Westlake University Team Develops AI Model Pinal That Generates Proteins from Text Descriptions

Researchers at Westlake University, in collaboration with Hong Kong University of Science and Technology (Guangzhou) and the Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, have developed an AI model named Pinal. The model can generate candidate protein sequences that differ from known natural proteins and may have practical functions, based solely on a sentence describing the desired protein. The findings were published online in Nature.

Protein design is challenging due to the precision of three-dimensional structures, where local changes can affect overall folding and function. While AI has already been used to predict the relationship between protein structure and sequence, the team aimed to go further by letting AI generate protein sequences directly from natural language descriptions of function. To address the scarcity of experimentally verified functional data, the researchers used protein three-dimensional structures as an intermediate bridge. They built a training dataset covering approximately 135 million protein sequences and constructed a synthetic dataset of about 1.7 billion protein-text pairs by associating each protein with multiple language descriptions.

The model works in three steps: generating a structural blueprint from text, arranging amino acids according to the structure, and scoring candidate sequences to select the best match. The team validated Pinal across four functionally distinct proteins, including fluorescent proteins, PET-degrading enzymes, alcohol dehydrogenases, and shuttle protein H. Results showed that Pinal could design proteins distinct from known natural sequences while still performing the target functions in experiments. Lead researcher Yuan Fajie noted that the model is still at the proof-of-concept stage, and more refined functional design requirements remain a challenge for the team and the broader scientific community.