Toward De Novo Protein Design from Natural Language

Pinal, a 16-billion-parameter model, translates natural language into diverse, functional proteins, with all four designed proteins demonstrating activity and enhanced performance.

• Why it matters: Current protein design methods are limited by the need for existing proteins or specialized expertise, hindering rapid and accessible creation of new functions.
• What they did: Researchers trained Pinal on 1.7 billion protein-text pairs to ground functional language in protein biophysics and tested it by designing four distinct proteins, including enzymes and a fluorescent protein.
• The result: All four proteins were functionally active, with two enzymes achieving catalytic turnover and the H-protein outperforming its natural version by 1.7 times, enabling direct, language-guided protein creation.