Generative microbiome language model writing disease-conditioned community profiles and predicting post-transplant composition for FMT donors.
No providers recorded yet. Browse all providers
Microbiome analysis is dominated by the abundance table: a matrix of taxa against samples, fed to a classifier. That framing supports discrimination well enough, but it has no natural way to express generation — producing a plausible community that does not exist yet — or intervention, predicting what a community becomes after something is done to it. Fecal microbiota transplantation, for instance, succeeds or fails largely on donor choice, and no assay says in advance which donor suits which recipient.
MGM 2.0 takes the language analogy seriously. A sample is a sentence and a genus is a word, so the tasks natural language processing already solves become available: prompt-conditioned generation for synthesising communities, and sequence-to-sequence question answering for predicting a transplant's outcome. It was developed in Kang Ning's lab at Huazhong University of Science and Technology and builds on that group's MGM, reusing its rank-based tokenisation and transformer backbone and extending them from classification into generative territory. It is distinct from MGM2, a later multimodal microbiome foundation model from the same lab that grounds each taxon token in sequence embeddings.
The encoding follows MGM: relative genus abundances are normalised against MicroCorpus-260K statistics, genera are ranked within each sample, and each rank becomes a discrete token, so the sentence records community structure while shedding the absolute abundances most distorted by sequencing protocol. The network is eight transformer blocks, each with eight attention heads, an embedding size of 256, a feed-forward width of 1,024, and GELU activations, implemented in PyTorch on the HuggingFace Transformers stack. Downstream classification swaps the language modelling head for a sequence classification head reading the end-of-sentence token.
On an exogenous colonisation dataset of 24 antibiotic-treated individuals challenged with Enterococcus faecium, the results section reports a five-fold cross-validated classification accuracy of 0.86 against 0.83 for a random forest on the raw abundance table, and a much larger gap on regression of post-intervention E. faecium abundance, where R² was 0.22 against 0.03. (The abstract labels 0.86 an AUROC; the results section gives it as accuracy.) For generation, the model was fine-tuned on 6,004 GMrepo gut samples spanning 17 diseases, split evenly for training and testing. Generated samples had lower perplexity than real training, test, or random samples; a disease classifier scored ROC-AUC 0.98 on real test data and 0.95 on generated data; and UMAP embeddings clustered by disease (ARI 0.085) far more strongly than by whether a sample was real (ARI 5.4 × 10⁻⁴). Under the paper's "Microbiome Turing Test", MGM 2.0 beat the GAN-based generator MB-GAN on cosine similarity to real composition (0.52 versus 0.046), taxon-taxon Spearman structure (0.51 versus 0.0018), downstream ROC-AUC (0.95 versus 0.49), and overlap of the top 500 biomarkers (440 versus 287). The FMT model was fine-tuned on 228 transplantation groups and validated on 15 inflammatory bowel disease pairs and 38 Clostridioides difficile infection pairs, with ROUGE-1 above 0.6 throughout. Scoring in-silico donor reassignments with a community-to-recipient metric flagged one donor whose average gain was 0.52 and another whose average was −2.20; SHAP attribution associated Desulfovibrio and Bifidobacterium with successful engraftment, Bilophila and Holdemania with failure.
The generative mode lets microbiome researchers augment small clinical cohorts with disease-conditioned synthetic samples and run in-silico experiments without new sequencing. The transplantation mode is aimed at FMT programmes, where predicting a recipient's post-transplant community would let a stool bank match donors rather than allocate them arbitrarily, and where the SHAP analysis suggests composition-based screening criteria. Colonisation prediction serves probiotic and engineered-consortium work, where the question is whether an introduced strain will establish at all.
MGM 2.0's contribution is showing that the sentence representation underlying MGM supports more than classification, and that prompt conditioning removes the per-condition retraining GAN-based microbiome generators require. The transplantation framing reuses an idea from machine translation for a clinical problem that has resisted predictive modelling. The work remains a preprint under a non-commercial, no-derivatives licence, and no code or checkpoint has been released for this version. The evaluations are also small: the colonisation cohort has 24 individuals and the external FMT validation totals 53 pairs.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.