The problem. Prediction and inverse folding both start from something — a sequence, or a backbone you already drew. Design’s most ambitious version starts from intent: “give me a protein that binds this target” or “that scaffolds this active site,” with no template. Can you generate a plausible new backbone from noise?

The idea. RFdiffusion adapts denoising diffusion — the generative recipe behind modern image models — to protein structure. Fine-tuned from the RoseTTAFold network, it starts from random residue coordinates and iteratively denoises toward a coherent backbone, optionally conditioned on constraints: bind this epitope, obey this symmetry, hold this functional motif fixed. You then hand the backbone to ProteinMPNN for a sequence and to a predictor for a check. The Baker lab validated the outputs experimentally — de novo binders (some at picomolar affinity), symmetric assemblies, and motif scaffolds — not just in silico plausibility.

Why it matters. This is the “DALL·E for proteins” moment, and the mechanism is genuinely the same math you already know from image diffusion — which is the point worth internalizing: the generative-ML toolbox transfers across modalities once you have the right representation and enough validated data. It’s the generative end of DBTL’s Design step, the part most plausibly tooled by software rather than a foundry, and it composes cleanly with the models above into a design → sequence → predict loop.

Verdict. Genuinely exciting and, unusually for a generative-model paper, backed by wet-lab success rather than pretty renders alone. Honest limits: success rates vary by task, hard functional constraints (catalysis, complex dynamics) remain difficult, and every generated design still needs experimental confirmation. Read it last in this batch — it’s where the whole design stack points.