Skip to content

Getting Started

Need Help?

The SuperCowPowers team is happy to give any assistance needed when setting up AWS and ADMET Workbench. So please contact us at workbench@supercowpowers.com or on chat us up on Discord

Install and run

pip install workbench
workbench

workbench starts the REPL. With no AWS configuration it comes up in local mode: artifacts live on your filesystem, so there's no account to create, no credentials to manage, and nothing to pay for.

The base install carries XGBoost, RDKit, and the descriptor stack — everything on this page. Two extras add the rest:

Extra Adds
workbench[modeling] Chemprop and PyTorch models (pulls torch, a large download)
workbench[ui] The dashboard and its plugin pages

Build a model

pub_data reads public datasets from S3 anonymously, and the local classes are already bound at the prompt, so this runs as-is:

df = pub_data.get("comp_chem/aqsol/aqsol_public_data")

ds = DataSource(df, name="aqsol_local")
fs = ds.to_features("aqsol_local_features", id_column="ID")
model = fs.to_model(
    "aqsol-local-reg",
    model_type=ModelType.REGRESSOR,
    model_framework=ModelFramework.XGBOOST,
    target_column="solubility",
    feature_list=["molwt", "mollogp", "tpsa", "numrotatablebonds"],
)

model.get_inference_metrics()

Training runs the same model script SageMaker runs, so a model that works here publishes and produces the same model. See Local Mode for scoring, endpoints, and the rest of the local API.

Or just ask

Bosco, the Workbench ML agent, runs in local mode too. Give it an Anthropic API key and it writes and runs that chain for you:

export ANTHROPIC_API_KEY=sk-ant-...
workbench

Bosco appears in the prompt, and anything you type that isn't Python goes to it:

Workbench:Bosco> build me a solubility model from the aqsol public data
Workbench:Bosco> show me a predicted vs actual plot
Workbench:Bosco> what are the nearest neighbors of the worst outlier?

It works in your live session, so the variables it creates stay in your namespace. Prompts go to the Anthropic API rather than an AWS account in this mode — status always names where they go. See Security & Admin.

Connect your AWS account

Publishing a model, deployed endpoints, monitoring, and the dashboard all need an AWS account. ADMET Workbench uses your existing AWS account/profile/SSO — if you don't have one yet, start with AWS Setup.

Then run aws_setup() from the REPL:

aws_setup()

AWS_PROFILE: my_aws_profile
WORKBENCH_BUCKET: my-company-workbench
[optional] REDIS_HOST(localhost): my-redis.cache.amazon (or leave blank)
[optional] REDIS_PORT(6379):
[optional] REDIS_PASSWORD():
[optional] DASHBOARD_URL():
[optional] ENABLE_BOSCO -- run the Bosco ML agent? (y/N):

It writes the config, prints the one environment variable to export, and exits. Add that line, open a new terminal, and run workbench again — you only do this once.

Data Scientists/Engineers

AWS Administrators

For companies that are setting up ADMET Workbench on an internal AWS Account: Company AWS Setup

Additional Resources