Chemprop

Chemprop implements directed message-passing neural networks (D-MPNNs) to predict chemical and molecular properties from structural representations for molecular property prediction and computational chemistry analyses.

Key Features:

• Directed message-passing neural networks (D-MPNNs): Implements D-MPNNs for graph-based deep learning on molecular structures.
• Supported data types: Handles multimolecule properties, reactions, atom-level properties, and spectra.
• Uncertainty quantification and calibration: Includes uncertainty quantification and calibration methods to assess prediction reliability.
• Pretraining and transfer learning: Supports pretraining and transfer learning workflows to leverage existing data and models.
• Hyperparameter optimization: Provides improved hyperparameter optimization techniques for model tuning.
• Customizable model components: Allows customization of loss functions and atom/bond feature sets.
• Benchmarking: Benchmarked on property prediction datasets including MoleculeNet and SAMPL.

Scientific Applications:

• Molecular property prediction: Predicts properties such as water-octanol partition coefficients.
• Reaction property prediction: Predicts reaction barrier heights.
• Atom-level property prediction: Predicts atomic partial charges.
• Spectral prediction: Predicts absorption spectra.
• Benchmarking and validation: Used for benchmarking on datasets including MoleculeNet and SAMPL.

Methodology:

Implements directed message-passing neural networks (D-MPNNs) and incorporates uncertainty quantification and calibration methods, pretraining and transfer learning workflows, hyperparameter optimization, and customization of loss functions and atom/bond features.