BondNet: molecular bond perception from 3D coordinates with explicit hydrogen context

Inferring molecular bonds from atomic coordinates is necessary when a reliable molecular graph is unavailable. We study the contribution of explicit hydrogens to this task using BondNet, a PaiNN-based network that jointly predicts connectivity and heavy–heavy bond types. On 27,240 GEOM-DRUGS test records, exact heavy–heavy graph recovery is 99.83% for clean coordinates and 69.56% at Gaussian coordinate noise σ =0.20 σ = 0.20 Å. The corresponding rates on a separately selected cohort of 10,000 PubChem3D molecules are 98.43% and 64.60%. At the higher noise level, default RDKit recovers 3.64% of PubChem3D graphs exactly; a Hückel configuration examined after the initial evaluation recovers 46.73%. A heavy-only BondNet reaches 50.70% exact match on GEOM at σ =0.20 σ = 0.20 , and the explicit-H advantage persists under matched heavy-atom perturbations. An oracle model supplied with reference hydrogen counts reaches 82.03%, suggesting that hydrogen-derived valence information contributes substantially to the observed benefit. Removing H atoms from the explicit-H model at inference sharply reduces bond-type accuracy. These results highlight the difference between bond-level and complete-graph accuracy. The GEOM test set was inspected during development, and the findings do not extend to charged molecules or authentic generated geometries. Scientific contribution We compare explicit-H and heavy-only input under the same training schedule and use an oracle hydrogen-count control to probe the information H atoms contribute. Heavy–heavy graph metrics, a separate PubChem3D cohort, and chemically normalized baselines show where learned bond perception succeeds and where high bond-level accuracy obscures reconstruction failures.