Computational & AI Journal of Cheminformatics (Springer Nature)
Explicit hydrogens help a neural network recover molecular bonds from noisy 3D coordinates
- BondNet, an equivariant network, predicts connectivity and heavy-atom bond types from 3D coordinates, and was used to test how much explicit hydrogens contribute.
- It was evaluated retrospectively on a conformer test set and a separate PubChem3D cohort, with and without added coordinate noise.
- Full-graph recovery was near-perfect on clean input and fell with noise, but stayed far above default RDKit; hydrogen-free inputs did worse. Limits: charged molecules and generated geometries untested.
Bond orders recovered from docked or generated poses are a familiar headache, so a learned alternative to RDKit's distance-based perception is welcome. But the untested cases, charged molecules and generated geometries, are where it is most needed, so it can't yet fix that workflow. ML-generated ligands, tautomers and conjugated systems would be the real test.
Paper: Journal of Cheminformatics (Springer Nature) ↗ BondNet: molecular bond perception from 3D coordinates with explicit hydrogen context