Score-Based Generative Models for Designing Binding Peptide Backbones

John D. Boom, Matthew Greenig, Pietro Sormanni, Píetro LióarXiv 2023 · DOI · arXiv

Abstract

Score-based generative models (SGMs) have proven to be powerful tools for designing new proteins. Designing proteins that bind a pre-specified target is highly relevant to a range of medical and industrial applications. Despite the flurry of new SGMs in the last year, there has been little systematic exploration of the impact of design choices in SGMs for protein design. Here we present LoopGen, a flexible SGM framework for the design of short binding peptide structures. We apply our framework to design antibody binding loop structures conditional on a target epitope and evaluate a variety of modelling choices in SGM-based protein design. We demonstrate that modelling residue orientations in addition to positions improves not only the quality of the output structures but also their diversity. Additionally, we identify variance schedules that result in significant performance improvements and observe patterns that may motivate the development of better schedules for protein design. Finally, we develop three novel tests to evaluate whether the model generates structures that are appropriately conditioned on an epitope, demonstrating that LoopGen's generated structures are dependent on the structure, sequence, and position of the epitope. Our findings will help guide future development and evaluation of generative models for binding proteins.

Introduction

In protein design, creating a "binder"—a molecule that adheres specifically to a target—is a fundamental challenge for drug discovery. Most current AI methods utilize a two-step process: generating a 3D "backbone" skeleton followed by "inverse folding" to determine the amino acid sequence. However, evaluating these backbones is difficult, as standard metrics like designability often fail for short, flexible peptides like antibody complementarity-determining regions (CDRs). Furthermore, many models focus solely on $C\alpha$ atom positions, neglecting residue orientations. This study introduces LoopGen to investigate how structural representation and noise schedules influence the quality and diversity of designed binders.

Methodology

LoopGen is a flexible SGM framework designed for conditional generation of peptide loops. Key technical innovations include:

  • $SE(3)$ Manifold Modeling: LoopGen can represent residues as "frames," incorporating both 3D position (translation) and orientation (rotation) as rigid bodies.
  • Geometric Vector Perceptrons (GVP): The core architecture uses a GVP-based Graph Neural Network to reason about geometric tensors and spatial relationships.
  • Diffusion on Rotations: The model utilizes the $IGSO(3)$ distribution to apply and reverse noise on 3D rotations, enabling diffusion on the full $SE(3)$ manifold.
  • Target Conditioning: Generation is specifically conditioned on a target epitope, ensuring the resulting structure is contextually relevant.

Results and Discussion

The Impact of Structural Representation

A central finding of this work is that modeling residue orientations is essential for structural diversity. While $C\alpha$-only models might achieve similar error rates relative to ground truth, they often produce "homogeneous" structures that fail to explore the natural shape space of binding loops.

Comparison of Frame-based vs. Coordinate-only models across three metrics: accuracy to ground truth (left), physical validity via bond lengths (middle), and structural diversity (right).

As shown in the figure above, frame-based models significantly outperform coordinate-only models in diversity (mean pairwise RMSD) and physical validity (adjacent $C\alpha$ distance), despite having similar mean RMSDs to the ground truth. This highlights the "RMSD Trap," where low error does not necessarily equate to a high-quality, diverse ensemble of structures.

Physical Plausibility and Noise Schedules

The choice of noise schedule—how noise is added and removed during training—is critical. The authors found that a quadratic schedule for translations and a logarithmic schedule for rotations produced the most physically realistic structures. With optimal scheduling, over 90% of generated structures were physically plausible, exhibiting valid bond lengths and minimal atomic clashes.

Ramachandran plot showing the distribution of dihedral angles (phi, psi) for real vs. generated structures, confirming physical plausibility.

The Ramachandran plot further validates these results, showing that the generated backbone geometries closely mirror the distribution of dihedral angles found in real protein structures.

Epitope Conditioning

To ensure the model was not simply "hallucinating" random loops, the researchers performed perturbation tests. By scrambling the epitope sequence or shifting its position, they observed measurable changes in the generated structures. This confirms that LoopGen is truly responsive to the geometry and chemistry of the target epitope.

Conclusion

LoopGen demonstrates that modeling full residue orientations and optimizing variance schedules are vital for generating high-quality, diverse binding peptide backbones. By moving beyond simple coordinate-based diffusion and implementing rigorous conditioning tests, this work provides a roadmap for the future development of generative models in therapeutic protein design.

Summary updated

Abstract

Score-based generative models (SGMs) have proven to be powerful tools for designing new proteins. Designing proteins that bind a pre-specified target is highly relevant to a range of medical and industrial applications. Despite the flurry of new SGMs in the last year, there has been little systematic exploration of the impact of design choices in SGMs for protein design. Here we present LoopGen, a flexible SGM framework for the design of short binding peptide structures. We apply our framework to design antibody binding loop structures conditional on a target epitope and evaluate a variety of modelling choices in SGM-based protein design. We demonstrate that modelling residue orientations in addition to positions improves not only the quality of the output structures but also their diversity. Additionally, we identify variance schedules that result in significant performance improvements and observe patterns that may motivate the development of better schedules for protein design. Finally, we develop three novel tests to evaluate whether the model generates structures that are appropriately conditioned on an epitope, demonstrating that LoopGen's generated structures are dependent on the structure, sequence, and position of the epitope. Our findings will help guide future development and evaluation of generative models for binding proteins.
BibTeX
@misc{boom2023score,
  author = {John D. Boom and Matthew Greenig and Pietro Sormanni and Píetro Lió},
  title = {{Score-Based Generative Models for Designing Binding Peptide Backbones}},
  year = {2023},
  eprint = {2310.07051},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2310.07051},
}
Figure