ProteinMPNN

Proteins are the machines of life. They consist of long chains of building blocks, the amino acids, and fold into a specific three-dimensional shape. This shape determines what a protein does. ProteinMPNN turns the question around: given a desired shape — what sequence of amino acids would produce this shape? The model was published in 2022 by researchers at the Baker Lab at the University of Washington and delivers suggestions in seconds that would have taken a biochemist weeks in the past.

Significance for Medicine and Biotechnology

Designing new proteins is one of the most important tasks in modern biology. Custom-made proteins can act as drugs, improve vaccines, or form new materials. The problem: the classical approach was laborious. Researchers had to adjust amino acid sequences by hand, produce them in the lab, and then test whether they actually folded as desired.

ProteinMPNN drastically shortens this process. It automatically suggests sequences that are highly likely to work. This doesn’t mean lab testing is eliminated — but now one tests ten promising candidates instead of a hundred random hits. In practice, teams using ProteinMPNN have already designed novel proteins that, in experiments, actually adopted the desired structure.

How ProteinMPNN Computes a Sequence

The model receives as input the exact atomic coordinates of a target structure. It treats the protein as a network: each amino acid position is a node, and nearby positions are connected by edges. The model then lets information travel across these edges — hence the name Message Passing, meaning “passing messages along.”

Over several rounds, each node gathers information from its neighbors and updates its own state. In the end, the model predicts for each position which of the twenty possible amino acids fits best there. It doesn’t proceed sequentially, but instead considers the entire three-dimensional environment simultaneously. An important difference from older methods: ProteinMPNN operates directly on the spatial geometry, not on a simplified two-dimensional representation.

A common misunderstanding is confusing ProteinMPNN with AlphaFold. AlphaFold takes the opposite approach: it takes a sequence and predicts what shape it will adopt. ProteinMPNN starts from the desired outcome and designs a matching sequence. Both models complement each other and are often used together in research.

ProteinMPNN in Research and Industry

ProteinMPNN is freely available and is used in university labs worldwide. It is part of a larger toolchain for so-called de novo protein design — that is, designing proteins that do not occur in nature at all. The Baker Lab, which is also behind the RoseTTAFold tool, has published several papers using ProteinMPNN in which entirely novel protein structures were confirmed in the lab.

In the pharmaceutical industry, interest in such models is growing because they accelerate the early phase of drug discovery. Anyone who can more quickly check whether a protein is a viable candidate saves time and money. ProteinMPNN is also already being tested in the field of vaccine development — for instance, for vaccines based on custom-designed protein shells.

For the broader public, ProteinMPNN became particularly visible in the context of the 2024 Nobel Prize. David Baker, head of the Baker Lab, received the Nobel Prize in Chemistry in part for work on the computational design of proteins — to which ProteinMPNN made a significant contribution.