MGM2 as a Multimodal Foundation Model for Microbial Community States and Responses

Microbial communities sustain ecosystems and influence host health, yet connecting their molecular diversity to responses under intervention remains difficult. Here we introduce MGM2, a multimodal foundation model integrating DNA language-model representations of microbial sequences, quantitative abundances and sample descriptions. After pretraining on 1.8 million microbiomes, its frozen representations achieved a mean macro-AUROC of 0.91 on later MGnify projects absent from pretraining. These representations also supported fecal microbiota transplantation outcome prediction, recovering 68% of observed high-change taxa. Adding molecular structures enabled prediction of drug effects in nested five-fold evaluation, while measured perturbed states informed subsequent recovery. To interpret the learned representations, sparse decomposition resolved taxonomic, ecological and abundance-related features. Together, these results show how a shared representation of molecular identity and community context connects large-scale microbiome learning to the prediction and interpretation of community change.