Megadose The AI curve, compressed.

Calpha: Toward Native Biomolecular Intelligence in a General-Purpose Language Model

· HF Daily Papers ·
Calpha produces single-chain protein structures through an LLM, but longer-chain performance and general reasoning accuracy take a hit.

Qwen3.5-9B reads protein sequences enriched with ESM2 features and passes its residue states to a pretrained SimpleFold structure head. For chains of 100–299 residues, comparisons with the baseline were inconclusive; at 300–600 residues, Calpha scored lower across all three evaluation sets, though training conditions also differed. Protein training reduced accuracy on a 350-question MMLU-Pro subset from 77.4% to 70.3%. Predicting complexes, understanding structures and showing that language reasoning changes predicted coordinates remain future work.

HF Daily Papers' note

score 5

Categories: Research