The problem. Structure prediction runs sequence → structure. Protein design needs the inverse: you have a backbone you want (a fold, a binder, a scaffold), and you need an amino-acid sequence that will actually fold to it. Physics-based design (Rosetta) worked but was slow and often failed.

The idea. ProteinMPNN is a message-passing graph neural network over the backbone: each residue is a node, geometric relationships are edges, and the model autoregressively predicts amino acids conditioned on the structure and the residues already chosen. It’s fast, order-agnostic in its decoding, and can tie or fix positions for symmetry and constraints. The headline numbers: native sequence recovery around 52% versus ~33% for Rosetta, and — more tellingly — it experimentally rescues designs that Rosetta and even AlphaFold-guided pipelines couldn’t get to fold, often just by redesigning the sequence for an existing backbone.

Why it matters. This is the “Learn”-adjacent design tool that pairs with generative backbone models: RFdiffusion (or another method) proposes a shape, ProteinMPNN dresses it in a sequence, and a predictor checks it. It’s the concrete counterpart to AlphaFold — same coordinate-and-graph world, opposite direction — and it’s cheap enough to run at scale, which is exactly the regime a computational shop can own without a wet lab.

Verdict. Quietly one of the most useful papers in the modern design stack — not because the architecture is exotic (it isn’t) but because it’s fast, robust, and validated in the lab, not just on held-out recovery. The caveat is scope: it designs sequences for a given backbone; it doesn’t decide what backbone you should want. Read it as one half of a pair with RFdiffusion.