All-Atom Structure Tokenizer: AtomLM-9b for Biomolecular Structure Prediction

Abstract

We develop an All-atom Structure Tokenizer that compresses biomolecular geometry into discrete structural tokens and achieves high-accuracy all-atom reconstruction. A hierarchical geometric encoder and finite scalar quantization map chemical units to a shared vocabulary of 8,192 tokens, while a joint coordinate de-tokenizer reconstructs backbone, side-chain, nucleic-acid, and ligand heavy atoms. The tokenizer achieves all-atom lDDT scores of 0.832 and 0.835 on CASP14 and CASP15, respectively. Building on this tokenizer, we train AtomLM-9b, an autoregressive language model based on Qwen3.5-9B that predicts structural tokens from molecular identities and connectivity and decodes them into complete atomic structures. We demonstrate, for the first time, protein structure prediction performance comparable to AlphaFold 3 with an all-atom-tokenizer-based structural language model: on the FoldBench protein-protein task, AtomLM-9b achieves a success rate of 75.01%, compared with the published AlphaFold 3 result of 72.93%. These results establish discrete all-atom tokenization as an effective interface between high-fidelity atomic reconstruction and language-model-based structure prediction.