Artificial intelligence is increasingly used to create proteins - molecules that accelerate reactions, glow or break down plastic. But to obtain a desired protein, a person usually has to understand molecular structures and geometric constraints. This may soon become optional.
A group of researchers led by scientists from Westlake University (China) created a large-scale AI model called Pinal. It can generate candidate proteins based only on a text description of the task. The work was published in the journal Nature.
The user only needs to write what the protein should do, for example: "create a fluorescent protein that glows green under illumination." The model consists of two modules. The first, T2struct, turns text into a one-dimensional structural "blueprint." The second, SaProt-T, reads this blueprint and, taking the original request into account, outputs a suitable amino acid sequence. The authors note that Pinal translates functional descriptions into candidate sequences, reducing reliance on manually specified templates and geometric constraints.
The model does not stop at the first option that comes to mind: it creates several structural blueprints, selects the best sequence for each and evaluates the pairs to choose the final design.
Training Pinal required enormous data. The scientists took 135 million protein sequences and, using AI and existing annotations, compiled descriptions of them in plain language. The result was a library of 1.7 billion "protein - text" pairs. The predicted three-dimensional structures of these sequences were taken from the AlphaFold database.
The researchers then asked Pinal to create four proteins: a fluorescent one, a plastic-degrading enzyme, alcohol dehydrogenase and a metabolic H-protein. The results were mixed. For the glowing protein, the model proposed many sequences, of which 29 were selected for testing. Only one fluoresced. The authors acknowledge that controlling specific properties is still difficult: for example, setting the exact emission wavelength remains a hard task.
Things went better with the metabolic H-protein: of 20 candidates, five worked, and the best increased the yield of the key carbon-binding process by 1.7 times compared with the natural E. coli protein.
Despite its promise, Pinal has limitations. Further refinements are needed before the model becomes an ordinary laboratory tool.