Structure prediction reads a sequence and guesses the fold. Protein design runs the arrow backwards — you name a shape, a function, or even a plain-language brief, and a generative model writes the protein to match. ProteinMPNN, RFdiffusion and Chroma made the inverse problem routine (Chroma alone experimentally characterized 310 designed proteins); ESM3 turned design into prompting a protein language model across sequence, structure and function — and produced esmGFP, a glowing protein only 58% identical to anything nature ever made, which its authors estimate is worth over 500 million years of evolution. The 2024 Nobel sealed the field’s coming-of-age, and the 2025 frontier is designed catalysis: working enzymes drawn from scratch. It’s generative AI meeting matter — and the design→build→test loop, plus the prove-it discipline, the lab keeps betting on.
