🤖 AI Summary
This study addresses the underutilization of phylogenetic information in 16S rRNA sequencing data by modeling evolutionary relationships among bacterial taxa as sequence string similarities. It introduces, for the first time in microbiome analysis, natural language processing–inspired string kernels—specifically spectrum and mismatch kernels—to explicitly encode multi-scale phylogenetic signals while preserving biological interpretability and enhancing statistical power. Based on this formulation, we develop StringPhylo: a kernel two-sample test and Gaussian process regression framework. In simulation studies, StringPhylo significantly improves statistical power and scale sensitivity compared to conventional methods. In real-world host phenotype prediction tasks, it outperforms UniFrac-based approaches combined with standard machine learning models. The core contribution lies in bridging phylogenetics and string kernel methodologies for microbiome statistical inference.
📝 Abstract
Bacterial community composition is measured using 16S rRNA (ribosomal ribonucleic acid) gene sequencing, for which one of the defining characteristics is the phylogenetic relationships that exist between variables. Here, we demonstrate the utility of modelling these relationships in two statistical tasks (the two sample test and host trait prediction) by employing string kernels originally proposed in natural language processing. We show via simulation studies that a kernel two-sample test using the proposed kernels, which explicitly model phylogenetic relationships, is powerful while also being sensitive to the phylogenetic scale of the difference between the two populations. We also demonstrate how the proposed kernels can be used with Gaussian processes to improve predictive performance in host trait prediction. Our method is implemented in the Python package StringPhylo (available at github.com/jonathanishhorowicz/stringphylo).
