CoordToken: Tokenization of 3D Molecules for Autoregressive Generation and Geometry Refinement

Atom-level generation of three-dimensional molecular structures is an increasingly active research direction. Because atomic coordinates are continuous, the field has been dominated by diffusion and flow-based methods, with comparatively few autoregressive alternatives. We argue that an accurate discrete tokenizer is an important missing ingredient for autoregressive 3D molecular modeling. We introduce CoordToken, an FSQ-based tokenizer trained on a corpus of 190 million conformers. Across diverse evaluation datasets, CoordToken achieves a micro-averaged reconstruction RMSD of 0.072 Å, while 98% of reconstructed structures pass all applicable PoseBusters checks. Replacing conventional coordinate representations with CoordToken in Qwen3- and BindGPT-style autoregressive models consistently improves conformer generation on GEOM-DRUGS, both in terms of precision and PoseBusters validity. Beyond generation, CoordToken functions as a conformer refiner. It repairs randomly perturbed structures, improves the chemical validity of conformers generated by diffusion and flow-based methods, and improves ligand-validity success for seven of eight protein–ligand pose-prediction methods. The tokenizer, its 190-million-conformer training corpus, and the autoregressive models trained with it will be publicly released.