Skip to content

Galaxy-ML2

A curated machine-learning tool suite within Galaxy — Galaxy-ML's pitch is reproducible, auditable, web-driven supervised ML for biomedical scientists. Wraps scikit-learn-compatible pipelines (preprocessing, feature selection, model training, hyperparameter search, evaluation) as Galaxy tools so end-to-end ML analyses ship as Galaxy workflows rather than ad-hoc notebooks.

Talk at GBCC2025

  • Speaker: Paulo Cilas Morais Lyra Junior (lead) with Junhao Qiu
  • Day / session: Day 2 (Tue Jun 24, 2025) — Oral Session 2A, Bush Hall
  • Talk title: "The Galaxy-ML2 tool suite: Using Galaxy to promote best practices in machine learning for biomedical data science"
  • Slides: TBD
  • Video: GBCC2025 playlist
  • Abstract / schedule: GBCC2025 program

What it does

Galaxy-ML2 is the next generation of Galaxy-ML: a unified scikit-learn-compatible pipeline framework exposed as Galaxy tools, with companion Galaxy-Ludwig deep-learning tools sitting alongside. Users assemble preprocessor + estimator + search-cv pipelines in the Galaxy UI, run them at scale on shared infrastructure, and ship the resulting workflow as a single shareable artefact. Validation is published in PLOS Comp Bio (Gu et al. 2021) and the suite is actively maintained.

How it works

Core idea. Galaxy-ML2 wraps scikit-learn-compatible pipeline components — preprocessors, feature selectors, estimators, hyperparameter searches — as Galaxy XML tools, so a Pipeline([preprocessor, estimator]) + GridSearchCV setup is composed in the Galaxy UI instead of a Python script, and the resulting workflow is shareable as a single Galaxy artefact.

Inputs / outputs. Input is a tabular dataset (features + label) plus a chosen pipeline configuration. Output is a trained model persisted to HDF5 (faster than pickle for array-heavy models), a hyperparameter-search report, cross-validation metrics, and the Galaxy workflow file describing exactly how the result was produced.

Key innovation. Reproducibility + auditability as a first-class property: because every step is a versioned Galaxy tool with logged parameters, the entire ML pipeline — preprocessing, hyperparameter grid, CV folds, evaluation metric — is recoverable from the Galaxy history alone, which matters when biomedical ML has to ship to regulated contexts (clinical decision support, FDA submissions).

Parameters / API surface worth knowing. - Estimators: scikit-learn (ensemble, linear, SVM, trees), XGBoost, MLxtend stacking, Keras (KerasGClassifier/KerasGRegressor), IRAPSClassifier. - Preprocessors: scikit-learn preprocessing/feature-selection/decomposition, imbalanced-learn samplers, skrebate relief algorithms, plus domain-specific encoders for genomic sequences, protein sequences, and image batches. - Hyperparameter search: GridSearchCV with parallelized fold evaluation. - Model persistence: HDF5 format for array-heavy models (preferred over pickle).

Canonical example. From the Galaxy-ML docs: in the Galaxy UI, chain a StandardScaler preprocessor → SelectKBest feature selector → RandomForestClassifier estimator, wrap the pipeline in SearchCV over a hyperparameter grid, point it at a labeled tabular dataset, and emit a fitted HDF5 model plus a CV report — the entire run is captured as a Galaxy workflow that another user can re-execute against their own data.

Where it fits in the corpus

  • AACR 2026: axes = bioinfo-tools and (loosely) agentic-AI — this is classical ML wrapped for reproducibility, not LM-driven, so it complements rather than overlaps GAIA / GalaxyMCP
  • GAIA (entry) and GalaxyMCP (entry) — same Goecks-lab umbrella; ML2 supplies the underlying tools an agent might invoke
  • Nextflow Summit 2026: thematic overlap with reproducible-ML pipeline storylines

Notes

Galaxy-ML's enduring pitch is reproducible/auditable ML pipelines — relevant when biomedical ML needs to ship in regulated contexts (clinical decision support, FDA-relevant work). The "v2" framing at GBCC suggests a re-architecture pass; verify scope from talk slides.