SuperWater

Ordered waters bridge a ligand to its pocket, hold two chains together across an interface, and pay part of the binding free energy. A crystallographer sees them; a modeler working from a predicted structure does not, and docking pipelines routinely delete them because nobody knows which ones a new structure would have.

SuperWater treats that solvent shell as something to generate rather than something to score. Existing deep-learning hydration predictors — HydraProt and GalaxyWater-CNN — represent the protein as a 3D image and classify voxels as water-occupied or not, so sub-angstrom accuracy demands a very fine grid and cost grows with the cube of the resolution. SuperWater instead learns the score function of the conditional density of water positions given a protein, then samples from it: candidate waters are scattered at random, reverse diffusion drives each toward an energetically favorable site, a confidence model discards the implausible ones, and the survivors are clustered into final positions. Coordinates stay continuous throughout, so precision is not bounded by a grid.

The framework was built at Vanderbilt University's Data Science Institute with Takeda and the Albert Einstein College of Medicine, posted as a preprint in November 2024 and published in Communications Chemistry in December 2025. It is the water-placement member of the same sample-filter-cluster family as SuperMetal, which places zinc ions, and it adapts the SE(3)-equivariant diffusion architecture of DiffDock.

#Key Features

• Continuous placement, not voxel classification: Positions are sampled in continuous space, so spatial accuracy does not trade against grid spacing or memory.
• The number of waters emerges: Candidates are seeded at fifteen times the residue count, then filtered and clustered, so the output count follows from the structure rather than a user-supplied number.
• An explicit precision-coverage dial: Raising the confidence cutoff tightens precision and lowers coverage along a smooth curve, letting a user pick an operating point.
• Interfacial waters evaluated separately: Dedicated benchmarks cover bridging waters at protein-protein and protein-ligand interfaces, the ones that matter most for design.
• Runs from shipped checkpoints: Pretrained score and confidence models ship in the repository, and one command hydrates a folder of PDB or mmCIF files.

#Technical Details

The score model is a variance-exploding stochastic differential equation over water oxygen coordinates conditioned on the protein, parameterized by an SE(3)-equivariant graph neural network built with e3nn over a heterogeneous graph of residues, atoms and candidate waters, with residues featurized by cached ESM-2 embeddings; it was trained for 300 epochs. The confidence model is a second equivariant convolutional network trained by mean squared error against a normalized distance to the nearest crystal water. Clustering is confidence-weighted, uses the 1.52 Å van der Waals radius of oxygen to define neighbors, and merges centroids closer than 2.2 Å so predictions cannot clash. Inference uses 20 reverse-diffusion steps.

Training data was curated from PDB entries better than 1.5 Å resolution — 23,189 files initially, narrowed to structures of 100 to 500 residues with a water-to-residue ratio above 0.6, giving 17,092 complexes split 8:1:1 into 13,674 training, 1,709 validation and 1,709 test structures. Only waters within 4 Å of the protein were kept, each as its oxygen alone. On the held-out set at a 1 Å matching radius and 50% coverage, SuperWater sustains above 70% precision against under 60% for HydraProt and roughly 55% for GalaxyWater-CNN, and at a confidence threshold of 0.5 its mean absolute deviation from crystallographic positions is 0.3 ± 0.06 Å. On 550 multi-chain complexes it holds about 25% coverage of bridging waters at 90% precision, roughly three times HydraProt; on 600 protein-ligand co-crystals it reaches 95% precision at 25% coverage. Case studies cover carbonic anhydrase II, the PHIP protein, and the NDM1-meropenem complex.

#Applications

The immediate use is preparing structures for work that depends on explicit solvent: seeding molecular dynamics, restoring the waters that structure-based drug design deletes, and identifying bridging waters a medicinal chemist can displace or hydrogen-bond to. Because it takes any PDB or mmCIF file, it also hydrates predicted structures, adding a solvent shell to output from co-folding models such as Boltz-2.

#Impact

SuperWater shows that the generative recipe established for ligand pose prediction transfers to solvent placement and beats voxel-based hydration predictors on precision, coverage and spatial accuracy at once. Its limits are stated plainly by the authors: each water is a single oxygen atom, so dipole orientation and hydrogen-bond directionality are absent; and because candidate sites scale with residue count, GPU memory and runtime grow steeply with protein size, making structures above about 1,000 residues impractical without manual chunking. Inference is CUDA-only, with no CPU path. Code and both checkpoints are MIT-licensed on GitHub with a Zenodo software archive, and the training set and processed data are archived on Zenodo, so the pipeline is reproducible end to end.