[Paper Review] Predicting mutational effects on protein-protein binding via a side-chain diffusion probabilistic model

Shiwei Liu, Zhu Tian+4 more|arXiv (Cornell University)|Oct 30, 2023

Protein Structure and Dynamics10 citations

TL;DR

SidechainDiff is a diffusion-based model that learns side-chain conformations at protein interfaces to predict mutational effects on binding (ΔΔG), achieving state-of-the-art performance on SKEMPI2 and SARS-CoV-2 datasets, with DiffAffinity leveraging SidechainDiff representations for binding-change prediction and antibody optimization.

ABSTRACT

Many crucial biological processes rely on networks of protein-protein interactions. Predicting the effect of amino acid mutations on protein-protein binding is vital in protein engineering and therapeutic discovery. However, the scarcity of annotated experimental data on binding energy poses a significant challenge for developing computational approaches, particularly deep learning-based methods. In this work, we propose SidechainDiff, a representation learning-based approach that leverages unlabelled experimental protein structures. SidechainDiff utilizes a Riemannian diffusion model to learn the generative process of side-chain conformations and can also give the structural context representations of mutations on the protein-protein interface. Leveraging the learned representations, we achieve state-of-the-art performance in predicting the mutational effects on protein-protein binding. Furthermore, SidechainDiff is the first diffusion-based generative model for side-chains, distinguishing it from prior efforts that have predominantly focused on generating protein backbone structures.

Motivation & Objective

• Motivate accurate prediction of mutational effects on protein-protein binding amid scarce labeled data.
• Develop a diffusion-based generative model for protein side-chains at interfaces.
• Learn structural-context representations of mutations to improve ΔΔG prediction.
• Demonstrate the utility of the learned representations for antibody optimization and SARS-CoV-2-related tasks.

Proposed method

• Introduce SidechainDiff, a conditional Riemannian diffusion model on the 4D torus (T^4) for side-chain rotamers conditioned on SE(3)-invariant structural context.
• Model the forward diffusion with Geodesic Random Walks on T^4 and learn the score function s_theta(X,t,Z) via an MLP, where Z encodes mutation-context.
• Use a conditional encoder (SE(3)-invariant IPA network) to produce Z from per-residue and pairwise structural features.
• Train SidechainDiff with refined PDB-REDO structures following the De Bortoli et al. (2022) implicit loss for score-based diffusion.
• Define DiffAffinity as the downstream predictor that uses SidechainDiff embeddings to predict ΔΔG, via an IPA-like transformer and MLP head.

Experimental results

Research questions

• RQ1 Can a diffusion-based model learn a generative distribution over protein side-chain conformations at interfaces?
• RQ2 Do learned side-chain representations improve accuracy in predicting mutational effects on protein-protein binding (ΔΔG) compared to existing methods?
• RQ3 Are SidechainDiff representations beneficial for downstream tasks such as antibody optimization and SARS-CoV-2 binding predictions?

Key findings

• DiffAffinity achieves state-of-the-art performance on SKEMPI2 across multiple metrics, outperforming energy-based, sequence-based, unsupervised, end-to-end, and pre-trained baselines.
• Across single- and multi-point mutations, DiffAffinity and DiffAffinity* outperform baselines, with DiffAffinity slightly superior to DiffAffinity* due to SidechainDiff representations.
• On SARS-CoV-2 RBD mutations, DiffAffinity shows higher Pearson correlation than competing methods for the 15 significant sites and 285 single-point mutations.
• DiffAffinity improves antibody optimization by ranking favorable mutations, consistently outperforming FoldX, RDE-Net, and DiffAffinity* in top-mutation rankings.
• SidechainDiff demonstrates competitive side-chain rotamer prediction with lower steric clash counts than baselines and favorable MAE relative to deep-learning methods.