis, such as missing temporal samples, complex correlation structure, and high dimensionality. Here, we present TEMPoral TEnsor Decomposition (TEMPTED), the only time-informed dimensionality reduction method that extracts the underlying microbial dynamics while overcoming the statistical challenges posed by this type of data. TEMPTED facilitates beta-diversity analysis at both sample- and subject-level and enables the transfer of the learned low-dimensional representation from training data to unseen test data. In data-driven simulations, TEMPTED enables host phenotype classification at 90% accuracy compared to random using existing methods. In real data we show that TEMPTED deconvolutes the dynamics of the vaginal microbiome during pregnancy, allowing for the detection of microbial signatures associated with term and preterm births that are reproducible across datasets.