Researchers Unveil All-Atom Tokenizer for Protein Prediction
A new structure tokenizer and language model report gains over AlphaFold 3 on a protein-protein prediction benchmark.
Updated on Nov. 4, 2026 in Biotech
Researchers introduced an All-atom Structure Tokenizer and AtomLM-9b, an autoregressive language model based on Qwen3.5-9B, for biomolecular structure prediction. AtomLM-9b reports a 75.01% success rate on the FoldBench protein-protein task, compared with AlphaFold 3's published 72.93%.
Why it matters
The work applies a language-model approach to atomic-scale biomolecular structure, a field where prediction accuracy shapes drug design and molecular research. Results reported so far are preprint findings and have not yet been independently replicated in peer-reviewed form.
The tokenizer maps chemical units to a shared vocabulary of 8,192 tokens and records all-atom lDDT scores of 0.832 on CASP14 and 0.835 on CASP15. AtomLM-9b, at 9 billion parameters, posts 75.01% on FoldBench protein-protein against AlphaFold 3's published 72.93%, and performance on other task types is not reported.
The details
A hierarchical geometric encoder and finite scalar quantization map chemical units to structural tokens, and a joint coordinate de-tokenizer reconstructs backbone, side-chain, nucleic-acid and ligand heavy atoms. The approach treats molecular structure as a sequence that a language model can predict, rather than relying solely on specialized structure-prediction architectures.
Timeline
• Article published on the bioRxiv preprint server on November 4, 2026.
Big Picture
The result measures AtomLM-9b directly against AlphaFold 3's published 72.93% on the FoldBench protein-protein task, the benchmark that now sets the comparison point for atomic-scale structure prediction. Whether the 75.01% margin holds under independent evaluation will determine how much weight the comparison carries.
For researchers and drug developers, the reported gains suggest a language-model approach may be worth evaluating alongside existing structure-prediction tools once the work is peer reviewed. Practical adoption depends on independent replication and on whether results extend beyond the protein-protein benchmark.
The takeaway
The reported results point to language models as a serious candidate for atomic-scale structure prediction, but the claims rest on preprint benchmarks. Readers should watch for independent replication and peer review before treating the gains over AlphaFold 3 as settled.
Further reading
For more on the research tools behind these advances, see Biotech.
Source note: This article includes information reported by Biorxiv.
