Towards a Generative Paradigm for Large-scale Microbiome Analysis by Generative Language Model

Abstract Microbiome analysis has traditionally relied on taxonomic abundance tables, which, while effective, often constrain the exploration of deeper contextual relationships. In this study, we present MGM 2.0, a novel framework that applies advanced natural language processing (NLP) techniques to microbiome research. By reimagining microbiome samples as sentences and microbial species as words, MGM 2.0 enabled the extraction of nuanced patterns and relationships. The model demonstrated robust predictive performance in identifying exogenous species colonization (AUROC = 0.86). Additionally, through prompt-guided microbiome data generation, MGM 2.0 produced realistic microbial profiles conditioned on disease labels. The framework further revolutionized donor selection in fecal microbiota transplantation (FMT) by framing it as a sequence-to-sequence prediction task, enabling the prediction of post-transplantation community compositions and the identification of super donors for personalized treatments (average increase in C2R = 0.52). This innovative integration of NLP and microbiome science provides a versatile toolkit for predictive modeling, data generation, and personalized medicine. Highlights Introduced MGM 2.0, a generative language model utilizing sentence-like representation for microbiome analysis and generation. Demonstrated that sentence-like representation preserves sample distinctions, enabling accurate microbial sample classification tasks, such as colonization prediction. Generated realistic, disease-specific microbiome profiles using a prompt-guided approach, validated by a novel “Microbiome Turing Test.” Applied MGM 2.0 to fecal microbiota transplantation (FMT) donor selection, accurately predicting post-transplant community compositions and identifying potential “super donors” for personalized treatment strategies.