Diffusion models for protein design

Generative deep learning has transformed protein design, enabling the creation of new protein structures, interactions, and functions with high precision, by using generative models such as diffusion models and flow matching. Geometric deep learning plays a crucial role as it enables to model symmetries to optimize model expressiveness and efficiency.

In this HITS Lab project, we aim at improving current generative models for protein design by revisiting the way equivariance is built into them and incorporating these insights into the generative process. We are specifically tackling the design of a protein with a trivalent crosslink—a complex chemical bond among three amino acids, a structure not yet successfully synthesized but known to exist in collagen. By addressing this challenge, we aim to both enhance diffusion models and deepen our understanding of this intricate biochemical process.

Contributors:

Frauke Gräter (MBM)

Jan Stühmer (MLI)

Members:

Leif Seute

Vsevolod Viliuga

Simon Wagner

Nicolas Wolf

Publications:
Seute et al.: Grappa — A Machine Learned Molecular Mechanics Force Field
https://arxiv.org/abs/2404.00050

Wagner S, Seute L, Viliuga V, Wolf N, Gräter F, Stühmer J: Generating Highly Designable Proteins with Geometric Algebra Flow Matching.
The Thirty-eighth Annual Conference on Neural Information Processing Systems (NeurIPS) 2024 (accepted).