ADMET Workbench

Getting Started

pip install workbench
workbench

Workbench Version: 0.8.475
Bosco: Claude via Anthropic API key

Welcome to Workbench!

🐶  New to Workbench? Ask me to walk you through building your first model.

Workbench:Bosco>

No AWS account needed to start. Local mode gives you the whole chain — DataSource → FeatureSet → Model → Endpoint — on your filesystem, training with the same generated script SageMaker runs. pub_data reads public ADMET datasets straight from S3, so there's real data to model from the first prompt:

Workbench:Bosco> build me a solubility model from the aqsol public data

Ask Bosco in plain language, or type Python — both land in the same session, and the variables persist for both of you. See Local Mode.

Giving Bosco a model

Bosco is the resident ML agent. He needs a path to Claude, and status always names the one he took:

• With an AWS account — he runs on Bedrock inside it, with Zero Data Retention. Nothing to install and nothing leaves your account.
• Without one — set ANTHROPIC_API_KEY and restart. Bring your own key, or email workbench@supercowpowers.com and we'll issue you a trial key.

Connecting your AWS account

• Initial Setup
• One-time AWS Onboarding
• Using Bosco  ·  Security & Admin

📝 Bosco: Workbench ML Agent — a real eleven-turn session, start to finish.

Chemprop Models

All the rage for the Open ADMET Challenge. ADMET Workbench supports:

• Single Task Chemprop Models
• Multi Task Chemprop Models
• Chemprop + Descriptors Models (MPNN + Molecular Descriptors)
• Foundation Chemprop Models (CheMeleon Pretrained)

Examples: Deploying Chemprop Models  ·  Deploying Foundation Chemprop Models

References

• Open ADMET Challenge
• ChemProp: Yang et al. "Analyzing Learned Molecular Representations for Property Prediction" J. Chem. Inf. Model. 2019 — GitHub | Paper
• CheMeleon Github

Chemprop Action Shots!

The Dashboard

Health monitoring, dynamic updates, and a high-level summary across the whole pipeline, with drill-down views for Incoming Data, Glue Jobs, DataSources, FeatureSets, Models, and Endpoints. Try it: Workbench Dashboard Demo.

Private SaaS Architecture

Secure your Data, Empower your ML Pipelines

Workbench deploys as an AWS Stack inside your own cloud environment (BYOC), so your data never leaves it — and plugin support lets you tailor it to your own workflows. See the Private SaaS Architecture deck.

Documentation

The ADMET Workbench framework makes AWS® both easier to use and more powerful: a simple Python API over Glue, Athena, Feature Store, Models, and Endpoints, with web interfaces on top of it.

Workbench Docs covers the Python API in depth, fully searchable, with code examples throughout. Runnable versions of every example live in examples/.

Workbench takes something genuinely complex — the full set of AWS ML services — and makes it less complex, so there's conceptual documentation too: Workbench Presentations.

AWS Marketplace

Workbench is on the AWS Marketplace as a Dashboard for ML Pipelines and can be billed through AWS.

Installation extras

pip install workbench               # API + REPL + orchestration (the default)
pip install 'workbench[ui]'         # + plotly, dash — the Workbench Dashboard
pip install 'workbench[modeling]'   # + torch, chemprop, ray[tune], optuna
pip install 'workbench[misc]'       # + umap-learn
pip install 'workbench[dev]'        # + pytest, coverage, flake8, black
pip install 'workbench[all]'        # everything above

Quotes are needed — shells read square brackets as globs.

Cleanlab/Datalab workflows aren't bundled in any extra. Install cleanlab[datalab]>=2.8.0 before calling cleanlab_model().

Questions?

The SuperCowPowers team is happy to answer any questions about AWS and Workbench — workbench@supercowpowers.com or Discord.

Beta Program — if your company would like to be a Workbench Beta Tester, get in touch at the same address.

Contributions are welcome, and fall under the existing project license.

® Amazon Web Services, AWS, the Powered by AWS logo, are trademarks of Amazon.com, Inc. or its affiliates