Plain-language prompts now steer protein design in a new Chinese-built tool

A tool called Pinal turns written descriptions of a protein's function into designs, though how many actually work is still unclear.

A large collaboration of researchers based in China has built Pinal, a protein design tool that accepts natural-language prompts describing the function you want. Separately, Hugging Face announced it is hosting a tool that does similar natural-language design at the DNA level, though the material we have gives no details on it.

The usual approach has two steps: one AI generates a protein shape that meets the user's constraints, and a second tool fills in an amino acid sequence that folds into it. Using those tools takes detailed knowledge, such as which amino acids are critical to a reaction and where they sit in 3D space. Pinal also runs in two stages. T2struct converts the prompt into a structural blueprint, a linear array of tokens describing each amino acid's relationship to its nearest neighbors, with probabilities attached. SAProt-T then combines the original text with that blueprint to build the final sequence. For training, the team used an LLM to turn function tags into plain-language descriptions. It drew on three databases, one of which held 800 million keyword–protein pairs, paired with AlphaFold-predicted structures.

The team tested four designs: a fluorescent protein, two enzymes, and part of a large multi-protein complex. For the fluorescent protein, Pinal was asked for 50,000 GFP-like candidates. After filtering out look-alikes of known proteins and requiring a barrel-like structure and the key modifiable amino acid, 29 high-priority candidates remained.

This matters because expertise is the bottleneck. If someone can describe a function in words and get plausible designs, more people can work on enzymes, drugs and crop traits, and the cost of early exploration falls.

The caveats are real. The source suggests the success rate may be modest, with many failed designs needing lab sorting. AlphaFold's imperfect predictions could hurt results for obscure proteins. The source text we have is truncated, so we don't know how the 29 candidates or the other tests turned out. That lab data is what to watch for. The full reporting at Ars Technica has the detail.

Source: Ars Technica