Feed-Forward Hydration Fields: Predicting Structural-Water Dynamics in One Pass

About 85% of protein–ligand binding interfaces are completed not by direct contacts but by ordered water molecules that bridge chemical gaps between protein and ligand (Rapid and Accurate Prediction and Scoring of Water Molecules in Protein Binding Sites). These structural waters decide whether a drug candidate displaces a tightly held water (paying an entropy debt) or recruits a loose one (gaining a hydrogen bond). Yet the two tools we have for knowing which waters sit where are badly mismatched to the need. Machine-learning water-placement models treat the protein as a frozen crystal and each water as a single oxygen sphere, trained on cryogenic X-ray coordinates that may not reflect room-temperature behavior. Molecular dynamics (MD) simulations — which track every atom by integrating Newton's equations at femtosecond timesteps — give the physical truth about water mobility, energetics, and residence times, but cost hours to days per system on large clusters.

This proposal takes a mechanism from an unrelated area of machine learning and points it at that gap. In computer vision, "feed-forward 3D Gaussian prediction" has replaced slow per-scene optimization: instead of iteratively refining millions of ellipsoid primitives against a rendering loss, a network regresses all primitive parameters (position, orientation, scale, opacity) from an image in a single forward pass, in under a second (Sharp Monocular View Synthesis in Less Than a Second; Learning Compact 3D Gaussians via Feed-Forward Point Fusion). I propose to do the same for protein hydration: a network that reads a dry protein–ligand structure and, in one pass, emits a compact set of oriented ellipsoid primitives representing the hydration shell — each carrying not just a position but an occupancy, a local free energy, and a residence time, all supervised by explicit-solvent MD trajectories rather than crystal waters.

Core Claim

The hydration shell of a binding pocket — where waters sit, how strongly each is bound, how long it stays before exchanging with bulk, and how its spatial and orientational mobility is distributed — can be regressed from the unsimulated structure in a single feed-forward pass, with MD trajectories acting as the teacher. Two existing lines stop just short of this and cannot reach it. SuperWater (Superwater as a generative AI framework to predict water molecule positions on protein structures) places waters accurately but does so with an iterative, score-based diffusion process (many denoising steps), conditions on a rigid protein, predicts only an unpolarized oxygen point, and is trained on static crystallographic waters. Grid Inhomogeneous Solvation Theory, or GIST (Structure and Thermodynamics of Molecular Hydration via Grid Inhomogeneous Solvation Theory), computes exactly the thermodynamic hydration quantities I want — local entropy, enthalpy, and free-energy densities — but only by post-hoc analysis of multi-nanosecond MD trajectories; it is a measurement, not a predictor. The claim is that a learned one-pass predictor, trained on MD-derived fields, can approach GIST's physical content at a tiny fraction of its cost, and that the dynamic quantities it predicts (occupancy, residence time, anisotropy) carry signal that crystal-trained placement models structurally cannot provide.

Research Program

The first deliverable is a training corpus of dynamic hydration. I will run explicit-solvent MD on a stratified subset of the PDBbind Refined Set (version 2016) — the same complexes used to train the strongest water-aware affinity model, Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands — with the CASF-2016 core 285 complexes held out entirely. From each trajectory I will extract per-site water oxygen density maps, GIST-style enthalpy/entropy/free-energy fields, water-identity residence times (tracking when a specific molecule leaves a site), and hydrogen-bond persistence statistics of the kind used to filter docking hits in Identification of highly effective inhibitors against SARS-CoV-2 main protease: From virtual screening to in vitro study. The large-scale protein–nucleic acid trajectory dataset of Beyond Static Snapshots: A Large-Scale Dataset for Dynamics-Aware Protein-Nucleic Acid Modeling will serve as an additional pretraining corpus for general protein-surface hydration before fine-tuning on the protein–ligand trajectories.

The second deliverable is the representation and model. The hydration shell is encoded as a set of anisotropic Gaussian primitives whose parameters the network regresses in one pass:

$$f_\theta(\text{protein}, \text{ligand}) = \{(\mu_i,\, q_i,\, s_i,\, o_i,\, \Delta G_i,\, \tau_i)\}_{i=1}^{N}$$

where $\mu_i$ is the primitive's 3D center, $q_i$ its orientation quaternion, $s_i$ its scale along three axes, $o_i$ the site occupancy (the fraction of MD frames a water occupies it), $\Delta G_i$ the local excess free energy, and $\tau_i$ the mean residence time. The ellipsoid axes and orientation directly answer SuperWater's acknowledged monoatomic-oxygen limitation: an elongated primitive with a preferred axis encodes both the anisotropic spatial spread of a mobile water and the directional preference of its hydrogen-bond geometry, which a scalar oxygen point cannot. The encoder is an SE(3)-equivariant graph network over protein and ligand atoms — the same symmetry property SuperWater uses, and one that vision-side feed-forward predictors like SHARP lack entirely, as Learning Compact 3D Gaussians via Feed-Forward Point Fusion does for optical scenes — with a compact query-based head that emits a bounded set of primitives per pocket rather than one per atom, following the compactness argument of that same paper. Training combines a density-matching term (rendering the Gaussian mixture into a 3D oxygen density and comparing it to the MD-derived density), a free-energy field term against the GIST-computed field, and a robust regression on $\log \tau$ for matched long-lived sites, plus a consistency term so that high-occupancy, long-residence predictions align with crystallographic waters where those exist.

A secondary thread, treated as a stretch goal, addresses the coupling gap that both Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands and SuperWater explicitly leave open: predicting small water-induced side-chain rearrangements alongside the hydration field, validated against the coupled water–side-chain motion that MD studies of interface hydration demonstrate (Computational simulations reveal the binding dynamics between human ACE2 and the receptor binding domain of SARS-CoV-2 spike protein; Sampling protein motion and solvent effect during ligand binding).

Baselines

The naive reference point is the dry molecular graph: the same Graphormer affinity model with no water nodes at all. This is not a strawman — the GraphWater-Net paper reports that removing water from its input drops the CASF-2016 Pearson correlation by 0.154 from its water-aware value of 0.868 (Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands), so the dry reference sits at roughly $R_p \approx 0.714$; any water model must clear that to justify its existence.

The strongest existing method on water placement is SuperWater, the score-based diffusion model with an SE(3)-equivariant scoring network, which reports a mean absolute deviation of $0.3 \pm 0.06$ Å from crystallographic waters and greater than 70% precision at 50% coverage with a 1.0 Å matching radius on its high-resolution benchmark (structures under 1.5 Å with a water-to-residue ratio above 0.6) (Superwater as a generative AI framework to predict water molecule positions on protein structures). On that same benchmark and protocol it reports under 60% precision for HydraProt and roughly 55% for GalaxyWater-CNN at the same coverage, which establishes those empirical and CNN-based tools as the weaker placement baselines. I will re-run SuperWater and both weaker tools on my held-out complexes under the identical matching radius, coverage, and compute budget, and also give each method equal tuning effort on the training split. Beating SuperWater means matching or exceeding its precision-at-coverage and MAD while producing its outputs in one pass instead of iterative denoising, and additionally predicting occupancy, free energy, and residence time — quantities SuperWater does not output at all, so on those metrics it cannot compete by construction.

The strongest downstream competitor is the two-stage GraphWater-Net pipeline: off-the-shelf HydraMap and 3D-RISM water predictions frozen as extra graph nodes, fed to a four-layer, twelve-head Graphormer that reaches $R_p = 0.868$ on CASF-2016 (Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands). I will plug my predicted waters — including their $\Delta G$ and $\tau$ as node features — into the identical Graphormer configuration, trained on the same PDBbind split, so the only difference is the water source and its differentiability. Beating it means exceeding 0.868 on the same benchmark, which would show that end-to-end trained dynamic waters beat frozen heuristic placements.

The physics reference is GIST computed from actual MD (Structure and Thermodynamics of Molecular Hydration via Grid Inhomogeneous Solvation Theory). Since GIST requires exactly the simulations I want to bypass, it serves as ground truth for held-out systems and as a cost-accuracy competitor at short simulation lengths: I will compare my prediction against both a converged long-trajectory GIST and a cheap 20-nanosecond GIST, all on identical systems. No reported accuracy number for GIST itself exists in this literature — it is an analysis method, so there is no reported result found on Lacuna for GIST prediction accuracy; the comparison must be constructed by me. Finally, the key ablation isolates my central ingredient: the identical feed-forward architecture and compute, trained only on static crystallographic waters with no MD targets. If that variant matches the MD-trained model on the dynamic metrics, the proposal's core claim is falsified rather than supported.

Why not just…?

Why not just run explicit-solvent MD? For one important system, yes — MD is the gold standard, and nothing here claims otherwise. But the reviewer's own framing in this literature concedes the scaling problem: femtosecond timesteps confine standard simulations to the nanosecond-to-microsecond regime on large clusters, and equilibrating deeply buried structural waters takes even longer or needs special insertion schemes. The Lupala et al. ACE2–spike analysis required hundreds of nanoseconds just to characterize one interface (Computational simulations reveal the binding dynamics between human ACE2 and the receptor binding domain of SARS-CoV-2 spike protein). The honest test is in my experiment two: if a 20-nanosecond GIST beats my one-pass prediction everywhere against a converged reference, the practical case collapses, and I will report it that way.

Why not just use 3D-RISM, HydraMap, or HydraProt? These are the incumbent tools and must be beaten, not dismissed. The evidence in the file says they fall short on discrete conserved-water identification: on the shared high-resolution benchmark, HydraProt stays below 60% and GalaxyWater-CNN near 55% precision at 50% coverage, versus above 70% for the learned continuous approach (Superwater as a generative AI framework to predict water molecule positions on protein structures). More fundamentally, they are used as frozen preprocessing in the strongest affinity pipeline (Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands), so no gradient from a downstream task can improve the water placement. I will still run them head-to-head rather than rely on reported numbers alone.

Why not just distill SuperWater into a one-pass student? This is the most serious alternative, and I treat it as a baseline rather than an argument. Distillation would fix the latency problem but inherit everything else: static cryo-crystal supervision, a rigid protein, a scalar oxygen with no orientation, and no kinetics or thermodynamics. My factorial ablation — feed-forward architecture trained on crystal targets versus the same architecture trained on MD targets — is precisely the experiment that separates "one-pass matters" from "MD supervision matters," and if the crystal-trained student matches the MD-trained model on free-energy and residence-time accuracy, the MD-training claim of this proposal is wrong.

Why not just use a 3D CNN on a voxelized solvent grid, or a co-folding model like AlphaFold-3? Voxel grids are the standard vision answer, but the file's evidence is direct: 3D CNNs are rotation-sensitive and memory-intensive at sub-angstrom resolution, forcing a resolution–memory trade-off that the continuous-coordinate formulation was built to escape (Superwater as a generative AI framework to predict water molecule positions on protein structures). Co-folding models fail differently: they systematically misplace ligands at underexplored interfaces (ACER: Towards Generalizable Protein-ligand Co-folding) and predict dry static structures with no explicit solvent, so they do not currently compete on this question at all.

Proposed Experiments and Decisive Checks

Experiment one evaluates static placement on the SuperWater benchmark (structures under 1.5 Å, water-to-residue ratio above 0.6), reporting precision at 50% coverage with a 1.0 Å matching radius and mean absolute deviation, against SuperWater's reported 0.3 ± 0.06 Å and greater-than-70% precision. A sharper, more novel check exploits the crystal-artifact tension directly: waters my model predicts to have low occupancy and sub-nanosecond residence should be exactly the crystallographic waters that exchange rapidly in held-out MD, testing whether MD supervision lets the model flag cryogenic freezing artifacts that crystal-trained models treat as ground truth.

Experiment two evaluates the dynamic field on held-out MD: Spearman correlation of predicted per-site $\Delta G$ against GIST values, correlation of predicted $\log \tau$ against observed residence times, and a density-field discrepancy between the rendered Gaussian mixture and the MD oxygen histogram, all benchmarked against short-trajectory GIST to locate the model on the cost–accuracy frontier. Experiment three evaluates downstream value: feeding predicted waters and their physical attributes into the Graphormer configuration of Incorporating Water Molecules into Highly Accurate Binding Affinity Prediction for Proteins and Ligands on CASF-2016, where the targets to beat are the reported $R_p = 0.868$ with heuristic waters and roughly 0.714 dry; a small auxiliary check reruns the pipeline on FreeSolv hydration free energies, for which no reported result is found on Lacuna in this literature, so I will establish my own reference numbers. Experiment four measures wall-clock inference per pocket and behavior beyond 1,000 residues, where SuperWater reports severe scaling bottlenecks; the target is sub-second inference on a single GPU, the standard set by feed-forward vision models (Sharp Monocular View Synthesis in Less Than a Second).

The decisive check for the whole program is the factorial ablation, run with equal data, compute, and tuning: two architectures (feed-forward and diffusion) crossed with two supervision sources (crystal waters and MD fields). The proposal succeeds if the MD-trained feed-forward model wins on thermodynamic and kinetic accuracy while matching placement precision, and it fails — in a publishable and informative way — if crystal supervision turns out to contain all the signal MD provides.