Chemprop: Benchmark Data for Molecular Property Prediction

Benchmark Data for Chemprop is a collection of datasets and pre-defined splits from the Chemprop machine learning package for chemical property prediction. It includes data for multiple benchmarking systems, such as HIV replication inhibition, biological activities from PCBA, DFT-calculated quantum properties from QM9, and reaction barrier heights. The collection provides train, validation, and test splits to facilitate the development and evaluation of predictive models in computational chemistry.

Use Cases

• Benchmarking machine learning models for HIV replication inhibition based on scaffold splits.
• Training multi-task models for DFT-calculated quantum mechanical properties from the QM9 dataset.
• Predicting water-octanol partition coefficients (logP) for molecules from the SAMPL6, 7, and 9 challenges.
• Evaluating model performance on reaction barrier heights for E2, SN2, and cycloaddition reactions.
• Comparing model architectures using biological activity data from MoleculeNet and OGB with random or scaffold splits.

Strengths

• Provides pre-defined train, validation, and test splits for multiple distinct chemical prediction tasks.
• Includes data from established, peer-reviewed sources such as MoleculeNet, OGB, and the SAMPL challenges.
• Covers a wide range of chemical prediction tasks, from biological activity to quantum-mechanical properties.

Limitations

• Key metadata such as row counts, column names, and specific file formats are not provided in the source descriptions.
• The exact date of the last update is unknown, which may affect confidence in the data's currency.
• The descriptions from multiple sources are highly repetitive and lack detailed documentation on data collection and preprocessing.

Provenance

Source

Massachusetts Institute of Technology (author: Esther Heid), derived from sources including MoleculeNet, OGB, and SAMPL challenges.

Collection Method

Data was compiled and splits were created for the Chemprop machine learning package manuscript.