[Paper Review] Out of Many, One: Designing and Scaffolding Proteins at the Scale of the Structural Universe with Genie 2

Yeqing Lin, Minji Lee+2 more|arXiv (Cornell University)|May 24, 2024

Genetics, Bioinformatics, and Biomedical Research11 citations

TL;DR

Genie 2 extends Genie to support single- and multi-motif protein scaffolding, achieves state-of-the-art designability, diversity, and novelty, and scales to larger structure spaces by using motif conditioning and large-scale augmentation with AlphaFold data.

ABSTRACT

Protein diffusion models have emerged as a promising approach for protein design. One such pioneering model is Genie, a method that asymmetrically represents protein structures during the forward and backward processes, using simple Gaussian noising for the former and expressive SE(3)-equivariant attention for the latter. In this work we introduce Genie 2, extending Genie to capture a larger and more diverse protein structure space through architectural innovations and massive data augmentation. Genie 2 adds motif scaffolding capabilities via a novel multi-motif framework that designs co-occurring motifs with unspecified inter-motif positions and orientations. This makes possible complex protein designs that engage multiple interaction partners and perform multiple functions. On both unconditional and conditional generation, Genie 2 achieves state-of-the-art performance, outperforming all known methods on key design metrics including designability, diversity, and novelty. Genie 2 also solves more motif scaffolding problems than other methods and does so with more unique and varied solutions. Taken together, these advances set a new standard for structure-based protein design. Genie 2 inference and training code, as well as model weights, are freely available at: https://github.com/aqlaboratory/genie2.

Motivation & Objective

• Expand the design space of protein structures by enabling motif scaffolding with unspecified inter-motif geometry.
• Improve unconditional protein generation quality to achieve higher designability, diversity, and novelty.
• Enable single- and multi-motif scaffolding to design proteins with multiple functional sites or interaction partners.
• Leverage large-scale augmentation from AlphaFold databases to broaden observed structure space.

Proposed method

• Use Genie’s diffusion framework with an SE(3)-invariant encoder and SE(3)-equivariant decoder to operate on backbone coordinates and reference frames.
• Introduce a multi-motif framework that allows conditioning on motif structures while leaving inter-motif geometry unspecified.
• Represent motif information as SE(3)-invariant pairwise distance matrices and one-hot residue encodings, enabling flexible conditioning.
• Train exclusively on conditional motif scaffolding tasks, with a loss that softly enforces motif constraints via a combined motif-and-scaffold objective.
• Augment training data with AlphaFold Database (AFDB) predictions clustered by FoldSeek and filtered by confidence (pLDDT>80) and length ≤ 256.
• Optimize the model with a denoising objective comparing predicted noise to ground truth noise across diffusion steps.

Figure 1: Genie 2 architecture (top), which extends Genie to enable scaffolding on (multiple) motifs. It consists of an SE(3)-invariant encoder that transforms input features into single residue and pair residue-residue representations, and an SE(3)-equivariant decoder that updates frames based on s

Experimental results

Research questions

• RQ1 Can Genie 2 outperform existing diffusion models on unconditional protein generation in terms of designability, diversity, and novelty?
• RQ2 How effectively can Genie 2 perform motif scaffolding, including multi-motif scenarios with unspecified inter-motif geometry?
• RQ3 What is the impact of large-scale AFDB data augmentation on the quality and diversity of generated structures?
• RQ4 How does Genie 2 scale to longer sequence lengths beyond training limits in unconditional and motif-scaffolding tasks?

Key findings

• Genie 2 matches or exceeds designability and achieves substantially higher diversity and novelty than Chroma and RFDiffusion on unconditional generation.
• Genie 2 outperforms competing methods on motif scaffolding, solving more tasks and yielding more unique designs, especially as sample size grows.
• Genie 2 enables single- and multi-motif scaffolding with under-specified inter-motif geometry, generating complex designs with multiple functional motifs.
• Training with AFDB augmentation expands the observed structure space and improves performance relative to training on PDB alone.
• Genie 2 maintains competitive performance across sequence lengths up to 500 residues despite being trained on ≤256 residue proteins.

Figure 2: Visualizations of in-distribution performance on unconditional generation. (A) Secondary structure distributions of proteins generated by Chroma, RFDiffusion and Genie 2. For reference, we also include the secondary structure distribution of 1,000 structures randomly drawn from AFDB (far r