When a protein or small molecule moves in water, the water does not simply push back like syrup. Water molecules in the first two or three layers around a biomolecule move at widely different speeds, some trapped for long stretches and others behaving almost like bulk water. This spread of speeds is why water near proteins and inside cells spreads more slowly than ordinary diffusion predicts, a behavior called subdiffusion. Fast machine-learned simulators that remove the water, so they only have to move the solute, discard this information. They either replace water with an instantaneous averaged force, or model its lagging effect with a "memory" computed only from the solute's own past motion. We propose to keep a cheap, coarse summary of the hydration shell: a per-region map of how mobile the nearby water currently is, updated at a slow fixed rate. The solute's fast dynamics look up this map through cross-attention, the same device automatic video-dubbing systems use to let a steady video frame rate set the timing of generated speech.
The core bet is falsifiable and narrow. When hydration-shell mobility is heterogeneous, the current state of the shell carries information about upcoming solute kinetics, such as how long before a conformational transition, that a short window of solute history cannot recover. If that holds, a two-stream simulator in which solute motion attends to the hydration-shell stream should reproduce transition times and velocity correlations that single-stream memory models get wrong. If it does not hold, the right conclusion is that solute-history memory is enough. The program is designed to settle that question before any architecture is scaled up.
Core Claim
A coarse-grained solute obeys, formally, a generalized Langevin equation (GLE), which is Newton's law plus a friction term that depends on past velocities and a correlated random force:
$$M\ddot{x}(t) = -\nabla U(x) - \int_0^t K(t-s)\,\dot{x}(s)\,ds + \eta(t),$$
where $x$ are solute coordinates, $M$ their masses, $U$ the averaged (potential of mean force) energy, $K$ the memory kernel, and $\eta$ the random force whose correlations are tied to $K$. The standard kernel is stationary and depends only on elapsed time. Our claim is that in heterogeneous hydration environments the effective kernel is modulated by a slowly varying hydration state $h(t)$, a vector of per-surface-patch water residence times and mobilities. We therefore replace $K(t-s)$ with a learned $K_\theta(t-s \mid h)$. The dependence on $h$ is implemented by solute-state queries attending over a sequence of hydration-state frames used as keys and values. A separate, cheap stochastic model advances the hydration stream at a coarse clock, roughly one frame per picosecond, while the solute advances at a much finer step.
What is new is not memory in learned molecular dynamics, and not attention over trajectories. It is the explicit claim, tested by ablation, that a separately evolved solvent-mobility stream is a necessary second input for kinetics.
Relation to Prior Work
Two papers motivate the gap. Rui Li et al. combined neutron scattering on perdeuterated proteins with all-atom simulation. They showed that water subdiffusion comes from a broad distribution of microscopic mobilities concentrated in the first few hydration layers, not from uniform slowing, and that this distribution is broader in E. coli than in protein solutions. Chu et al. showed that transient, rearranging energy traps formed by mobile charged lipids produce subdiffusion (MSD exponent near 0.38) in long polyelectrolytes, even while displacement distributions stay Gaussian. Both treat heterogeneity as something to measure. Neither builds a model that uses it to predict solute motion. Chu et al. also treat water only implicitly.
Two papers constrain the design. ISSNet learns a solvent-free potential of mean force by force matching on explicit-water trajectories of capped alanine and chignolin, and beats Generalized Born continuum models on thermodynamics. By construction, though, it is memoryless, so it can be right about equilibrium populations and wrong about rates. Xiantao Li et al. fit a rational-function memory kernel for one collective variable of benzyl bromide in 1,011 waters. Markovian models failed to reproduce mean first-passage times, 1st- and 2nd-order kernels missed a low-frequency peak, and 3rd–4th order kernels matched explicit MD within a 95% confidence interval. This is the strongest existing evidence that memory matters. It is also a warning: a well-fit stationary kernel already works for a small solute, so any gain from the hydration stream must be shown where heterogeneity is large.
The method comes from video dubbing. VisualTTS lets text queries attend to lip-video embeddings so that the fixed frame rate constrains durations and prevents decoder drift. Neural Dubber inverts this: video frames query the script, and the result is upsampled by the exact ratio of the audio and video rates. We borrow the multi-rate structure. The hydration stream plays the role of video, with a fixed coarse clock, and the fast solute stream plays the role of audio. Upsampling is fixed by the ratio of the two integration steps, and there is no separate duration predictor. Hassid et al. found that attention trained on paired data learns synchronization without an explicit sync loss, which suggests we need not hand-design how $h$ modulates $K$. Neural Dubber's reported weakness, degradation when the script diverges from the video, has a direct analogue here: the learned hydration stream may drift from the true solvent state. We measure this failure mode directly.
Learned trajectory models without solvent set the context. GF-NODE models vacuum molecules with graph-Fourier modes and a neural ODE, stable to 15,000 steps. I-MLFF speeds up Markovian force evaluation. OneProt with MD tokens aligns solvent-stripped trajectory embeddings contrastively. Piochi et al. report that AlphaFlow underestimates per-residue flexibility in 91% of protein–nucleic-acid complexes. All of these drop water, so none can test our claim. We do not claim hydration memory explains the AlphaFlow gap. That dataset strips solvent, so it cannot answer the question.
Research Program
The first stage is a go/no-go analysis on reference data and involves no learning. We run explicit-water all-atom MD of three systems of increasing heterogeneity. The first is benzyl bromide in water, replicating Li et al. so the GLE baseline is on its home turf. The second is chignolin, matching the ISSNet setting. The third is chignolin in a crowded solution of inert proteins at reduced hydration, chosen to mimic the low-hydration regime of Rui Li et al. Before trusting the crowded system, we check that its water MSD exponents fall in the range those authors report, from about 0.784 at h=0.4 to 0.945 at h=4.0.
From each trajectory we define the hydration state $h$ as per-patch water residence times and short-window mobilities in the first and second shells. We then ask one question: holding the recent solute history fixed (matched on a window of collective-variable values and velocities), does the distribution of first-passage times to the next conformational basin still depend on $h$? We estimate this as a conditional mutual information using a nearest-neighbor estimator, with a surrogate test that shuffles $h$ across matched windows. If the dependence is indistinguishable from zero in the crowded system, the project stops and reports a clean negative result. That result is useful in itself: it would say solute-only memory models are sufficient even under strong heterogeneity.
If the first stage passes, the second stage builds the two-stream simulator. The potential of mean force is learned ISSNet-style by force matching. The fluctuating force is learned on residuals, with attention from solute queries over the last $L$ hydration frames. The random force is generated by auxiliary variables whose correlations are tied to the learned kernel, following the extended-Markovian construction of Li et al., so the model satisfies fluctuation–dissipation by construction. The hydration stream is a small stochastic model (a graph over surface patches, advanced one frame per picosecond) trained to reproduce residence-time distributions and to be conditioned on the current solute pose. Two-way coupling is essential: water rearranges when the solute moves.
Baselines
The naive reference point is the Markovian Langevin simulator on the learned potential of mean force, with a scalar friction fit to the solute velocity autocorrelation. This is ISSNet's approach with constant friction. It fixes equilibrium quality, so any kinetic difference is attributable to memory. Every model shares this potential of mean force, trained once on the same data.
The strongest existing competitor is the rational-kernel GLE of Xiantao Li et al., at 3rd and 4th order. On benzyl bromide it matched explicit-MD first-passage times within a 95% confidence interval. The file gives no absolute first-passage-time value, and no reported result found on Lacuna for this method on chignolin or crowded systems. We re-implement it on identical collective variables and training trajectories, and choose its interpolation points by validation rather than by hand. Beating it means that on the crowded system our model's first-passage-time estimate falls inside the explicit-MD confidence interval while the rational GLE's falls outside. On benzyl bromide we expect a tie, and we report it as such.
The ablation that isolates the key ingredient is a solute-history-only attention memory model. It has identical architecture, parameter count, training steps and hyperparameter search budget, but its keys and values are past solute states instead of hydration frames, with the history length extended to match total context. A second ablation feeds the true hydration frames time-shuffled within trajectories. This tests whether the model uses hydration state or just extra input. A third feeds oracle hydration frames from explicit MD instead of the learned stream. This separates the value of the information from the error of the cheap hydration model. All models receive the same GPU-hour tuning budget.
Why not just…?
Why not just fit a higher-order stationary GLE kernel? This is the most serious alternative, and on benzyl bromide it already works per Li et al.. The physical reason it might fall short elsewhere is that a stationary kernel averages over hydration states. When residence times span orders of magnitude, as Rui Li et al. report, the rate of the next transition depends on which state the shell is currently in. An averaged kernel can then get the mean first-passage time right but its spread wrong. We therefore compare full first-passage-time distributions stratified by initial hydration state, not only means. If the stationary kernel matches those stratified distributions, it wins.
Why not just give a long-context sequence model the solute history? In the exact theory, solute history contains all the bath information the memory needs, so this could suffice. The first-stage conditional mutual information test and the solute-history-only ablation answer this directly. If the stage-one test finds no residual dependence on $h$, we adopt this simpler model and drop the hydration stream.
Why not just keep explicit water, or subsample it? Explicit water is the ground truth, and we use it as such. The motivation for removing it is cost, since water is typically 80–90% of the particles. The hydration stream is justified only if it is much cheaper than explicit water. We report wall-clock time per nanosecond against explicit MD. We also report how kinetic accuracy degrades as the hydration stream's frame rate drops, which traces the accuracy–cost curve.
Why not just use a continuum solvent such as Generalized Born? ISSNet already beats it on thermodynamics, and it has no memory at all, so it inherits every kinetic failure of the naive baseline. We include it only as a reference line.
Proposed Experiments
The decisive checks come in order. First, the stage-one conditional mutual information test on all three systems, with a shuffled-surrogate null. Second, mean first-passage time and its full distribution between conformational basins for all models against explicit-MD confidence intervals. Third, solute velocity autocorrelation and reconstructed memory kernels. Fourth, the solute MSD exponent over long rollouts, and whether the model preserves subdiffusion when the Markovian baseline does not. Fifth, stratified first-passage times by initial hydration state, the test that most directly distinguishes our model from a stationary kernel. Sixth, the oracle-versus-learned hydration stream gap, which plays the role of Neural Dubber's script-divergence failure.
Long-horizon stability is tracked out to rollout lengths comparable to GF-NODE's 15,000 steps. Its benzene coordinate error of 1.52 is a vacuum result and is cited only as context, not as a target.
The project succeeds if the crowded system shows a nonzero hydration dependence, and if the two-stream model then beats both the 4th-order GLE and the solute-history ablation on stratified first-passage times at a large cost reduction relative to explicit MD. It fails informatively if the dependence is absent.
References
- Anomalous sub-diffusion of water in biosystems: From hydrated protein powders to concentrated protein solution to living cells
- Anomalous Diffusion of Polyelectrolyte Segments on Supported Charged Lipid Bilayers
- Machine learning implicit solvation for molecular dynamics
- Data-driven molecular modeling with the generalized Langevin equation
- Visualtts: TTS with Accurate Lip-Speech Synchronization for Automatic Voice Over
- Neural Dubber: Dubbing for Videos According to Scripts
- More than Words: In-the-Wild Visually-Driven Prosody for Text-to-Speech
- Graph Fourier Neural ODEs: Bridging Spatial and Temporal Multiscales in Molecular Dynamics
- Implicit Machine Learning Force Fields Accelerate Molecular Dynamics Simulations
- When Protein Dynamics Matter: Integrating Molecular Dynamics into Protein Foundation Models
- Beyond Static Snapshots: A Large-Scale Dataset for Dynamics-Aware Protein-Nucleic Acid Modeling