MGM2

MGM2 is a multimodal microbiome foundation model that learns transferable representations of whole microbial communities. Most computational microbiome work still reduces a community to a cohort-specific abundance table, where each taxon is an anonymous feature column and nothing about its underlying sequence is available to the model. Representations built that way rarely survive the move to a new study, so each new question gets a new model.

The model was developed in Kang Ning's lab at Huazhong University of Science and Technology, the same group behind the earlier MGM, a microbiome community foundation model pretrained on roughly 260,000 samples. MGM2 scales the corpus by nearly an order of magnitude and adds a sequence modality: rather than treating each operational taxonomic unit (OTU) as an arbitrary vocabulary index, it grounds each token in sequence embeddings derived from NTv3, then conditions on measured abundance and aligns the result with community-level semantics. That makes it a sequence-aware counterpart to community-composition models such as Waypoint and to abundance-only embeddings like Microbiome Self-Supervised Learning, and a community-level complement to genome models such as GenomeOcean.

The central claim is reuse rather than task-specific engineering. A single frozen MGM2 backbone supplies the representation for biome classification, paired-community prediction, and longitudinal forecasting, with no retraining of the backbone between task families.

#Key Features

• Sequence-aware community tokens: Each of 225,067 OTUs, clustered at 99% sequence similarity, is represented by a sequence-derived embedding coupled with abundance conditioning, so taxon identity carries sequence information rather than an index.
• Large-scale pretraining: Trained on 1,821,291 MicrobeAtlas samples, spanning host-associated and environmental communities.
• Frozen-representation transfer: Frozen features outperformed DeepPhylo by 0.06–0.21 macro-AUROC across five temporally held-out MGnify biome hierarchy levels, with the largest gains on rare and fine-grained labels.
• Paired-community prediction: On fecal microbiota transplantation, MGM2-XLarge reached a response ROC AUC of 0.79 and reduced post-transplant Bray-Curtis distance by 15% relative to the recipient baseline.
• Longitudinal forecasting: The same representation supported ASV-level trend forecasting across 24 wastewater treatment plants.
• Interpretable feature dictionary: Sparse autoencoder analysis resolved token states into a 4,096-feature dictionary spanning taxonomic identity, abundance state, ecological context, and technical variation.

#Technical Details

MGM2 produces both sample-level and token-level representations over a vocabulary of 225,067 OTUs, pretrained on 1,821,291 MicrobeAtlas samples. The multimodal design combines three ingredients: NTv3-derived embeddings of the microbial sequences behind each token, conditioning on the abundance at which that token was observed, and an alignment objective that ties token states to community-level semantics. The preprint reports several model scales; MGM2-XLarge is the configuration used for the transplantation and interpretability analyses.

Evaluation is organized around frozen features and temporal holdouts. Across five levels of the MGnify biome hierarchy, linear readouts on frozen MGM2 representations improved macro-AUROC over DeepPhylo by 0.06–0.21. In the transplantation setting, MGM2-XLarge predicted donor–recipient response at 0.79 ROC AUC and cut post-transplant Bray-Curtis distance by 15% against the recipient baseline. Interpretability comes from a sparse autoencoder trained on MGM2-XLarge token states, which decomposes them into 4,096 features that separate biological signal from technical variation.

#Applications

The model targets researchers who work with 16S/amplicon and community-profile data and want representations that carry across cohorts. Uses demonstrated in the preprint span biome annotation over the MGnify hierarchy, predicting the outcome of a paired community intervention such as fecal microbiota transplantation, forecasting taxon-level trends in engineered systems such as wastewater treatment plants, and feature discovery through the sparse-autoencoder dictionary. Because the backbone stays frozen, small labeled cohorts can be analyzed by fitting only a lightweight head.

#Impact

MGM2 pushes microbiome foundation models past the abundance table by giving each community token a sequence-derived identity, and it makes the case with a single frozen backbone serving classification, intervention-response prediction, and forecasting. Its interpretability analysis is unusual for this domain and offers a route to auditing whether a microbiome representation has learned ecology or batch structure. The work is a preprint awaiting peer review, and its benchmark comparisons are reported against DeepPhylo rather than a broad panel of community foundation models. Code and pretrained checkpoints for all four model sizes are published, but the repository and the weight release both omit a license, which leaves reuse terms unsettled; the pretraining corpus and evaluation splits are not distributed, and the preprint carries a CC-BY-NC license.