Researchers introduced Pinal, a 16-billion-parameter AI model. It designs proteins from plain-language descriptions of desired function. Pinal was trained on 1.7 billion synthetic protein–text pairs. It links functional intent to protein sequence and structure. Computational tests showed high predicted foldability and alignment with descriptions. Four designed proteins were experimentally validated. A Pinal-designed H-protein boosted CO2-fixation yield by 1.7 times. The model reduces reliance on manual structural constraints. Pinal is open-sourced under Apache 2.0. Model weights and a web server are publicly available.
What changed
Protein design required natural proteins as starting points or manual constraints.
What it unlocks
Designing functional proteins directly from natural-language descriptions.
- 16 billion parameters
- 1.7 billion synthetically annotated protein–text pairs
- 1.7-fold product titre increase
Sources