Phylogeny-Guided Global-Local Shrinkage for Microbiome Feature Selection
- 1Case Western Reserve University
Abstract
Microbiome feature selection is challenging because sequencing counts are sparseand compositional, while evolutionarily related taxa may exhibit similar associations with an exposure. We propose phylogeny-guided global--local shrinkage (PGL), a Bayesian regression method for identifying exposure-associated differences in the relative abundances of individual amplicon sequence variants (ASVs). A multinomial logit-normal working model analyzes relative abundance among ASVs detected in each sample. A zero-sum constraint gives the regression coefficients a log-ratio interpretation relative to the mean across retained ASVs. Global--local shrinkage accommodates weak and large associations, while an estimated weight controls information sharing along a supplied phylogenetic tree. Simulation ablation studies showed that phylogenetic borrowing improved detection power and estimation of nonzero coefficients when associations aligned with the tree, with little change under a tree-independent generator. Estimated borrowing decreased as the supplied tree was corrupted. Comparisons with benchmark methods revealed different trade-offs between detection power and false discovery rates across simulation settings. In an HIV oral microbiome application, ablation analysis showed that phylogenetic borrowing preserved the overall pattern of estimated associations while changing posterior uncertainty and selecting three additional ASVs. PGL provides ASV-level inference with adaptive information sharing across evolutionarily related features.
Related
- QBayMic: Quantum-coupled variational Bayes for clustering and feature selection in low-signal microbiome data
- Creator 1.0: Phylogeny-Guided Stagewise Pretraining for Plant Genome Foundation Models across 1,000 Species
- Pathogen-mediated selection and phylogeny shape MHC architecture in 33 bat genomes
- A multimodal deep learning for rice crop health monitoring using vegetation index feature selection
- ROOFS: RObust biOmarker Feature Selection
BibTeX
@misc{zhu2026phylogenyguided,
title = {{Phylogeny-Guided Global-Local Shrinkage for Microbiome Feature Selection}},
author = {Z. Zhu and S. Zhang and X. Tang and L. Zhang},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.28.762580},
url = {https://www.biorxiv.org/content/10.64898/2026.10.28.762580v1},
note = {CC BY-NC-ND},
}
RIS
TY - UNPB
TI - Phylogeny-Guided Global-Local Shrinkage for Microbiome Feature Selection
AU - Z. Zhu
AU - S. Zhang
AU - X. Tang
AU - L. Zhang
PY - 2026
DA - 2026/10/28
PB - bioRxiv
DO - 10.64898/2026.10.28.762580
UR - https://www.biorxiv.org/content/10.64898/2026.10.28.762580v1
AB - Microbiome feature selection is challenging because sequencing counts are sparseand compositional, while evolutionarily related taxa may exhibit similar associations with an exposure. We propose phylogeny-guided global--local shrinkage (PGL), a Bayesian regression method for identifying exposure-associated differences in the relative abundances of individual amplicon sequence variants (ASVs). A multinomial logit-normal working model analyzes relative abundance among ASVs detected in each sample. A zero-sum constraint gives the regression coefficients a log-ratio interpretation relative to the mean across retained ASVs. Global--local shrinkage accommodates weak and large associations, while an estimated weight controls information sharing along a supplied phylogenetic tree. Simulation ablation studies showed that phylogenetic borrowing improved detection power and estimation of nonzero coefficients when associations aligned with the tree, with little change under a tree-independent generator. Estimated borrowing decreased as the supplied tree was corrupted. Comparisons with benchmark methods revealed different trade-offs between detection power and false discovery rates across simulation settings. In an HIV oral microbiome application, ablation analysis showed that phylogenetic borrowing preserved the overall pattern of estimated associations while changing posterior uncertainty and selecting three additional ASVs. PGL provides ASV-level inference with adaptive information sharing across evolutionarily related features.
ER -