Citation

Cite this work

If you use the embeddings, trait tables or model, please cite the paper below. If you use results from this website, please also report the version information listed below.

Paper

Ecological embeddings reveal microbial social niches that generalize across diseases.

Manuscript in submission. The DOI and author list will be added upon publication. Until then, please cite the code repository and the version information below.

BibTeX

@article{microbial_social_niches,
  title   = {Ecological embeddings reveal microbial social niches that
             generalize across diseases},
  note    = {Manuscript in submission},
  url     = {https://github.com/xu-research-lab/microbial-embeddings}
}

Data and code availability

ResourceWhere
Analysis code xu-research-lab/microbial-embeddings
Website and export scripts xu-research-lab/sne-website
Problems with this site, or suggestions for it Open an issue
Data files used on this site /download

Version

Model—
Reference cohort—
Leave-one-disease-out AUC—
Vocabulary—
OTUs with a trained embedding—
Maximum input length (OTUs per sample)—

Fold checkpoints, per-disease AUCs and the validation AUC used to select each fold are listed in metrics.json.

Notes on reporting results

Trait values in the atlas are predictions, not measurements, and each is reported with its cross-validated AUC. The AUC is estimated by leave-one-phylum-out cross-validation, in which the model is evaluated on a phylum excluded from training. This is stricter than a random split and is appropriate here, because most predicted OTUs are uncultured and lie outside the training distribution.

The dysbiosis score is a percentile within a reference cohort, not a probability. On a disease outside the 13 in the cohort the AUC is 0.64, against 0.67 for the pooled leave-one-disease-out model reported in the paper on the same cohort; both are held-out estimates. When reporting generalization to new diseases, please use the leave-one-disease-out AUC (0.64). The AUC of the deployed ensemble measured on the cohort it was trained on is not a generalization estimate and should not be quoted as one.