Datasets

SPICE provides quantum-mechanical energies, forces and other molecular properties for training machine learning potentials, with an emphasis on drug-like molecules and protein interactions.

Quantum Chemistry · Scientific Data

ANI-1 provides calculated off-equilibrium molecular conformations, with Python readers for accessing HDF5 files containing coordinates and energies for organic molecules.

Quantum Chemistry

GEOM provides 37 million energy- and statistical-weight-annotated molecular conformations for over 450,000 molecules, with MessagePack data, RDKit objects, and loading and analysis tutorials.

Molecular Generation · Computational Chemistry

Matbench Discovery benchmarks machine-learning models for crystal stability and atomistic simulation tasks, using an interactive leaderboard to compare accuracy, robustness, and computational cost.

Materials Discovery

Matbench provides 13 curated materials-science machine learning tasks for benchmarking property-prediction methods, with benchmark data, leaderboards and a pip-installable package.

Materials Discovery

Open Reaction Database provides structured reaction datasets as Parquet files containing Protobuf records, with mirrored downloads and ord_schema workflows for streaming access and text or JSON conversion.

Reaction Prediction · Scientific Data

Therapeutics Data Commons provides therapeutic machine learning datasets, Python loaders, data splits, evaluation metrics and benchmarks for prediction and molecule-generation research.

Molecular Property Prediction · Drug Discovery

MOSES combines a ZINC-derived molecular dataset, generation baselines and evaluation metrics to benchmark the validity, diversity and novelty of generated molecules.

Molecular Generation

GuacaMol is a Python benchmarking package for de novo molecular design, with distribution-learning and goal-directed evaluations plus standardized ChEMBL-derived datasets.

Molecular Generation