Context-Aware Hydrophobicity Modeling: HydroMap and FastHydroMap.

Hydrophobicity governs a vast range of phenomena, from protein-protein interactions to nanomaterial assembly, and can be rigorously quantified by the dewetting free energy (Fdewet) of a molecule or surface. However, hydrophobicity remains widely treated as an additive property of amino acid identity, obscuring the fact that water's response is a collective property of the surface, shaped by curvature, chemical patterning, and neighboring residues. Direct calculation of Fdewet via specialized molecular simulations captures this collective behavior but is prohibitively slow, leaving in place broadly-used, decades-old sequence-based hydropathy scales that neglect the physics of solvation. Here we show that residue-level Fdewet can be predicted from a compact set of local water features (water structural signatures and residue-water potential energy) extracted from a brief and inexpensive all-atom simulation. We embed this insight in two models: HydroMap, which predicts Fdewet directly from water features, and FastHydroMap, a computationally inexpensive graph neural network surrogate trained on HydroMap that requires no solvent simulation. HydroMap and FastHydroMap capture context-dependent hydrophobicity that classical, sequence-only hydropathy scales miss. We demonstrate this across three protein systems: on an α-synuclein amyloid filament, strongly dewetting interfaces align with unassigned peptide densities, revealing hidden binding sites; in calmodulin, hydrophobicity redistributes upon Ca