ProteinReDiff: Complex-based ligand-binding proteins redesign by equivariant diffusion-based generative models

Viet Thanh Duy Nguyen, Nhan D. Nguyen, Truong Son HyStructural 2024 · DOI

TL;DR

This paper introduces ProteinReDiff, a novel diffusion-based generative model for redesigning ligand-binding proteins. It achieves high-affinity protein design by leveraging protein sequences and ligand SMILES strings, bypassing the need for detailed structural information, and integrating key architectural components from AlphaFold2.

What Problem Does This Paper Address?

Designing proteins that can bind to specific molecules (ligands) is crucial for drug development and various biotechnological applications. Traditional methods often rely on extensive experimental trial-and-error or computationally intensive simulations that require detailed structural information of the protein and its binding site. This reliance on structural data can be a bottleneck, especially when dealing with novel targets or when detailed structural information is unavailable. Furthermore, many existing computational approaches focus solely on protein or ligand design independently, neglecting the synergistic potential of co-designing both for optimal binding. This paper aims to address these limitations by developing a computational framework that can efficiently redesign ligand-binding proteins using only sequence and ligand information, while also considering both protein and ligand aspects simultaneously.

Methods

ProteinReDiff is built upon an equivariant diffusion-based generative model. The core innovation lies in integrating architectural components from AlphaFold2 (AF2) into a diffusion framework. Specifically, it incorporates:

  • Input Featurization: Proteins are represented using masked amino acid sequences, with features derived from the ESM-2 language model. Ligands are represented by their SMILES strings, with comprehensive atomic and bond property embeddings.
  • Residual Feature Updates: This module adapts and enhances components from AF2's Evoformer, including:
  • Single Representation Attention (SRA): Captures long-range interactions within the protein-ligand complex.
  • Outer Product Update (OPU): Merges single and pair representations to generate plausible structural hypotheses.
  • Triangle Multiplicative Updates: Refines pair representations by accounting for physical and geometric constraints.
  • Equivariant Denoising: The model iteratively denoises the protein-ligand complex representation, ensuring that the generated structures are equivariant to rotations and translations (SE(3) equivariance). This process allows for the generation of novel protein sequences that are predicted to bind to the given ligand.

The framework also incorporates a joint sequence and structural diffusion module and a blind docking capability, meaning it doesn't require pre-defined binding pockets.

Datasets

The study utilizes the following datasets:

  • PDBBind v2020: A comprehensive database of protein-ligand complexes with experimentally determined binding affinities. This dataset was used for training, validation, and testing, with specific filtering applied to maintain sequence similarities between 40% and 50%.
  • CATH 4.2: A database of protein structures. This dataset was used for training and validation, focusing on smaller proteins (<400 amino acids) with less similar sequences (<90%). For CATH 4.2, ligand SMILES strings were represented as masked tokens, as actual ligands were not always specified. CATH 4.2 was excluded from the test set due to the lack of corresponding ligands.

Key Insights and Ideas

  • Bypassing Structural Information: The most significant insight is the successful demonstration that high-affinity ligand-binding proteins can be designed without requiring detailed 3D structural information of the protein or its binding site. This opens up new avenues for designing binders for targets where structural data is scarce or difficult to obtain.
  • Synergistic Protein-Ligand Design: By treating protein and ligand as a complex from the outset and using their respective representations, ProteinReDiff implicitly co-designs both components for optimal interaction, rather than optimizing them in isolation.
  • Leveraging AlphaFold2 Architecture: The integration of AF2's architectural components (SRA, OPU, Triangle Multiplicative Updates) into a diffusion model proves highly effective for capturing intricate protein-ligand interactions. This suggests that these architectural motifs are generalizable beyond pure structure prediction.
  • Optimal Masking Ratio: The study highlights that a specific masking ratio (around 15%) is crucial for achieving the best balance between enhancing ligand binding affinity, maintaining sequence diversity, and preserving structural integrity. Extreme masking ratios lead to suboptimal performance in one or more of these aspects.

Key Results

  • Enhanced Ligand Binding Affinity (LBA): ProteinReDiff, particularly with a 15% masking ratio, significantly improves ligand binding affinity compared to baseline methods and the original protein designs. The best LBA achieved was -6.803 ± 0.329 kcal/mol.
  • Competitive Sequence Diversity and Structure Preservation: The model demonstrates strong performance in sequence diversity (up to 190.425) and structure preservation metrics (TM Score up to 0.864 ± 0.022, CO up to 0.942 ± 0.007), indicating that it can generate novel sequences while maintaining the core structural integrity of the proteins.
  • Robustness to Ablations: Ablation studies showed that removing key components like SRA or OPU modules significantly degraded performance, underscoring their importance. Excluding ligand features also led to a reduction in performance, emphasizing the need for integrated protein-ligand representations.
  • Outperforming Baselines: ProteinReDiff generally outperformed established methods like CARP, MIF, ESMIF, and even the original DPL model in key metrics, especially ligand binding affinity.

Limitations and Future Work

  • Masking Ratio Optimization: While 15% masking showed optimal results, the study suggests that further exploration of masking strategies could lead to even better performance.
  • Generalizability: While the model shows promise, further validation on a wider range of protein-ligand complexes and different types of interactions would be beneficial.
  • Computational Cost: Although designed to be more efficient than traditional methods, diffusion models can still be computationally intensive. Future work could focus on further optimizing the model's efficiency.
  • Experimental Validation: The study relies on computational predictions. Experimental validation of the designed proteins is the crucial next step to confirm their efficacy in real-world applications.
  • Ligand Representation: While effective, exploring more sophisticated ligand representations could potentially further enhance performance.

Summary updated

Abstract

Proteins, serving as the fundamental architects of biological processes, interact with ligands to perform a myriad of functions essential for life. Designing functional ligand-binding proteins is pivotal for advancing drug development and enhancing therapeutic efficacy. In this study, we introduce ProteinReDiff, an diffusion framework targeting the redesign of ligand-binding proteins. Using equivariant diffusion-based generative models, ProteinReDiff enables the creation of high-affinity ligand-binding proteins without the need for detailed structural information, leveraging instead the potential of initial protein sequences and ligand SMILES strings. Our evaluations across sequence diversity, structural preservation, and ligand binding affinity underscore ProteinReDiff's potential to advance computational drug discovery and protein engineering.
BibTeX
@inproceedings{nguyen2024proteinrediff,
  author = {Viet Thanh Duy Nguyen and Nhan D. Nguyen and Truong Son Hy},
  title = {{ProteinReDiff: Complex-based ligand-binding proteins redesign by equivariant diffusion-based generative models}},
  booktitle = {Structural},
  year = {2024},
  doi = {10.1063/4.0000271},
}
Figure