Programmable design of functional proteins from natural language.

Pinal, a 16-billion-parameter model, successfully designs functional proteins from natural-language descriptions, including enzymes with catalytic activity, without iterative optimization.

• Why it matters: Current protein design methods are limited by their reliance on existing proteins or manual constraint formulation, hindering rapid, flexible, and high-level functional specification.
• What they did: Researchers trained Pinal on 1.7 billion synthetically annotated protein-text pairs to link functional descriptions to protein sequences and structures, enabling de novo candidate generation.
• The result: Pinal-produced proteins demonstrated high predicted foldability, functional alignment, and sequence diversity, with experimental validation showing functional activity and a 1.7-fold increase in product titre for an engineered H-protein.