🧬Chinese Researchers Build Pinal, Turning Text Prompts Into Working Proteins

Your protein design pipeline might soon take plain English

A big research collaboration in China just built a tool that takes a plain-English prompt and spits out a protein with the function you asked for. It's called Pinal, and it works in two steps: a text model turns your prompt into a structural blueprint, then a second model builds the amino acid sequence to fit that blueprint. Why should you care? Protein design has always demanded deep structural biology knowledge. Pinal flattens that barrier. The team also extended it to DNA, introducing Carbon X, an open-weight model that runs on local hardware, so teams without cluster budgets can experiment. The numbers are humbling. Pinal was trained on protein sequences from three databases paired with AlphaFold-predicted structures. The team asked it for four proteins: a fluorescent protein, two enzymes, and a component of a large multi-protein complex. After filtering out sequences resembling known fluorescent proteins, 29 candidates went into E. coli. Only one fluoresced. Carbon X was retrained with DNA data, then reinforcement learning and supervised fine-tuning. Watch for independent replications, because AlphaFold's imperfect predictions could hurt results on obscure proteins.