PDC-Net: Probability Density Cloud Representations of Proteins for Mutation Effect Prediction

Abstract

Understanding the ramifications of mutations at a protein level can have significant implications in various domains such as drug development, disease pathways, and the broader field of genomics. Despite the promise of data-driven and deep learning (DL) strategies, existing algorithms still face a significant challenge in integrating the dynamic changes of biomolecules to accurately predict protein-protein interaction binding affinity changes following mutations ($\Delta \Delta G$). Within this study, we introduce an inventive approach aimed at capturing the equilibrium fluctuations and discerning induced conformational changes at the interface, which is particularly important for forecasting mutational effects on binding. This novel technique harnesses probability density clouds (PDC) to describe the magnitude and intensity of their movement during and after the binding process and puts forth aligned networks to propagate distributions of the equilibrium of molecular systems. To fully unleash the potential of PDC-Net, we further present two physics-inspired pretraining tasks to employ the molecular dynamics (MD) simulation trajectories and the extensive collection of static crystal protein structures. Experiments demonstrate that our approach surpasses the performance of both empirical energy functions and alternative DL methods.