Skip to content

Advanced Models

OpenADMET Challenge

ChemProp was used by many of the top performers on the OpenADMET Leaderboard. Workbench makes it easy to train and deploy these models to AWS®.

Workbench supports advanced model frameworks beyond the default XGBoost. All frameworks use the same API — just change the model_framework parameter.

Available Model Frameworks

Model Framework Description Details
XGBoost Gradient boosted trees on RDKit molecular descriptors Default framework
PyTorch Neural network on RDKit molecular descriptors Good for nonlinear descriptor interactions
ChemProp Message Passing Neural Network (MPNN) on molecular graphs Learns directly from molecular topology
ChemProp Hybrid MPNN combined with top RDKit descriptors Best of both worlds
ChemProp Multi-Task Single MPNN predicting multiple endpoints Related endpoints with shared chemistry
ChemProp Foundation Pretrained MPNN weights (CheMeleon) Transfer learning for small datasets
Meta Model Ensemble aggregating multiple endpoints Combines frameworks for lower error

Quick Example

from workbench.api import FeatureSet, ModelType, ModelFramework

fs = FeatureSet("admet_features")

# ChemProp model
chemprop_model = fs.to_model(
    name="logd-chemprop",
    model_type=ModelType.UQ_REGRESSOR,
    model_framework=ModelFramework.CHEMPROP,
    target_column="logd",
    feature_list=["smiles"],
)

# PyTorch model (same API, different framework)
pytorch_model = fs.to_model(
    name="logd-pytorch",
    model_type=ModelType.UQ_REGRESSOR,
    model_framework=ModelFramework.PYTORCH,
    target_column="logd",
    feature_list=fs.feature_columns,
)

# Deploy either one
endpoint = chemprop_model.to_endpoint()

For detailed documentation on ChemProp model types (ST, MT, Hybrid, Foundation), see ChemProp Models. For ensemble models that combine multiple endpoints, see Meta Models.

Beta Software

Workbench is currently in beta. We're actively looking for beta testers! If you're interested in early access, contact us at workbench@supercowpowers.com.

Questions?

The SuperCowPowers team is happy to answer any questions you may have about AWS® and Workbench.

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