Decoding the Microbiome-Disease Axis with Interpretable Graph Neural Networks

Abstract

The human gut microbiome is a complex ecosystem whose disruption is implicated in a wide spectrum of diseases, yet translating microbiome research into actionable therapeutics is hindered by a critical trade-off: existing models either prioritize predictive accuracy at the expense of interpretability or sacrifice performance for mechanistic insight, limiting their ability to pinpoint specific disease-driving microbial interactions and taxa. To address this, we introduce Graph neural network for Interpretable Microbiome (GIM), a graph neural network framework that integrates minimally processed taxonomic metadata as sparse node embeddings within an unweighted complete graph, enabling direct modeling of high-order microbial interactions through message passing. GIM achieves state-of-the-art classification performance on microbiome-disease prediction tasks (e.g., healthy vs. allergic states) while generating fine-grained, experimentally validated attributions at the level of taxonomic ranks, driver microbes, and putative microbe-to-microbe interactions. By bridging the gap between predictive accuracy and biological interpretability, GIM overcomes a key limitation in current approaches, offering a unified framework to both predict dysbiosis-associated disease states and identify actionable microbial targets for therapeutic intervention. This dual capability represents a critical advance toward precision microbiome engineering and scalable hypothesis generation in translational microbiome research.