Here, using fecal metagenomics data of 2,320 individuals, the authors develop a microbiome-based machine learning approach showing high accuracy for multi-class disease diagnosis, highlighting its potential application in improving noninvasive diagnostics and monitor responses to therapy.