MoleculeNet: a benchmark for molecular machine learning.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Machine learning (ML) in molecular science has advanced, improving molecular property predictions.
• Lack of standardized benchmarks hinders comparison of new ML algorithms for molecular tasks.

Purpose of the Study:

• Introduce MoleculeNet, a comprehensive benchmark for molecular machine learning.
• Facilitate reproducible evaluation and comparison of diverse molecular ML algorithms.

Main Methods:

• Curated multiple public datasets for molecular property prediction.
• Established standardized evaluation metrics and protocols.
• Provided open-source implementations of molecular featurization and learning algorithms within DeepChem.

Main Results:

• MoleculeNet benchmarks show learnable representations generally yield superior performance.
• Performance of learnable representations is limited in data-scarce or imbalanced classification scenarios.
• Physics-aware featurizations outperform algorithm choice for quantum mechanical and biophysical datasets.

Conclusions:

• MoleculeNet establishes a crucial benchmark for advancing molecular machine learning.
• Learnable representations are powerful but require careful consideration for complex tasks.
• Physics-informed approaches are vital for specific scientific domains in molecular ML.