How to Use SuperWater

SuperWater predicts water-molecule positions on protein surfaces using a generative diffusion workflow. It is designed for structural biology and protein-engineering tasks where ordered or semi-ordered hydration waters can affect binding, stability, catalysis, or interpretation of a protein surface.

On Neurosnap, researchers upload one or more Input Structures, choose whether returned structures should include the protein coordinates, and tune sampling parameters in Advanced Options when they need stricter confidence filtering or broader water coverage. The service returns final centroid water coordinates as a CSV table together with structure files containing predicted HOH waters for visualization and downstream analysis.

SuperWater uses a score-based diffusion model with equivariant neural networks to move initially sampled water particles toward likely hydration sites around a protein. A separate confidence model scores candidate waters, and a clustering step produces the final centroid positions. This architecture is useful because hydration-site prediction is a geometric and chemical problem: water placement depends on the protein surface, local interaction patterns, and three-dimensional context rather than sequence alone.

In practical use, Water Ratio controls how many candidate waters are sampled per residue, Inference Steps controls reverse-diffusion refinement depth, and Confidence Cutoff determines how strictly candidate waters are filtered before the final outputs are produced. Lower cutoffs can recover more potential hydration sites, while higher cutoffs produce a more conservative set.

Researchers should interpret SuperWater outputs as structural hypotheses for hydration-site review. Predicted waters can help inspect protein-ligand interfaces, polar cavities, catalytic sites, mutation effects, and regions where solvent-mediated contacts may influence design decisions, but they should still be checked against experimental density, simulations, or downstream energetic analysis when available.

• Structure-first workflow: SuperWater starts from uploaded PDB or mmCIF structures, matching how hydration questions are usually framed in structural biology.
• Generative water placement: Diffusion sampling explores candidate solvent positions around the protein surface instead of only applying fixed geometric rules.
• Confidence-guided filtering: The confidence cutoff lets researchers tune whether they want exploratory coverage or a smaller high-confidence set of waters.
• Visualization-ready outputs: Neurosnap returns coordinate tables and centroid structure files so predicted waters can be inspected alongside the protein in standard molecular viewers.