The central test here is not whether a language model can produce protein coordinates, but whether its own representations can serve as a useful bridge between language and structural biology. Calpha demonstrates that pathway, while its results also show that connecting the modalities does not by itself improve folding accuracy.
Key Findings
- Qwen3.5-9B receives protein sequences enriched with residue features from ESM2-3B; a generated trigger token routes its residue states to a pretrained SimpleFold-3B structure head. The head receives protein information through those states rather than a separate sequence input.
- The system reports mean TM-scores of 0.808 on FoldBench-334, 0.635 on CASP16 domains, and 0.900 on a CAMEO-style PDB set. These figures establish a workable integration route, not a new accuracy record.
- Comparisons with an ESM2-to-SimpleFold baseline are mixed. For 100–299-residue chains, paired confidence intervals include zero; at 300–600 residues, Calpha scores lower, with reported mean differences of -0.044, -0.160, and -0.025 across the three sets.
- Protein training reduced accuracy on a 350-item MMLU-Pro subset from 77.4% to 70.3%. This makes the trade-off between adding a specialized capability and preserving general-purpose performance concrete.
Who It Suits—and the Trade-Off
A good fit for researchers studying how general-purpose LLM representations can connect to scientific prediction systems, especially protein-structure modeling. Look elsewhere if you need state-of-the-art folding, reliable long-chain results, complexes, ligands, or evidence that natural-language reasoning improves predicted structures; those capabilities are not established here.