ProteinMPNN · AIxSci

What it is

Proteins are long chains of small building blocks called amino acids. The chain folds into a particular shape, and the shape decides what the protein does. The path the chain traces through space is called the backbone. Working out the shape from the sequence is one problem; going the other way, from a wanted shape to a sequence that will adopt it, is called sequence design. ProteinMPNN is a network trained on known protein structures to do the second task. Given a backbone, it proposes amino acids position by position, using what it has learned about which residues sit comfortably in which local surroundings.

The network describes the backbone as a graph: each residue is a point, and the links between nearby points carry numbers for the distances between their backbone atoms, how far apart they are along the chain, and whether they belong to the same chain. It never sees absolute coordinates, so a structure turned or shifted in space looks the same to it. Because the method is fast and the network can be asked repeatedly with different random choices, designers can generate many candidate sequences for one shape and then test them, often by predicting each candidate's structure and keeping those that fold back to the target.

Why scientists use it

Designing a protein usually starts with a shape rather than a sequence. A researcher may want a scaffold with a particular fold, a surface that grips another molecule, or a repeating architecture. The number of possible sequences for even a short chain is far beyond anything that could be tried one at a time, so some way of narrowing the field is needed. Older approaches searched through amino acid choices while scoring their physical energies, which is slow and sensitive to how the energy is written down. A trained network replaces that search with a learned guess about what real proteins look like at each position.

Because it gives a quick answer for any backbone, ProteinMPNN also serves as a general tool inside larger pipelines. It can fill in sequences for shapes invented by other programs, supply plausible sequences to be folded and scored, or act as a reference point when a new method for judging mutations or designing binders is being assessed.

In the record

structural-biologyNeural network writes amino acid sequences to fit fixed protein backbonesTrained on protein assemblies clustered by sequence identity, with backbones given only as graph edges: interatomic distances, a capped positional encoding, and a same-chain indicator.structural-biologyTwo protein design models redesigned T cell receptor contact points on solved structuresUsed in designing T cell receptors, with the test set of solved receptor complexes filtered to exclude structures from before the network's training cut-off date.structural-biologyRepeating-coil proteins designed from random sequences using AlphaFold2 in an evolution loopApplied in work designing repeating solenoid proteins, where candidate sequences were optimised against a structure prediction and a solenoid classifier.structural-biologyNeural networks trained on atomic surroundings score how mutations change proteinsNamed alongside a model that encodes each residue's surrounding atoms on a Zernike basis to predict mutational effects and stability.

Limits

The network learns from structures that have already been solved, so it reflects the proteins nature and laboratories have supplied, and it may do less well on shapes unlike those. It proposes sequences that look plausible for a backbone; it does not show that a designed protein will fold, stay soluble, survive in a cell or do anything useful, and those questions are settled by prediction and experiment afterwards. Scoring a method against structures the network saw during training flatters it, which is why such structures are excluded in careful tests.

Written by AIxSci: the background from general knowledge, the examples from checked records.