Generating Novel, Designable, and Diverse Protein Structures by Equivariantly Diffusing Oriented Residue Clouds

TL;DR

The paper tackles de novo protein design by introducing Genie, a diffusion-based model that operates on Cartesian Cα coordinates while employing SE(3)-equivariant reasoning over oriented residue frames. By combining a forward diffusion process with a frame-aware, IPA-enabled reverse denoiser, Genie achieves high designability, diversity, and novelty, outperforming short-model baselines and offering competitive results against larger long-model diffusion methods. The work demonstrates Genie's ability to generate designable structures across a range of lengths, visualize design space, and highlight the trade-offs between designability, diversity, and sampling speed. It also outlines directions for scaling, conditional generation, and integrating functional constraints to further enhance de novo protein design workflows.

Abstract

Proteins power a vast array of functional processes in living cells. The capability to create new proteins with designed structures and functions would thus enable the engineering of cellular behavior and development of protein-based therapeutics and materials. Structure-based protein design aims to find structures that are designable (can be realized by a protein sequence), novel (have dissimilar geometry from natural proteins), and diverse (span a wide range of geometries). While advances in protein structure prediction have made it possible to predict structures of novel protein sequences, the combinatorially large space of sequences and structures limits the practicality of search-based methods. Generative models provide a compelling alternative, by implicitly learning the low-dimensional structure of complex data distributions. Here, we leverage recent advances in denoising diffusion probabilistic models and equivariant neural networks to develop Genie, a generative model of protein structures that performs discrete-time diffusion using a cloud of oriented reference frames in 3D space. Through in silico evaluations, we demonstrate that Genie generates protein backbones that are more designable, novel, and diverse than existing models. This indicates that Genie is capturing key aspects of the distribution of protein structure space and facilitates protein design with high success rates. Code for generating new proteins and training new versions of Genie is available at https://github.com/aqlaboratory/genie.

Figures (18)

• Figure 1: Architecture of SE(3)-equivariant denoiser, including SE(3)-invariant encoder (bottom left) and SE(3)-equivariant decoder (bottom right). Notation: $r$: number of residues, $c_s$: dimensionality of single representation, $c_p$: dimensionality of pair representation.
• Figure 2: Analysis of structures generated by Genie and short models. (A) Heatmap of the relative frequencies of generated domains with specific combinations of highest scTM and pLDDT values achieved by ProtDiff, FoldingDiff, and Genie. (B) Heatmap of relative frequencies of confidently designable domains with specific combinations of fractional SSE content. The number of designed domains for each model is shown in parentheses. Heatmap of relative frequencies of the SCOPe dataset is provided in Figure \ref{['appendix_results']}C (Appendix \ref{['appendix_c1']}) for reference. (C) Histogram of confidently designable domains as a function of sequence length. (D) Bar chart of number of designable domains generated by different methods out of a fixed budget of 780 attempted designs per method.
• Figure 3: Heatmap of relative frequencies of confidently designable structures with specific combinations of fractional SSE content for Genie and long models.
• Figure 4: Design space of Genie-SwissProt. 818 confidently designable structures were embedded in 2D space using multidimensional scaling (MDS) with pairwise TM scores as the distance metric. Domains are colored by their maximum TM score to PDB structures (central panel), fraction of helical residues (top left panel), fraction of beta strand residues (middle left panel), and sequence length (bottom left panel). Eight novel designed domains are shown as representatives.
• Figure 5: Diffusion of protein backbone in Cartesian space.