MoleculeACE

MoleculeACE benchmarks machine learning methods for their ability to detect and predict activity cliffs—pairs of structurally similar molecules with large differences in biological activity—to improve evaluation of predictive models in drug discovery.

Key Features:

• Benchmarking Platform: Benchmarks the performance of machine learning methods on curated bioactivity data with emphasis on activity-cliff cases.
• Comprehensive Evaluation: Evaluates 24 distinct machine and deep learning approaches across curated datasets from 30 macromolecular targets to assess ability to predict activity-cliff compounds.
• Performance Insights: Reports that methods based on molecular descriptors tend to outperform more complex deep learning models when predicting activity-cliff compounds.

Scientific Applications:

• Model evaluation in drug discovery: Evaluates model reliability on activity cliffs to inform selection and refinement of predictive models used in drug discovery.
• Molecule discovery and optimization: Provides targeted assessment of activity-cliff prediction to aid molecule discovery and optimization workflows.
• Metric and algorithm development: Informs the development of activity-cliff-centered metrics and algorithms to better capture abrupt activity changes among structurally similar compounds.

Methodology:

Utilizes curated bioactivity data from diverse macromolecular targets, advocates inclusion of "activity-cliff-centered" metrics during model development and evaluation, and supports development of algorithms specifically aimed at predicting properties associated with activity cliffs.