PCQM-Contact

Rank atom pairs that are more than five covalent-bond hops apart but less than 3.5 angstrom apart in the molecule’s computed 3D geometry. The benchmark frames this contact-map problem as link prediction on previously unseen molecules.

The 529,434 molecules come from PCQM4Mv2 training molecules with available 3D structures and at least one qualifying contact. Unlike most molecular graph benchmarks, hydrogen atoms are explicit nodes because hydrogen bonding is relevant to the contact target.

Inputs are 2D covalent graphs using the OGB nine-field atom encoder and three-field bond encoder. The 3D coordinates determine labels but are not input features, and the predicted contact itself has no edge type.

Molecules are randomly divided 90%/5%/5% into 476,490 training, 26,472 validation, and 26,472 test graphs. Evaluation uses filtered MRR and Hits@1/3: when ranking a true contact, other known true contacts for the query atom are masked rather than counted as negatives.

Training supplies positive contacts without authoritative hard-negative labels; an unobserved pair is not necessarily a chemically meaningful negative. The random molecule split does not enforce scaffold separation. Results depend on filtered ranking and explicit hydrogens, and therefore are not comparable to generic binary link prediction or PCQM4Mv2 HOMO-LUMO regression.