Adaptive graph learning of microbial phylogeny enables accurate and interpretable microbiome-based host phenotype prediction.

The human microbiome is inherently structured by phylogeny, yet most predictive models treat microbial taxa as independent features, thereby underusing evolutionary information that may improve disease classification. While recent deep learning approaches have attempted to incorporate phylogeny, they generally rely on projecting phylogenetic trees into Euclidean spaces, which can distort the intrinsic topology of evolutionary relationships. To address this limitation, we propose PhyloGCNE, a framework that models microbiome samples directly as graphs and employs edge-aware graph convolution to integrate phylogeny. Unlike previous methods that rely on fixed, distance-based aggregation, PhyloGCNE learns how phylogeny-informed edge attributes should influence signal propagation across evolutionary hierarchies. We further introduce a Phylogenetic Saliency Propagation (PSP) framework for model interpretation, which attributes importance scores to microbial taxa by integrating gradient sensitivity with evolutionary context. Benchmarked against one synthetic and eight real-world data sets spanning inflammatory bowel disease, colorectal cancer, type 2 diabetes, oral squamous cell carcinoma, gastric cancer, and dietary fiber intervention, PhyloGCNE consistently outperforms existing state-of-the-art approaches. Together, these results establish PhyloGCNE as an accurate and interpretable phylogeny-aware framework for microbiome-based host phenotype prediction.IMPORTANCEThe human microbiome is a complex ecosystem closely linked to physiological health, yet traditional analysis often treats microbes as isolated features, ignoring their shared evolutionary history. This study introduces PhyloGCNE, a novel framework that integrates the evolutionary tree directly into the analysis of microbiome data. By modeling microbial communities as interconnected networks rather than independent entities, this approach captures shared biological traits across related lineages. We demonstrate that this method significantly improves the accuracy of predicting host phenotypes, such as inflammatory bowel disease and colorectal cancer. Crucially, unlike many "black box" artificial intelligence models, this tool identifies specific, biologically relevant microbial signatures driving these predictions. This advancement provides a powerful, interpretable approach for deciphering the complex links between the human microbiome and host phenotypes.