Westlake University team publishes Pinal model in Nature: a single sentence can generate candidate proteins

On October 22, reporters learned from Westlake University that a research team from the university, in collaboration with teams from The Hong Kong University of Science and Technology (Guangzhou) and the Shenzhen Institute of Advanced Technology of the Chinese Academy of Sciences, has built an artificial intelligence (AI) model named Pinal. Given only a single sentence of text description, the model can generate candidate proteins that differ from known natural protein sequences and may possess practical functions. The related research findings were recently published online in the journal Nature. Proteins are biological macromolecules that carry out various specific functions through the precise spatial structures formed by layer upon layer of folding. "In the three-dimensional structures reconstructed by cryo-electron microscopy, proteins present complex and intricate folded forms, with tens of thousands of atoms arranged in an orderly manner at the nanoscale," said Yuan Fajie, a distinguished research fellow at Westlake University's School of Engineering. He noted that the difficulty of protein design lies precisely in the highly intricate nature of protein structures, where local changes may affect overall folding and function. In recent years, with the rapid development of technology, AI has become capable of predicting the relationship between protein structures and sequences. The Westlake University research team hopes to go a step further by directly using natural language descriptions of protein functions to enable AI to generate the corresponding protein sequences. Concept image of Pinal. (Image generated by GPT-Image-2) Image courtesy of Westlake University. During the research, the team found that this attempt required bridging a huge data gap: the scale of known protein sequence data is enormous, but data with experimentally validated functions is relatively scarce, making it difficult for AI models to establish a reliable correspondence between functional text and protein sequences. To overcome this challenge, the research team proposed a technical approach using protein three-dimensional structures as an intermediate bridge to develop and train the Pinal model. They integrated protein data from multiple sources and constructed a training dataset covering approximately 135 million distinct protein sequences. By equipping the same protein with multiple linguistic descriptions, the team ultimately built a synthetic AI training dataset containing approximately 1.7 billion "protein-text" pairs, providing a foundation for the Pinal model to learn the associations between linguistic descriptions and proteins. The model's operation roughly consists of three steps: first, generating a structural blueprint of the protein based on the text; then arranging amino acids according to the structure; and finally scoring numerous candidate amino acid sequences to select the candidate sequence that best matches the target. The research team selected four types of proteins with vastly different functions for experimental validation: fluorescent proteins, PET plastic-degrading enzymes, alcohol dehydrogenases, and shuttle H proteins. Together, the four experiments confirmed that the Pinal model can design proteins that differ from known natural protein sequences and perform target functions in experiments. Yuan Fajie stated that protein design models based on natural language descriptions are still at the proof-of-concept stage, and that more refined protein function design requirements—such as specifying the emission wavelength of fluorescent proteins through language instructions—remain a research challenge for the team and the scientific community at present.