New AI tool designs proteins from a plain-English request

A team of researchers based in China has built an AI tool that designs proteins from a plain-English request. It is called Pinal. Instead of feeding in technical specifications, you describe the function you want. The catch is that many of the designs may not work, so someone with a lab still has to sort the good ones from the failures.

What changed

Proteins are the tiny machines of living things. Today's best design tools work in two steps. First, you set constraints, such as binding another molecule. One AI tool then generates a shape, and a second tool fills in the amino acids, the building blocks that make the protein fold into that shape. This takes real expertise in protein chemistry.

Pinal also uses two steps, but the starting point is different. You simply say what you want. A module called T2struct turns the request into a structural blueprint. A second system, SAProt-T, then builds the amino acid sequence from the original text and that blueprint.

How it was trained

The team used large collections of protein sequences, each tagged with notes about what the protein does. For each set of tags, an AI language model wrote a plain-language description. Both the tags and the descriptions went into training.

• Proteins came from three databases.
• One database alone held 800 million keyword and protein pairs.
• Structure data came from AlphaFold, an AI system that predicts protein shapes.

AlphaFold is not 100 percent accurate. That could be a problem for obscure proteins. Many common proteins have real structural data, which should help.

The test

The researchers asked Pinal to design four very different proteins: a fluorescent one, two enzymes, and one that takes part in a large multi-protein complex. For the fluorescent one, they asked for 50,000 designs that could glow like green fluorescent protein, or GFP. The discovery of GFP won a Nobel.

They then filtered that list. They threw out anything whose sequence looked like a known fluorescent protein. They kept designs with a barrel-like structure and a key amino acid that can be chemically modified to enable glowing. The goal was something like GFP in general, but not a copy.

What to keep in mind

The tool lowers the skill needed to ask for a design. It does not remove the need to test. Separately, Hugging Face said it is now hosting a tool that does similar things for DNA. The material does not say how many of Pinal's designs worked in the end, so that is the number to look for.