A statistical design approach termed 'Constraint Distillation' that learns generative models from strain presence/absence patterns alone—without requiring biological priors about strain phenotypes, interactions, or mechanisms—enables engineering of a 15-member synthetic microbiome (SynCom15) that suppresses multi-drug resistant Klebsiella pneumoniae and other gut pathogens more effectively than fecal microbiota transplant across diverse environments through environmental acidification and fatty acid production, demonstrating that functional synthetic microbiomes can be created by solving the inverse problem of inferring compositional constraints from observational data rather than encoding forward models of biological mechanism.

Engineering functional microbiomes is challenging due to complex interactions between bacteria and their environments. Using a set of 848 gut commensal strains and clearance of multi-drug resistant Klebsiella pneumoniae (Kp-MH258) as a target function, we engineered a functional 15-member synthetic microbiome—SynCom15—through a statistical approach agnostic to strain phenotype, mechanism of action, bacterial interactions, or composition of natural microbiomes. Our approach involved designing, building, and testing 96 metagenomically diverse consortia, learning a generative model using community strain presence/absence as input, and distilling model constraints through statistical inference. SynCom15 cleared Kp-MH258 across in vitro, ex vivo, and in vivo environments, matching the efficacy of a fecal microbiome transplant in a clinically relevant murine model of infection. The mechanism of suppression by SynCom15 was related to fatty acid production coupled with environmental acidification. SynCom15 also suppressed other pathogens—Clostridioides difficile, Escherichia coli, and other K. pneumoniae strains—but through different mechanisms. Sensitivity analysis revealed models trained on strain presence/absence captured the statistical structure of pathogen suppression, illustrating that community representation was key to our approach succeeding. Our framework, ‘Constraint Distillation’, could be a general and efficient strategy for building emergent complex systems, offering a path towards synthetic ecology more broadly.