SJTU's Calpha Folds Proteins Inside Qwen3.5-9B at 0.808 TM
TL;DR
- Calpha wires a SimpleFold structure head into Qwen3.5-9B so the LLM itself emits single-chain protein coordinates from its own residue states.
- The model reaches mean TM-scores of 0.808 on FoldBench-334, 0.900 on a CAMEO-style PDB set and 0.635 on CASP16 domains.
- Protein training cut MMLU-Pro accuracy on a 350-item subset from 77.4% to 70.3%, and long-chain scores trail an ESM2-3B plus SimpleFold-3B baseline.
Calpha, from a Shanghai Jiao Tong University group led by Shuangjia Zheng, wires a protein structure head directly into Qwen3.5-9B so the language model itself emits single-chain coordinates from its own residue states. On a paper posted to Hugging Face, the authors report mean TM-scores of 0.808 on the FoldBench-334 development set, 0.900 on a CAMEO-style PDB set and 0.635 on CASP16 domains.
The pipeline is unusual. Qwen3.5-9B reads a sequence enriched with ESM2 residue features and can either respond in text or emit a token that invokes a pretrained SimpleFold structure head. That head, the paper notes, "receives information about the protein only through Qwen's residue states" — the LLM is the sole channel between sequence and structure.
There is a bill. Protein training dropped accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. Against a baseline that pairs ESM2-3B with a fine-tuned SimpleFold-3B head, Calpha is within noise on 100-to-299-residue chains, where "all three paired confidence intervals include zero." At 300-to-600 residues it trails: mean TM differences of -0.044, -0.160 and -0.025 across the three benchmarks.
The authors state the scope plainly: "An LLM can therefore output single-chain structures from its own states, with lower long-chain scores in the current comparisons and reduced accuracy on the evaluated MMLU-Pro subset." Complex prediction, structure understanding and the effect of language reasoning on predicted coordinates are all deferred to future work. The paper lands amid a steady run of Chinese-lab open releases on our open-source tracker, including the same-day MBZUAI K2-Horizon fleet.
Originally reported by huggingface.co
Read the original article →Original headline: Shanghai Jiao Tong's Calpha Bakes Protein Folding Into Qwen 3.5-9B, Hits 0.808 TM-Score