Molecular property prediction gets one shared testbed: 17 datasets and 800,000 records, with physics constraints cutting violations to 1.8%

Molecular property prediction now has a unified benchmark spanning single molecules and multi-component systems, placing density, conductivity, solubility, viscosity, enthalpy of mixing and other properties into one schema and one splitting rule across 10 property tracks. Earlier benchmarks mostly focused on a single property, a specific molecular task or a particular chemical domain, with inconsistent datasets and evaluation protocols that hindered systematic cross-task comparison. Chem World integrates 17 public datasets with more than 800,000 molecular and chemical-system instances; for comparison, QM9 holds about 130,000 molecules and CheMixHub is at the 500,000 scale.

PACM combines pretrained molecular representations, pairwise interaction modeling between components, and task-selected physical constraints, so molecular-level representations extend to multi-component systems. Simply averaging component representations can overlook composition-dependent contributions and non-additive mixture behavior, and can yield physically implausible predictions. On Chem World, MolFormer with PACM reaches a macro normalized RMSE of 0.367, below 0.433 for MolFormer with DeepSets and 0.416 for MolT5 with Set Transformer.

Removing the physical constraints raises the error, so these constraints are not decoration. Full PACM has a macro normalized RMSE of 0.367; removing physical-boundary constraints raises it to 0.412, the largest degradation in the ablation. In the same ablation, removing pairwise interactions raises it to 0.389, removing composition regularization to 0.398, and removing applicable temperature constraints to 0.382.

The evaluation covers the 17 public datasets and 10 property tracks integrated into Chem World, including single-molecule and multi-component systems; predictions are positioned for early-stage scientific screening and experimental planning, not as a replacement for wet-lab validation, safety evaluation, or regulatory and clinical decision-making.

The integrated datasets vary in measurement protocols, conditions and data quality; group-aware splitting reduces leakage between related observations but does not guarantee generalization to unseen molecules, compositions or experimental conditions; the physical constraints rely on task-specific assumptions and may not capture all complex molecular interactions or thermodynamic behaviors.