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

AtomLM-9b achieves 75.01% success on the FoldBench protein–protein task, matching AlphaFold 3’s reported 72.93% using an all-atom structural tokenizer interface.

• Why it matters: Protein structure prediction needs a bridge between high-fidelity atomic reconstruction and language-model sequence learning, where discrete geometry representations could improve usability and accuracy.
• What they did: The work builds an all-atom structure tokenizer using a hierarchical geometric encoder and finite scalar quantization to map chemical units into 8,192 tokens, then trains AtomLM-9b to predict these tokens from molecular identities and connectivity.
• The result: The tokenizer reconstructs atomic structures with all-atom lDDT of 0.832 on CASP14 and 0.835 on CASP15, and AtomLM-9b’s FoldBench success of 75.01% is comparable to AlphaFold 3’s 72.93%.

AI summary of the abstract · may contain errors · not medical advice · read the paper.