Dynamics-inspired Structure Hallucination for Protein-protein Interaction Modeling

TL;DR

This work tackles predicting mutation effects on protein-protein binding by addressing two bottlenecks: the absence of mutant complex structures and the dynamic nature of PPIs. It introduces Refine-PPI, which jointly refines unseen mutant structures through Mask Mutation Modeling and predicts $\Delta\Delta G$ via a shared encoder-predictor framework, coupled with the novel Probability Density Cloud Network (PDC-Net) to capture conformational dynamics and atomic uncertainty. The approach achieves state-of-the-art correlation metrics on SKEMPI.v2 and strong zero-shot performance on ProteinGym, while providing meaningful uncertainty estimates that align with MD-derived fluctuations. These contributions offer a robust pathway for structure-aware, dynamic mutation effect prediction in PPIs, with potential impact on antibody engineering and drug design.

Abstract

Protein-protein interaction (PPI) represents a central challenge within the biology field, and accurately predicting the consequences of mutations in this context is crucial for drug design and protein engineering. Deep learning (DL) has shown promise in forecasting the effects of such mutations, but is hindered by two primary constraints. First, the structures of mutant proteins are often elusive to acquire. Secondly, PPI takes place dynamically, which is rarely integrated into the DL architecture design. To address these obstacles, we present a novel framework named Refine-PPI with two key enhancements. First, we introduce a structure refinement module trained by a mask mutation modeling (MMM) task on available wild-type structures, which is then transferred to produce the inaccessible mutant structures. Second, we employ a new kind of geometric network, called the probability density cloud network (PDC-Net), to capture 3D dynamic variations and encode the atomic uncertainty associated with PPI. Comprehensive experiments on SKEMPI.v2 substantiate the superiority of Refine-PPI over all existing tools for predicting free energy change. These findings underscore the effectiveness of our hallucination strategy and the PDC module in addressing the absence of mutant protein structure and modeling geometric uncertainty.

Figures (9)

• Figure 1: Geometric deep learning is applied to optimize the antibody sequences and achieve desired properties (e.g., better affinity and specificity).
• Figure 2: Performance of Refine-PPI on SKEMPI.v2 compared to other energy-based or pretrained baselines.
• Figure 3: A. The overall pipeline of our Refine-PPI. The given wild-type structure and the masked mutant structure are subsequently fed into weight-shared equivariant neural networks. The masked region is reconstructed, and the mutation effect is predicted by comparing the features of the two resulting complexes. B. The procedure of the deep learning architecture. The particles in the complex are represented as probability density clouds (PDCs), where each atom moves according to some geometric distributions instead of being immobile. Then, the natural parameters, including mean, variance, and covariance, are updated and propagated throughout PDC-Network.
• Figure 4: A. Visualization of correlations between experimental $\Delta\Delta G$ and predicted $\Delta\Delta G$. B. The scatter plot shows that the recovery error of the wild-type structure has a positive relation with the error of $\Delta\Delta G$ prediction.
• Figure 5: Visualization of learned uncertainty. A darker color corresponds to a more flexible protein segment.