[paper] Design, Processing, and Modeling for Longitudinal Multiomics Microbiome Data

Our mini review is out in Frontiers in Cellular and Infection Microbiology! Written with Margaret Thairu and Kris Sankaran, it is a practical guide to analyzing microbiome studies that follow many omic layers over time.

Why this is hard

Microbial communities interact with each other and with their hosts, and multiomic studies try to link these interactions to phenotypes. Collecting the data has become much easier, but the statistical methods for integrating it have lagged behind. Adding time on top of multiple omic layers brings two kinds of problems.

Time. Repeated samples from the same subject are not independent, and an effect may only appear in a certain window or pass along as a delayed cascade. In practice, samples are rarely collected on a neat, evenly spaced schedule. Many methods still assume uniform spacing, so irregular sampling times need interpolation or models that can handle them directly.

Crossing omic layers. Each layer (16S, metagenomics, metatranscriptomics, metaproteomics, metabolomics) has its own quirks: compositional counts, different kinds of zeros, and its own distributions. Cost also limits how often each layer is measured, so the layers end up sampled at different frequencies and the combined dataset has gaps. Earlier reviews treated cross-omic integration and time-aware modeling mostly separately, not together.

What the review covers

• Study design and data management: blocking, randomization, controls and replication, plus tools such as MultiAssayExperiment and workflow managers like Snakemake and targets.
• Technical variation: interpolation for irregular time points, normalization, and batch correction.
• Integration strategies: late integration (model each layer separately, then relate them), early integration (one joint model), and intermediate or “mosaic” approaches for incomplete data.
• Temporal and cross-layer models: mixed-effects models, interval-specific tests, functional PCA, MOFA2/MEFISTO, DIABLO, dynamic Bayesian networks, network methods, and causal mediation analysis.

One takeaway I like: the right method depends on the question and on how densely you sampled. Simple repeated-measures approaches work well for finding biomarkers. Interval tests tell you when an effect happens. Full generative models need denser data to work.

Direct Link