MoleculeNet is a large-scale benchmark for molecular machine learning, curated by Stanford University researchers. It integrates multiple public datasets, establishes evaluation metrics, and provides high-quality open-source implementations of various molecular featurization and learning algorithms. The benchmark aims to facilitate the development of molecular machine learning methods by addressing key issues such as limited data, diverse output predictions, and complex molecular structures. MoleculeNet includes over 700,000 compounds across four categories: quantum mechanics, physical chemistry, biophysics, and physiology. It offers a suite of software that implements known featurizations and evaluates algorithms using different splits and metrics. The benchmark demonstrates that learnable representations are powerful for molecular machine learning but highlights challenges in handling complex tasks under data scarcity and imbalanced classification. Physics-aware featurizations are particularly important for quantum mechanical and biophysical datasets. MoleculeNet's comprehensive approach and detailed results aim to trigger breakthroughs in molecular machine learning, similar to the impact of ImageNet in computer vision.MoleculeNet is a large-scale benchmark for molecular machine learning, curated by Stanford University researchers. It integrates multiple public datasets, establishes evaluation metrics, and provides high-quality open-source implementations of various molecular featurization and learning algorithms. The benchmark aims to facilitate the development of molecular machine learning methods by addressing key issues such as limited data, diverse output predictions, and complex molecular structures. MoleculeNet includes over 700,000 compounds across four categories: quantum mechanics, physical chemistry, biophysics, and physiology. It offers a suite of software that implements known featurizations and evaluates algorithms using different splits and metrics. The benchmark demonstrates that learnable representations are powerful for molecular machine learning but highlights challenges in handling complex tasks under data scarcity and imbalanced classification. Physics-aware featurizations are particularly important for quantum mechanical and biophysical datasets. MoleculeNet's comprehensive approach and detailed results aim to trigger breakthroughs in molecular machine learning, similar to the impact of ImageNet in computer vision.