Learning pairwise energies to improve protein design

What is it about?

A common approach to designing proteins is to generate an amino acid sequence that folds into a desired shape. Our AI model, PottsMPNN, learns the energy of interactions between pairs of amino acids in a structure, not just the preferences at each position on its own, as previous state-of-the-art models do. We also trained it on families of related natural sequences that share a shape, rather than on a single sequence from nature. PottsMPNN designs sequences that fold reliably into their target shapes and predicts how mutations affect stability more accurately. Because its predictions break down into physically meaningful interactions, researchers can see why a sequence is favored.

Featured Image

Photo by Anirudh on Unsplash

Photo by Anirudh on Unsplash

Why is it important?

Sequence design AI models sit inside almost every modern protein engineering pipeline. Nearly all of them are trained to reproduce the exact sequence that nature paired with a known structure, a task called native sequence recovery, even though many different sequences fold into the same shape. We find that the models scoring highest on this task are not the best at the two applications that matter most to protein designers: generating sequences that fold as intended, and predicting the effects of mutations. By learning pairwise interactions and drawing on evolutionary information, our model, PottsMPNN, recovers fewer native amino acids than others yet performs better at both applications.

Perspectives