Generative Modeling of Solvated Biomolecules

We introduce BoltzW, the first generative framework for solvated biomolecules. BoltzW models the joint distribution of proteins, nucleic acids, small molecules, and water. Explicit solvent modeling poses novel challenges, each concerning limited modeling capacity, data inconsistency, physical indistinguishability, and variable cardinality. We address these by selecting structured solvents, designing permutation-invariant training objectives, and formulating unified structure targets that bypass cardinality prediction. BoltzW improves biomolecular structure prediction with minimal finetuning from Boltz-2, while achieving state-of-the-art water placement even compared to imputation methods that have access to the ground truth protein structure.