Understanding When Poisson Log-Normal Models Outperform Penalized Poisson Regression for Microbiome Count Data

2026-04-04

AI Summary

This study addresses the lack of clear guidelines for choosing between multivariate count models that account for latent dependencies—such as the Poisson log-normal (PLN) model—and simpler penalized Poisson regression (e.g., GLMNet) in microbiome count data analysis. The authors propose a unified leave-one-out cross-validation framework to systematically evaluate PLN and GLMNet across large-scale real datasets in terms of count prediction and microbial interaction network inference. Results demonstrate that PLN consistently outperforms GLMNet in most prediction tasks, with improvements up to 38%, and that networks derived from PLN (PLNNetwork) are well-suited for inferring generalized undirected microbial interactions. In contrast, GLMNet excels at capturing local or directed effects. This work further identifies key factors—such as the sample-to-taxon ratio—that govern model selection, establishing the first empirical benchmark for microbial interaction network modeling.

Abstract

Multivariate count models are often justified by their ability to capture latent dependence, but researchers receive little guidance on when this added structure improves on simpler penalized marginal Poisson regression. We study this question using real microbiome data under a unified held-out evaluation framework. For count prediction, we compare PLN and GLMNet(Poisson) on 20 datasets spanning 32 to 18,270 samples and 24 to 257 taxa, using held-out Poisson deviance under leave-one-taxon-out prediction with 3-fold sample cross-validation rather than synthetic or in-sample criteria. For network inference, we compare PLNNetwork and GLMNet(Poisson) neighborhood selection on five publicly available datasets with experimentally validated microbial interaction truth. PLN outperforms GLMNet(Poisson) on most count-prediction datasets, with gains up to 38 percent. The primary predictor of the winner is the sample-to-taxon ratio, with mean absolute correlation as the strongest secondary signal and overdispersion as an additional predictor. PLNNetwork performs best on broad undirected interaction benchmarks, whereas GLMNet(Poisson) is better aligned with local or directional effects. Taken together, these results provide guidance for choosing between latent multivariate count models and penalized Poisson regression in biological count prediction and interaction recovery.

Problem

• microbiome count data
• Poisson Log-Normal model
• penalized Poisson regression
• model comparison
• sample-to-taxon ratio

Innovation

• Poisson Log-Normal
• penalized Poisson regression
• microbiome count data
• held-out evaluation
• network inference