Chemprop: A Machine Learning Package for Chemical Property Prediction

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Drug Discovery Software

Background:

• Deep learning models, particularly directed message-passing neural networks (D-MPNNs), are increasingly vital for predicting molecular properties.
• Existing software solutions often lack the versatility and user-friendliness required for nonexpert operation.
• There is a growing demand for accessible, open-source tools to leverage machine learning in chemical research.

Purpose of the Study:

• To introduce significant enhancements to the Chemprop software package, expanding its capabilities beyond basic molecular property prediction.
• To provide users with advanced features for uncertainty quantification, transfer learning, and handling complex chemical data like reactions and spectra.
• To benchmark the performance of D-MPNN models trained with the updated Chemprop on various challenging chemical prediction tasks.

Main Methods:

• Implementation of new functionalities in Chemprop, including support for multi-molecule properties, reactions, atom/bond-level predictions, and spectral data.
• Integration of uncertainty quantification and calibration methods with relevant performance metrics.
• Development of pretraining and transfer learning workflows, alongside improved hyperparameter optimization and customization options for loss functions and molecular features.

Main Results:

• Chemprop demonstrates state-of-the-art performance on benchmark datasets (MoleculeNet, SAMPL) for predicting key molecular properties.
• Specific achievements include high accuracy in predicting water-octanol partition coefficients, reaction barrier heights, atomic partial charges, and absorption spectra.
• The enhanced software facilitates out-of-the-box training for D-MPNN models across a wide range of chemical prediction problems.

Conclusions:

• The updated Chemprop software provides a powerful, user-friendly, and open-source solution for advanced molecular property prediction using D-MPNNs.
• Its expanded capabilities, including reaction and spectra analysis, position it as a valuable tool for both expert and nonexpert researchers in computational chemistry.
• Chemprop's robust performance and versatility contribute to the broader adoption of machine learning in chemical discovery and development.