Workflow and outputs

Stages


configured MPMA sweep → ensemble sweep → explainability → report

Stage	Command	Purpose
Evaluate	`mllabiome config.py --stage evaluate`	Materialise MPDRs and evaluate every MPMA candidate
Ensemble	`mllabiome config.py --stage ensemble`	Search candidate MPMA ensembles and select MPMA-E
Explain	`mllabiome config.py --stage explain`	Run out-of-fold SHAP, LIME, permutation, ALE, and interactions for configured targets
Report	`mllabiome config.py --stage report`	Write the HTML report and manuscript-ready tables/figures
All	`mllabiome config.py`	Run the full pipeline

Main output locations

Each experiment writes under EXPERIMENT_DIR.


EXPERIMENT_DIR/
  configs.tsv
  configs.db
  tables/
    mpma_rankings.tsv
    qualification_gate.tsv
    representation_impact_cells.tsv
  figures/
    mpma_top_metric.png
    representation_impact.png
    representation_impact.pdf
    representation_impact.svg
  ensembling/
    final_model_comparison.tsv
    selected_unit.json
    ensemble_candidates.png
  explainability/
    <target>/
      feature_importance.tsv
      feature_importance_shap.tsv
      feature_importance_lime.tsv
      feature_importance_permutation.tsv
      feature_importance_ale.tsv
      oof_predictions.tsv
      explained_unit.json
      figures/
  report.html

Resume behavior

The stages are restartable. A configuration can be run repeatedly while the sweep space is being expanded. Completed MPMA/split pairs are retained, and newly enabled resolutions, transformations, or learners are evaluated on rerun.

If evaluation is complete but the report needs a refresh, rerun only:


mllabiome config.py --stage report

If the selected ensemble exists and only explanations need to be regenerated:


mllabiome config.py --stage explain