Pinal turns plain-language prompts into protein designs

Protein design tools have generally required users to specify detailed constraints, such as which amino acids must be placed where in three-dimensional space, before the AI can build a structure around them. Pinal changes the starting point by letting users describe a function in ordinary language, relying on a two-step process where T2struct translates the prompt into a structural blueprint and SAProt-T builds the amino acid sequence. The training drew on protein sequences from three databases paired with AlphaFold-predicted structures, and one database alone contained 800 million keyword–protein pairs. The low hit rate in testing reflects a deliberate filter: the researchers rejected candidates that looked too similar to known fluorescent proteins, so many likely-working designs were weeded out before testing. In a separate but related announcement, Hugging Face, the Zhongguancun Academy, and the University of Naples “Federico II” introduced Carbon X, an open-weight DNA model that can run on local hardware. Carbon X is based on Qwen3-14B and was retrained on DNA, then fine-tuned, to handle both DNA and natural-language prompts. The team has not released a publication describing its behavior, but internal performance is described as competitive with large commercial models like GPT-6.1 Sol, with cautions about position-specific accuracy and hallucinations. Both tools suggest that natural-language interfaces are reaching into biology, but the need for lab validation remains a limiting step.