Foundation-model representations

mllabiome can benchmark frozen pretrained embeddings alongside conventional microbiome transformations. The package does not distribute model weights.

Adapter	Default resolution	Requirements
`WaypointEmbedding`	Genus	Foundation dependencies; access to the selected Hugging Face checkpoint
`MGMEmbedding`	Genus only	Foundation dependencies and upstream model resources
`MGM2Embedding`	Species	Foundation dependencies and upstream checkpoints
`MBEmbed`	Species only	Compatible, portable author-provided checkpoint
`QwenOllamaEmbedding`	Genus	Local Ollama with a Qwen3 embedding model

Configuration

From the repository checkout, install optional dependencies with uv sync --extra foundation. Waypoint may also require Hugging Face authentication; download and usage conditions are set by the model providers.


from mllabiome import mll
 
RESOLUTIONS = (
    ("genus", ("genus",)),
    ("species", ("species",)),
)
 
COUNT_TRANSFORMATIONS = (
    mll.Transformation("relative_abundance"),
    mll.WaypointEmbedding(),
    mll.MGMEmbedding(),
    mll.MGM2Embedding(),
)

Each adapter runs only at its configured resolution. Activate only the adapters for which the required checkpoints and dependencies are available. Representations are evaluated through the same train/test splitting and model-selection workflow as other transformations.

For local Qwen embeddings, install Ollama, pull qwen3-embedding:0.6b, and add mll.QwenOllamaEmbedding() to COUNT_TRANSFORMATIONS. Its default endpoint is 127.0.0.1:11434.

MBEmbed limitation: the publicly available standard checkpoint is not portable in a clean Python environment; use a compatible author-provided checkpoint if available. These are representation adapters, not clinical prediction models.

The optional OllamaDecisionClassifier is a separate experimental local prediction adapter, not a frozen embedding transformation.