PCQM4Mv2

Predict each molecule’s HOMO-LUMO energy gap, a continuous quantum-chemical property calculated with density functional theory. Mean absolute error in electron volts is the official metric.

PCQM4Mv2 contains 3,746,619 PubChemQC molecules. Training molecules also have DFT-optimized 3D structures, while SMILES describe the final optimized connectivity; test labels are held by OGB-LSC.

The default 2D graph conversion uses OGB’s nine categorical atom and three categorical bond fields from SMILES. Coordinates are optional additional inputs for training molecules and are unavailable in the same form for blind test molecules.

Molecules are ordered and split by PubChem compound identifier into 90% train, 2% validation, 4% test-dev, and 4% test-challenge. Despite an original intention to use scaffolds, the released split is CID-based because of a preprocessing bug.

PCQM4Mv2 corrected an approximately 10% mismatch in PCQM4M between initial SMILES and gaps computed after geometry optimization, so the versions are not interchangeable. The CID split is not a scaffold split. Test-dev and test-challenge are distinct blind sets, and using supplied training coordinates creates an input-availability asymmetry.