Tell this AI what you want a protein to do—and it designs one

This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: In a short space of time, artificial intelligence has become a whiz at designing proteins, molecular machines that perform various functions such as speeding up reactions, producing fluorescent signals and breaking down plastic. But getting AI to design the specific protein you want usually requires the user to have a strong understanding of molecular structures and geometric constraints to guide the models correctly.

Well, that requirement may soon be a thing of the past. Scientists led by a team at Westlake University in China have developed a large-scale AI model called Pinal that can generate candidate proteins from natural language prompts alone. They detail their research and how their AI model works in a paper published in the journal Nature.

All a user has to do is enter a written description of the job they want the protein to do. For example, they might type something like "design a fluorescent protein that emits green light under illumination." Pinal, which uses two main modules, takes this description and builds a 1D structural blueprint using its first module, T2struct. A second module, SaProt-T, reads that blueprint and uses the original text prompt to generate a matching amino acid sequence.

"Pinal translates high-level functional descriptions into candidate protein sequences, thereby reducing reliance on manually specified structural templates or geometric constraints," wrote the study authors. One of the many clever things about Pinal is that it doesn't just pick the first shape that comes along. It creates several structural blueprints, designs the best sequence for each and scores each pair to determine the best final design.

Training Pinal required vast amounts of data. The team started with 135 million protein sequences and used AI models and existing protein annotations to build plain-language descriptions of what those proteins do. This resulted in a massive library of 1.7 billion protein-text pairs.

They then obtained the predicted 3D structures for those sequences from an existing AlphaFold database. To test Pinal, the researchers asked it to design four different proteins. These were a fluorescent protein, a plastic-breaking enzyme, an alcohol dehydrogenase and a metabolic H-protein.