README.md

MoleculeNet Examples

The repository includes the following example scripts:

• fingerprint
• gnn
• acnn

Fingerprint

To train a random forest on 1024-bit ECFP2 fingerprint

python fingerprint.py

The feasible arguments include:

• Dataset: -d dataset
  • Specifies the dataset to use, which can be one of the following:
    • BACE_classification
    • BACE_regression
    • BBBP
    • ClinTox
    • HOPV
    • SIDER
    • Lipo
• Hyperparameter Search (optional): -hs
  • Perform a hyperparameter search using Bayesian optimization. It determines the best hyperparameters based on the validation metric averaged across 3 runs.
  • If not specified, the script uses default hyperparameters.
• Number of Hyperparameter Search Trials (optional): -nt num_trials [default=16]
  • The number of trials for hyperparameter search. This comes into effect only when the user specifies -hs as described above.

GNN

To train a GNN using GraphConv featurization

python gnn.py

The feasible arguments include:

• Dataset: -d dataset
  • Specifies the dataset to use, which can be one of the following:
    • BACE_classification
    • BACE_regression
    • BBBP
    • ClinTox
    • HOPV
    • SIDER
    • Lipo
• Hyperparameter Search (optional): -hs
  • Perform a hyperparameter search using Bayesian optimization. It determines the best hyperparameters based on the validation metric averaged across 3 runs.
  • If not specified, the script uses default hyperparameters.
• Number of Hyperparameter Search Trials (optional): -nt num_trials [default=16]
  • The number of trials for hyperparameter search. This comes into effect only when the user specifies -hs as described above.

ACNN

To train an ACNN using AtomicConv featurization

python acnn.py

The feasible arguments include:

• Dataset: -d dataset
  • Specifies the dataset to use, which can be one of the following:
    • PDBbind
• Hyperparameter Search (optional): -hs
  • Perform a hyperparameter search using Bayesian optimization. It determines the best hyperparameters based on the validation metric averaged across 3 runs.
  • If not specified, the script uses default hyperparameters.
• Number of Hyperparameter Search Trials (optional): -nt num_trials [default=16]
  • The number of trials for hyperparameter search. This comes into effect only when the user specifies -hs as described above.

Results

The scripts will save the training results in results/, including:

• configure.json: the file that stores the training configuration
• eval.txt: the file that stores the final validation and test performance

If the scripts perform a hyperparameter search, there will be subfolders under results/ corresponding to the search trials. One subfolder per trial. In this case, results/configure.json and results/eval.txt contain the results for the best trial.