NovoMCP Open Corpus (lite)

122,454,458 PubChem compounds with RDKit physicochemical descriptors, ADMET/toxicity predictions from NovoMCP's own models (most trained on the public Therapeutics Data Commons ADMET benchmark; a few base heads on other public toxicology data such as Tox21), a machine-learned logP, and PAINS structural-alert flags. One row per PubChem CID, stored as Apache Parquet (Snappy) for direct in-place querying (Athena, DuckDB, PyArrow, pandas). An analysis-ready cheminformatics substrate for drug discovery: virtual screening, drug-likeness (QED) filtering, similarity search, chemical-space exploration, and training ADMET models at scale — so you don't recompute descriptors and property predictions across all of PubChem yourself. Per-endpoint model provenance is documented in the dataset card. Computed properties and model predictions for research use — not experimental measurements or regulatory determinations.