Abstract

The complex geometric structure of proteins is key to their function and specificity. Sixty years ago, designing proteins seemed nearly impossible, but today we can create fully synthetic proteins.

With the rapid expansion of structural databases, deep learning-based protein design methods are gaining attention. But can these generative models produce “evolved” samples that follow physical and chemical rules, fold correctly, remain stable, and offer diversity and novelty?

We used Score Matching and Flow Matching methods to train generative models for monomeric protein backbone structures on the SE(3)-invariant Riemannian manifold, and tested these models on different monomeric proteins, including cytochrome c, green fluorescent protein, monoglucosidase, and $\beta$-lactamase, to see how well they perform.