Abstract
Water plays a fundamental thermodynamic role in determining protein structure, function, and stability. Rather than merely acting as a passive solvent, water actively participates in entropy-driven free energy changes, which are crucial for stabilizing protein conformations. However, existing deep learning models for predicting protein thermal stability primarily focus on internal geometric and topological features, while neglecting the hierarchical hydration environment and its coupling with protein structure. To address this limitation, we propose a Hydration-Aware Geometric Graph Neural Network (HAGGNN), which explicitly integrates hydration environment into geometric deep learning. HAGGNN introduces a unified Hydration-Geometry Co-Modeling framework that combines invariant and equivariant GNNs. HAGGNN enables the joint capture of geometric dependencies and hydration effects, providing a more comprehensive understanding of protein-water interactions. Experiments on a large-scale protein dataset demonstrate that HAGGNN achieves superior predictive performance compared with models that do not incorporate hydration thermodynamics. Ablation studies further confirm the essential contributions of each module. HAGGNN provides a new computational paradigm that integrates geometric learning with hydration environment, offering mechanistic insights into protein thermal stability prediction.