Foundation Models for Microbiome Research: From Sequence Semantics to Community Dynamics and Multimodal World Models

Microbiome sequencing has advanced faster than microbiome understanding. Although large-scale 16S, metagenomic, metatranscriptomic, and proteomic datasets have accumulated rapidly, most analyses remain cohort-specific and association-driven, limiting mechanistic insight, cross-study transferability, and robustness to technical confounding. Foundation models offer a new computational framework by learning reusable biological representations from large unlabeled datasets. In this Review, we present microbiome foundation models as a hierarchy spanning biological scales. Sequence-centric models capture the syntax and semantics of DNA and proteins for taxonomic inference, functional annotation, and generative design. Community-centric models learn ecological structure from abundance profiles, while addressing compositionality, sparsity, and the unordered nature of microbial communities. Emerging multimodal frameworks integrate sequence-derived functional potential with community-level ecological dynamics under host and environmental context. We discuss key design choices, including tokenization, representation granularity, self-supervised objectives, and evaluation strategies, and highlight challenges in interpretability, domain shift, causal reasoning, and biological validation. Finally, we propose a transition from static representation learning toward intervention-aware microbiome world models capable of simulation, digital twinning, and generative microbiome engineering.