Uni-Bond: Learning Chemical Bonds from Atomic Coordinates

Recovering a molecular graph from 3D atomic coordinates is a fundamental step in computational chemistry, linking continuous geometry to the discrete graphs on which chemoinformatics and property prediction depend. Rule-based bond perception is brittle to geometric distortion and strained conformations, motivating learned alternatives. We introduce \textbf{Uni-Bond}, which predicts typed bonds directly from atom types and Cartesian coordinates by combining a pretrained Uni-Mol encoder with a lightweight pairwise classification head. On the GEOM scaffold split, Uni-Bond reduces the exact-match error rate by roughly 19$\times$ relative to the strongest baseline, and transfers zero-shot to peptides, water clusters, and molecular dimers, reliably distinguishing covalent bonds from close non-covalent contacts. We further study direct geometry-to-SMILES generation with a coordinate-tokenized language model: it performs well on GEOM but collapses under chemical domain shift, indicating that structured geometric foundation models offer a more reliable route from 3D coordinates to molecular graphs.