[Paper Review] Full-Atom Peptide Design with Geometric Latent Diffusion

Xiangzhe Kong, Jia, Yinjun+2 more|arXiv (Cornell University)|Feb 21, 2024

Machine Learning in Materials ScienceMaterials Science3 citations

TL;DR

This paper introduces PepGLAD, a geometric latent diffusion model for end-to-end full-atom peptide design conditioned on target binding sites. It addresses full-atom geometry and variable binding geometry via a variational autoencoder with fixed-size latent representations and a receptor-specific affine transformation to standardize binding site geometry, achieving 18% higher diversity, 8% higher in silico success rate, and 26% improved recall of reference binding conformations.

ABSTRACT

Peptide design plays a pivotal role in therapeutics, allowing brand new possibility to leverage target binding sites that are previously undruggable. Most existing methods are either inefficient or only concerned with the target-agnostic design of 1D sequences. In this paper, we propose a generative model for full-atom extbf{Pep}tide design with extbf{G}eometric extbf{LA}tent extbf{D}iffusion (PepGLAD) given the binding site. We first establish a benchmark consisting of both 1D sequences and 3D structures from Protein Data Bank (PDB) and literature for systematic evaluation. We then identify two major challenges of leveraging current diffusion-based models for peptide design: the full-atom geometry and the variable binding geometry. To tackle the first challenge, PepGLAD derives a variational autoencoder that first encodes full-atom residues of variable size into fixed-dimensional latent representations, and then decodes back to the residue space after conducting the diffusion process in the latent space. For the second issue, PepGLAD explores a receptor-specific affine transformation to convert the 3D coordinates into a shared standard space, enabling better generalization ability across different binding shapes. Experimental Results show that our method not only improves diversity and binding affinity significantly in the task of sequence-structure co-design, but also excels at recovering reference structures for binding conformation generation.

Motivation & Objective

• To enable end-to-end co-design of 1D peptide sequences and 3D full-atom structures conditioned on target binding sites.
• To address the challenge of variable-size full-atom residues in diffusion models by learning fixed-dimensional latent representations via a VAE.
• To improve generalization across diverse binding geometries by transforming 3D coordinates into a shared standard space using receptor-specific affine transformations.
• To construct a new training dataset combining PDB and literature data for full-atom peptide design.
• To achieve higher diversity, binding affinity, and conformational accuracy in peptide generation compared to existing methods.

Proposed method

• A variational autoencoder (VAE) is trained to encode variable-sized full-atom residues into fixed-dimensional latent vectors containing 3D coordinates and hidden features, enabling diffusion in a consistent latent space.
• The VAE encoder and decoder are specifically designed to handle full-atom input and output, preserving atomic-level interactions during generation.
• A receptor-specific affine transformation is derived from the binding site’s center offset and Cholesky decomposition of its covariance matrix, mapping data coordinates to a standard Gaussian space.
• The diffusion process is performed in the shared standard space, improving generalization across diverse binding site geometries and enabling transfer learning to unseen shapes.
• The model conditions both the VAE and diffusion process on the receptor’s 3D structure to guide sequence and structure generation toward stable, high-affinity interactions.
• The framework integrates with Rosetta for energy scoring and binding affinity validation, using PyRosetta to compute dG_separated values for binding success.

Experimental results

Research questions

• RQ1 Can a latent diffusion model effectively generate full-atom peptide structures while preserving atomic-level interactions?
• RQ2 How can variable-sized full-atom residues be modeled within a fixed-size diffusion framework?
• RQ3 Can a receptor-specific affine transformation improve generalization across diverse protein-peptide binding geometries?
• RQ4 To what extent does the proposed method improve diversity and in silico binding success compared to existing baselines?
• RQ5 Can the model accurately recall the reference binding conformation of known peptide-receptor complexes?

Key findings

• PepGLAD improves sequence-structure co-design diversity by 18% compared to baseline methods.
• The in silico success rate for binding affinity prediction increases by 8% in sequence-structure co-design.
• The model achieves a 26% absolute improvement in recalling the reference binding conformation of known complexes.
• The use of receptor-specific affine transformation enables better generalization across diverse binding site geometries, even for unseen shapes.
• The VAE-based latent representation successfully preserves full-atom geometry during diffusion, supporting high-fidelity structure generation.
• The model outperforms strong baselines including RFDiffusion, FlexPepDock, AlphaFold2, and HSRN in both diversity and binding conformation recall.