Calpha uses Qwen3.5-9B to output single-chain protein structures
Original title: Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model
A SimpleFold head reads only Qwen residue states; protein training cut a 350-item MMLU-Pro subset from 77.4% to 70.3%.
Calpha is a first step toward a general model that both reasons in language and predicts biomolecular structure. Qwen3.5-9B reads a sequence enriched with ESM2 residue features and can answer in text or call a pretrained SimpleFold head that sees the protein only through Qwen's residue states. Under a fixed folding request, mean TM-scores are 0.808 on FoldBench-334, 0.635 on CASP16 domains, and 0.900 on a CAMEO-style PDB set. For chains of 300–600 residues the reported mean gaps versus an ESM2-3B plus fine-tuned SimpleFold-3B baseline are negative, and complex prediction remains future work.
Unconfirmed: Score gaps are not a matched comparison because training setup and crop length differ.
- Four English request types triggered folding on every evaluated example.
- For chains of 100–299 residues, all three paired confidence intervals include zero.
- At 300–600 residues, mean differences are -0.044, -0.160, and -0.025.
- The compared systems differ in trainable weights, training duration, and crop length.
- AI models
- Qwen3.5-9BESM2SimpleFoldCalpha
From the original article
1 min read at huggingface.coBiomolecular foundation models predict the structures of proteins and their complexes accurately, and large language models (LLMs) reason in natural language, but the two remain separate: the abilities of the former are not part of a general-purpose LLM, and the reasoning of the latter does not enter a structure prediction.
By Yong Liu, Zhanpeng Shi, Peicong Lin · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.