Wiki · Bio × AI Methods · concept
Protein design
Building new proteins that don't exist in nature to do a specific job — the inverse of structure prediction.
Protein design is the inverse of structure prediction: instead of asking what shape a given sequence folds into, you start from the job you want done and invent a protein to do it. Nature spent billions of years evolving a tiny fraction of the proteins that are chemically possible. Design goes off that map — a molecule that grips a specific drug target, catalyzes a reaction no natural enzyme runs, or props up a vaccine — a protein that has never existed, built to specification.
How it works
Modern de novo design runs as a pipeline that mirrors and reuses the structure- prediction stack. First, generate a backbone: a diffusion model such as RFdiffusion (from the Baker lab, built on the RoseTTAFold network) starts from noise and denoises its way to a plausible 3D fold shaped for the task — for a binder, one that clamps onto the target surface. That backbone is just a shape with no sequence, so next comes sequence design: ProteinMPNN reads the geometry and picks the amino acids most likely to actually fold into it. Finally, validate in silico: run the designed sequence through AlphaFold and check that its predicted structure snaps back to the intended backbone. It’s a generate → predict → filter loop — propose many candidates, keep the few whose predictions agree, then take those to the lab.
Why it matters (for bio × AI)
This is biology’s generative frontier. Rather than discovering a binder or enzyme by screening what already exists, you design one to order — custom therapeutics, biosensors, and industrial catalysts drawn up rather than found. The Baker-lab lineage turned this from a curiosity into a working discipline, and it is the clearest case of generative AI making genuinely new molecules, not just describing old ones.
Related
See diffusion model for the generator behind the backbones, AlphaFold for the predictor that validates them, protein language model for the sequence-level view of the same design problem, and molecular docking for the complementary small-molecule side of structure-based design.