HEIWA: Hybrid Explicit-Implicit Water Attention Affects Force Field Accuracy in Structure-Based Protein Language Model ESM3

Protein language models have achieved remarkable success in structure prediction and de novo design. However, their understanding of physical interactions remains limited. For structure-based models, we hypothesize that this limitation stems from the absence of explicit solvent modeling: water mediates hydrogen bonding, hydrophobic burial, and electrostatic interactions that govern protein thermodynamics and kinetics, yet current structure-based models treat proteins as if they exist in a vacuum. Using thermophilic protein classification as a proxy task for force-field awareness, per-residue logit mis-correlated with H-bonds as a destabilizing attribute and exhibits significant solvent-accessible surface area (SASA)-dependent error, despite high overall classification accuracy. These artifacts point to the water-protein interface as a source of systematic bias. To address this, we propose HEIWA (HybridExplicit-Implicit Water Attention), a novel attention module that injects solvent-mediated force-field information into static ESM3 residue embeddings via an all-query-to-all-key water-residue attention. Biophysics-guided model selection identifies configurations whose residue-level outputs are more consistent with the RMSF benchmark and selected biophysical attributes, while classification performance remains broadly comparable to the baseline. By forcing the model to predict the protein thermophilicity from the contributions of individual residues, HEIWA learns biophysical patterns that can be extended to SUPFAM absent from the training data set. HEIWA also largely preserves the embedding structure of ESM3. Molecular dynamics simulations show improved kinetic stability under thermophilic conditions, suggesting the potential to generate structures with greater kinetic stability. To our knowledge, this is the first study to incorporate explicit solvent encoding into a residue-resolution protein language model.