Overview
mllabiome is a Python library for building, comparing, combining and explaining machine-learning models from microbiome abundance data. It supports phenotype prediction, model benchmarking, ensemble learning, multimodal integration, cross-cohort validation, explainable AI, robustness analysis, and reproducible reporting.
Use it when the scientific question is not only which model predicts well?, but also does the full microbiome modelling workflow generalize to observations that were not used to construct it?
Workflow
explore → evaluate → ensemble → inference → explain → robustness → reportEach stage can be run independently after the required upstream artefacts exist. The complete pipeline is launched with:
mllabiome path/to/config.pyCore terms
| Term | Meaning |
|---|---|
| MPDR | Microbiota profile data representation: a microbiome representation defined by taxonomic resolution and abundance transformation |
| MPMA | Microbiota profile modelling algorithm: an MPDR paired with a learner |
| MPMA-B | Base MPMA: the selected individual modelling algorithm |
| MPMA-E | Ensemble MPMA: a selected ensemble of eligible base algorithms |
| OOF | Out-of-fold: predictions or explanations generated for observations not used to fit the corresponding model |
| LODO | Leave-one-dataset-out: each dataset or cohort is held out in turn for evaluation |
What the package controls
- fold-local preprocessing and feature filtering;
- grouped splitting for repeated subjects;
- nested model and MPMA selection;
- LODO / cross-cohort evaluation;
- classification and regression;
- ensemble selection from inner out-of-fold evidence;
- multimodal feature and prediction integration;
- held-out explainability with SHAP, permutation importance, ALE, LIME, and ALE interactions;
- robustness, statistical summaries, figures, tables, and HTML reporting.
v0.2.0 highlights
- Frozen foundation-model embeddings (Waypoint, MGM, MGM2, and optional MBEmbed or local Qwen3).
- Expanded ensemble selection, benchmarking, and explainability reporting.
- Optional federated-learning simulation and local interactive inference interface.
- Improved pipeline inference and exploratory differential-abundance analysis.
See Foundation models, Federated simulation, and Inference.
What it does not do
mllabiome is not a raw sequencing pipeline. Read QC, assembly, taxonomic profiling, and generation of abundance or feature tables happen upstream.
It is also not a clinical decision-support system. Predictions and feature attributions are research outputs and require independent validation before any translational interpretation.
Start
Continue with the Quickstart, then read Core concepts and Evaluation and comparison.