betaFeedback

WorksPhysics and Astronomy

GPU infrastructure for ML/MM with electrostatic embedding for enzyme chemistry

Abstract

We present a fully GPU-native framework for machine learning/molecular mechanics (ML/MM) enhanced sampling simulations with electrostatic embedding, demonstrated on an enzymatic reaction. The framework combines a JAX port of the pretrained MACE-POLAR-1-M model, an electrostatic embedding scheme compatible with the MACE-POLAR architecture, and integration with JAX-MD and PLUMED-JAX for end-to-end compiled enhanced sampling. The implementation is developed using GPT-6.1-Sol within Codex, with reasoning effort set to high. Using replica exchange umbrella sampling, we calculate the free energy profile of nucleophilic substitution in haloalkane dehalogenase in approximately 200 H100 GPU hours, treating the entire protein and reactants with MACE-POLAR-1-M-JAX and most of the solvent with molecular mechanics.

Introduction

Simulating enzyme reactions requires describing both changes in chemical bonding and interactions with the surrounding protein and solvent. Quantum mechanics/molecular mechanics methods address this challenge by treating a selected region with electronic structure theory and its surroundings with a classical forcefield [34, 35]. Replacing electronic structure calculations with machine learning interatomic potentials (MLIPs) can substantially reduce computational costs, but conventional MLIPs lack explicit electronic degrees of freedom, making it challenging to capture the influence of the environment on the charge distribution of the reacting region. Furthermore, biochemical reactions often involve rare events on long timescales, requiring both efficient simulation implementations and enhanced sampling methods to obtain converged free energies.

We note that the first challenge can be addressed by the MACE-POLAR architecture, which conditions atomic features on the local electrostatic potential [36], allowing atoms, in principle, to respond to the surrounding molecular mechanics (MM) environment. The use of electrostatic-potential features to describe long-range interactions builds on the long-distance equivariant (LODE) representation introduced by Grisafi and Ceriotti [37]. The second can be addressed using JAX-MD [38] for efficient molecular dynamics and PLUMED-JAX [39], a port of selected PLUMED functionality [40, 41], for enhanced sampling. To combine these capabilities in an end-to-end compiled graphics processing unit (GPU) workflow, we develop MACE-POLAR-1-M-JAX, a JAX [42] implementation of the pretrained MACE-POLAR-1-M model, couple it to fixed MM charges, and integrate it with JAX-MD and PLUMED-JAX. The implementation is developed using GPT-6.1-Sol within Codex, with reasoning effort set to high. We describe these components separately before demonstrating their application to nucleophilic substitution in haloalkane dehalogenase.

Figure 1. MACE-POLAR forward pass (A), mechanical embedding (B), and electrostatic embedding (C). Solid arrows show values passed and dashed arrows the next polarization update.

Methods and results

MACE-POLAR-1-M-JAX

The MACE-POLAR architecture is summarized in Figure 1A. The pretrained MACE-POLAR-1-M model [36] instantiates this architecture with weights trained on Open Molecules 2025 data at the ω\omegaB97M-V/def2-TZVPD level [43].

Its inputs are the atomic numbers ZiZ_i, positions Ri\mathbf R_i, cell h\textsf{h}, total charge, and spin multiplicity, where ii is the index of an atom. Local MACE layers combine information from neighbouring atoms to describe their geometric environments [44]. Learned functions of the resulting atomic features, denoted by hi\mathbf h_i in the schematic, predict initial charges and dipoles in two spin channels. Their coefficients are ci(u)\mathbf c_i^{(u)} at polarization update uu, and ρσ\rho_\sigma denotes the density in spin channel σ\sigma. Atom-centred Gaussian functions represent these coefficients as a smooth charge density, and a global correction enforces the specified total charge and spin multiplicity. The model then calculates the electric potential generated by this density and integrates it against atom-centred Gaussian functions to obtain features describing the potential around each atom. These features include contributions from distant atoms and are combined with the local geometric features to update the charge and dipole coefficients. The model performs two successive learned polarization updates, each using the potential generated by the current density and followed by a correction of the channel charges. Finally, the energy combines local atomic contributions, a learned response contribution, and the electrostatic energy of the final density, including its interaction with an applied potential. The mathematical formulation and pretrained model settings are given in the original MACE-POLAR paper [36].

MACE-POLAR-1-M-JAX retains this forward pass and the pretrained weights. It supports tensor operations through the cuEq bindings for JAX [45]. Automatic differentiation supplies the forces and the derivatives used to calculate periodic stress. Neighbour construction and model evaluation can be compiled together using either the original reciprocal space Fourier sum or an additional generalized particle-particle-particle-mesh (PPPM) solver.

Numerical verification

We compare MACE-POLAR-1-M-JAX with MACE-POLAR-1-M evaluated in PyTorch [46], using the same checkpoint, input precision, neighbour graphs, and periodic images.

We validate the implementation using two sets of configurations:

  1. For molecules, we use all 46 HR46 geometries [47] from the GSCDB138 input archive [48], centred on their molecular centroids with the supplied charges and spin multiplicities. Each is evaluated at zero field and positive and negative fields along each Cartesian axis, with magnitudes of 0.0010.001 and 0.0005 V A˚−10.0005\,\mathrm{V\,\mathring{\mathrm A}^{-1}}, giving 13 fields and 598 molecule–field combinations.

  2. For periodic liquid water, we use three frozen configurations at each of six sizes: 192, 375, 648, 999, 1,536, and 3,000 atoms, with a volume of 28.8209 A˚328.8209\,\mathring{\mathrm A}^3 per water molecule. The 999-atom inputs comprise the 300 K water configuration released with the MACE-POLAR study [36] and two small rigid-molecule perturbations of it. For the other sizes, water geometries from that configuration are randomly oriented around slightly displaced cubic-grid positions. The periodic comparisons use the original reciprocal solver in both programs, with a cutoff of approximately 1.85 A˚−11.85\,\mathring{\mathrm A}^{-1}.

Table tab:model-validation reports differences between the pretrained MACE-POLAR-1-M model evaluated using PyTorch and MACE-POLAR-1-M-JAX using cuEq, separately in float64 and float32 floating-point arithmetic. Energies are divided by the number of atoms before averaging over configurations, while vector and tensor differences include every Cartesian component. The dipole comparison uses identical coordinates and origin conventions. Stress is obtained from derivatives under a joint deformation of coordinates and cell.

PyTorch--JAX
float64 & 2cPyTorch--JAX
float32
QuantityRMSMaxRMSMax
Molecules: 598 molecule--field combinations
Total energy (meV/atom)5.77×10−65.77\times10^{-6}2.49×10−52.49\times10^{-5}0.2000.2000.9770.977
Forces (meV/)8.39×10−68.39\times10^{-6}5.07×10−55.07\times10^{-5}0.2060.2062.462.46
Atomic charges (ee)1.73×10−91.73\times10^{-9}1.38×10−81.38\times10^{-8}6.58×10−76.58\times10^{-7}5.63×10−65.63\times10^{-6}
Model dipole (ee)4.10×10−94.10\times10^{-9}1.77×10−81.77\times10^{-8}4.25×10−64.25\times10^{-6}2.93×10−52.93\times10^{-5}
Periodic water: 18 configurations
Total energy (meV/atom)1.14×10−61.14\times10^{-6}1.28×10−61.28\times10^{-6}4.254.259.799.79
Forces (meV/)2.45×10−52.45\times10^{-5}4.54×10−44.54\times10^{-4}1.31×10−31.31\times10^{-3}0.01140.0114
Atomic charges (ee)3.09×10−103.09\times10^{-10}1.01×10−81.01\times10^{-8}2.59×10−82.59\times10^{-8}6.56×10−76.56\times10^{-7}
Model dipole (ee)3.82×10−73.82\times10^{-7}2.29×10−62.29\times10^{-6}1.02×10−41.02\times10^{-4}2.72×10−42.72\times10^{-4}
Stress (meV/3^3)8.73×10−88.73\times10^{-8}2.41×10−72.41\times10^{-7}1.68×10−51.68\times10^{-5}6.15×10−56.15\times10^{-5}

The float64 results show close numerical agreement across the tested quantities. The larger float32 differences in energies are dominated by rounding in the large, position independent atomic reference, which does not affect forces.

Performance efficiency in model evaluation and molecular dynamics throughput

We first measure the cost of a complete model evaluation, including neighbour construction, graph assembly or transfer, energy, forces, and stress, with outputs retained on the GPU. Figure 2A compares six periodic water sizes containing 192–3,000 atoms at 1.04 g cm−31.04\,\mathrm{g\,cm^{-3}} on one RTX A6000, using cuEq and matched float64 or float32 arithmetic. The original MACECalculator constructs neighbours on the central processing unit (CPU) and transfers the graph to the GPU, whereas JAX compiles neighbour construction and model evaluation together. The tested PyTorch configuration runs in eager mode because it does not support torch.compile. Both implementations use the original reciprocal solver, with additional JAX curves showing the cost of generalized PPPM. Across the six sizes, JAX is 1.88–2.09 times faster in float64 and 2.10–6.37 times faster in float32, with respective speedups of 1.90 and 2.10 at 3,000 atoms, where PPPM and reciprocal timings are similar.

We then compare molecular dynamics throughput for isolated benzene (12 atoms), periodic water (999 atoms), and isolated haloalkane dehalogenase (DhlA; 4,858 atoms, charge −17e-17e), where ee is the elementary charge (Figure 2B). The pretrained model runs through MACECalculator with the Atomic Simulation Environment (ASE) [49] and the official MaceTorchSimModel interface with TorchSim [50], both in eager mode, while MACE-POLAR-1-M-JAX uses JAX-MD with neighbour construction, force evaluation, and integration compiled together. All routes use cuEq on the same RTX A6000, float32 model arithmetic, float64 integration, and a 0.5 fs velocity Verlet time step, starting from identical positions and momenta initialized at 300 K, without a thermostat, barostat, or constraints. Water uses the original reciprocal solver, while benzene and the protein use real space electrostatics; initial forces and trajectories are compared for benzene and water to check the common settings. Throughput is measured from seven blocks after warm-up, excluding loading, compilation, trajectory writing, and thermodynamic logging. Compiled JAX-MD reaches 24.4 ns/day for benzene and 0.788 ns/day for water, respectively 33.0 and 2.17 times the TorchSim rates. The larger gain for benzene is consistent with reducing host orchestration around short force calculations. The isolated protein runs at 0.0760 ns/day with MACE-POLAR-1-M-JAX, while both tested PyTorch/cuEq routes exhaust GPU memory during the initial force evaluation.

Figure 2. Computational performance on one RTX A6000: model evaluation (A) and molecular dynamics throughput (B), with cuEq in both implementations. Arrows in A compare the original reciprocal solver at matched precision; error bars show the interquartile range of seven timing blocks. Structures are rendered with VMD [51], with the water cell outlined and the protein shown as a backbone cartoon. OOM denotes out of memory during the initial force evaluation.

Embedding MACE-POLAR-1-M-JAX in molecular mechanics

Mechanical embedding

A machine learning/molecular mechanics (ML/MM) calculation divides the atoms into a machine learning region AA and a classical molecular mechanics region BB, with positions RA\mathbf R_A and RB\mathbf R_B. The mechanical embedding approach evaluates region AA without the electric field of region BB and adds classical interactions between the regions:

Umech=UML0(RA)+UMM(RB)+ULJAB+UCAB.\begin{equation*} U_{\mathrm{mech}}= U_{\mathrm{ML}}^{0}(\mathbf R_A)+U_{\mathrm{MM}}(\mathbf R_B) +U_{\mathrm{LJ}}^{AB}+U_{\mathrm{C}}^{AB}. \end{equation*}

Here, UML0U_{\mathrm{ML}}^0 is the model energy without the classical environment, UMMU_{\mathrm{MM}} is the energy within region BB, and ULJABU_{\mathrm{LJ}}^{AB} and UCABU_{\mathrm{C}}^{AB} are the cross-region Lennard–Jones and electrostatic interactions. The charges defining UCABU_{\mathrm{C}}^{AB} can be fixed forcefield charges or geometry-dependent charges. The latter allow charge redistribution along a reaction but not polarization by the classical environment, as in the chorismate-mutase calculations of Sun and co-workers [52]. Figure 1B illustrates how mechanical embedding can be implemented with MACE-POLAR.

Electrostatic embedding

Electrostatic embedding includes the response of the learned region to its classical surroundings. This response can be supplied by a separate physical model: electrostatic machine learning embedding [53] combines a gas-phase potential with permanent electrostatics and coupled, Thole-damped induced dipoles, as used by Gradisteanu and co-workers [54] and refined by Morado and co-workers [55]. Alternatively, environmental information can enter the MLIP: Lei and co-workers [56] use spatial derivatives of the classical potential, while AMP-BMS/MM [57] introduces field-induced dipole and quadrupole features during message passing, which is related to the approach used here.

Figure 1C illustrates electrostatic embedding with MACE-POLAR. The potential ϕB\phi_B generated by the fixed classical charges enters the learned polarization updates of MACE-POLAR-1-M-JAX. It is shared equally between the two spin channels, whose total charges are denoted by QσQ_\sigma. This allows the model charges and dipoles to respond to the environment even when the geometry of region AA is unchanged, while the classical charges remain fixed. The total energy is

Uelec=UMLemb(RA;ϕB)+UMM(RB)+ULJAB.\begin{equation*} U_{\mathrm{elec}}= U_{\mathrm{ML}}^{\mathrm{emb}}(\mathbf R_A;\phi_B) +U_{\mathrm{MM}}(\mathbf R_B)+U_{\mathrm{LJ}}^{AB}. \end{equation*}

Here, UMLembU_{\mathrm{ML}}^{\mathrm{emb}} includes the model energy, its response to ϕB\phi_B, and the cross-region electrostatic interaction, so no separate UCABU_{\mathrm{C}}^{AB} term is added. In our implementation, MACE-POLAR-1-M-JAX supplies the embedded model energy, while JAX-MD evaluates the classical interactions. Cross-region Lennard–Jones interactions use arithmetic mixing of size parameters and geometric mixing of well depths. For periodic systems, generalized PPPM evaluates the Gaussian multipole potential and coupled electrostatic energy, while the classical region uses point-charge particle mesh Ewald [58].

Enzyme free energy calculations

System

To demonstrate the proposed ML/MM scheme, we consider the first substitution step catalysed by DhlA with 1,2-dichloroethane. The carboxylate group of aspartate residue 124 (Asp124) attacks a substrate carbon while chloride departs, forming a covalent ester intermediate [59, 60]. This process combines bond formation, bond breaking, and a changing charge distribution in a protein environment.

Plotnikov [59] calculated a B3LYP/MM free energy profile, and Schanstra and co-workers [60] measured the rate of the chemical step, providing computational and experimental comparisons. Following Ref. , we describe the progress of the reaction using s=d(C−Cl)−d(C−O)s=d(\mathrm{C{-}Cl})-d(\mathrm{C{-}O}), where C is the attacked substrate carbon, Cl is the departing chlorine, O is the attacking carboxylate oxygen OD2 of Asp124, and dd denotes an interatomic distance. Negative values describe reactants, and positive values describe the ester intermediate, which is the product considered here.

Figure 3. DhlA free energy profiles (A) and sampling diagnostics (B). Blue in A shows electrostatic machine learning/molecular mechanics and red the B3LYP/MM profile from Plotnikov [59]; shading and error bars are block-jackknife standard errors, and the experimental marker is rate-derived [60]. Snapshots rendered with VMD [51] show capped reacting fragments and neighbouring residues at reactant (1), near-barrier (2), and ester intermediate (3) geometries; snapshot 2 is not an optimized transition state. Panel B shows 0.5 ps block means and normalized production histograms for all 72 umbrellas; the dashed line marks the end of 10 ps equilibration.

Molecular dynamics setup

The solvated system is prepared from Protein Data Bank entry 2DHC and contains 33,305 atoms in a periodic cubic cell of side 70.8685 Å. The electrostatically embedded region contains the entire protein, the substrate, and two waters, giving 4,872 atoms with charge −17e-17e and spin multiplicity one. The remaining 28,433 atoms are classical water and counterions, and the complete cell is neutral.

The classical solvent uses TIP3P parameters [61], and nonbonded interactions use a 10 Å cutoff and a 1083108^3 electrostatic mesh. The Gaussian multipole mesh has target spacing 0.5 Å and assignment order eight. Separate neighbour lists for modelled, classical, and cross-region pairs use a 0.5 Å buffer and are updated when any atom in the local batch moves more than 0.25 Å from its reference position. The smooth distance-ordering restraint is based on the flexible inner region ensemble separator [62], using distances from the Asp124 CG atom to the water oxygens. It acts when a smooth estimate of the largest distance among modelled waters exceeds a smooth estimate of the smallest distance among classical waters, with force constant 5 kJ mol−1 A˚−25\,\mathrm{kJ\,mol^{-1}}\,\mathring{\mathrm A}^{-2}. Classical waters are rigid under the SETTLE position and velocity constraints [63], while modelled waters are flexible.

Enhanced sampling setup

Following Ref. , we use umbrella sampling to estimate the reaction free energy profile. However, we also use replica exchange, which attempts Metropolis exchanges between neighbouring umbrellas at a common temperature [64]. Using a setup similar to that of Ref. , we run 72 umbrellas for 10 ps equilibration and 20 ps production each, using Langevin dynamics with a target temperature of 300 K. The 64 central umbrellas have centres from −1.125-1.125 to 2.0252.025 Å in increments of 0.05 Å. Four additional umbrellas at each end cover −1.525-1.525 to −1.225-1.225 Å on the reactant side and 2.1252.125 to 2.4252.425 Å on the product side, both in increments of 0.10 Å. All 72 umbrellas use a force constant of 1046 kJ mol−1 A˚−21046\,\mathrm{kJ\,mol^{-1}}\,\mathring{\mathrm A}^{-2}. Dynamics use a 0.5 fs time step and Langevin friction 1 ps−11\,\mathrm{ps}^{-1}. Exchanges are attempted every 200 steps, and the reaction coordinate is recorded every 20 steps. The free energy surface is estimated using the multistate Bennett acceptance ratio approach [65].

Distributed execution

We execute all 72 umbrellas in a single job, using JAX sharding across 18 NVIDIA H100 GPUs, with four umbrellas batched on each GPU using jax.vmap. Within each umbrella, JAX-MD advances the dynamics with biases supplied by PLUMED-JAX, using jax.lax.scan inside blocks compiled by jax.jit. Replica exchanges swap umbrella labels rather than transferring atomic configurations between devices, with bias forces updated after each accepted exchange. Sampling all 72 umbrellas requires approximately 200 H100 GPU hours, including initialization, compilation, 10 ps equilibration, and 20 ps production per umbrella.

Free energy profile and comparisons

Figure 3A shows the reconstructed free energy profile relative to the reactant minimum. The forward barrier is 85.2±0.4 kJ mol−185.2\pm0.4\,\mathrm{kJ\,mol^{-1}}, and the ester intermediate minimum lies 57.9±0.9 kJ mol−157.9\pm0.9\,\mathrm{kJ\,mol^{-1}} below the reactant minimum. The B3LYP/MM profile digitized from Figure 8 of Plotnikov [59] gives a barrier of 36.8 kJ mol−136.8\,\mathrm{kJ\,mol^{-1}} and an approximate product–reactant difference of −84.1 kJ mol−1-84.1\,\mathrm{kJ\,mol^{-1}}, without reported statistical uncertainties. The experimental chemical-step rate of 50±10 s−150\pm10\,\mathrm{s}^{-1} at 303.15 K [60] corresponds to a barrier of 64.4±0.5 kJ mol−164.4\pm0.5\,\mathrm{kJ\,mol^{-1}} using transition state theory with transmission coefficient one and uncertainty propagated from the rate. Our barrier exceeds this rate-derived estimate by 20.8 kJ mol−120.8\,\mathrm{kJ\,mol^{-1}}. For comparison, Gao and co-workers [66] report discrepancies ranging from 3.84 kJ mol−13.84\,\mathrm{kJ\,mol^{-1}} for poly(ethylene terephthalate) hydrolase to 17.5 kJ mol−117.5\,\mathrm{kJ\,mol^{-1}} for nucleoside diphosphate kinase, the latter approaching the discrepancy observed here (Table 1). Their calculations use the pretrained eSEN-omol potential, trained at the same ω\omegaB97M-V/def2-TZVPD reference level, to describe both enzyme and solvent. They also use the Eyring relation with transmission coefficient one to relate rates and barriers, although the experimental estimate for nucleoside diphosphate kinase comes from a turnover rate and represents an upper bound on its chemical barrier.

SystemCalculatedRate-derivedDifference
DhlA, this work85.2±0.485.2\pm0.464.4±0.564.4\pm0.520.820.8
Chorismate mutase75.2±2.7275.2\pm2.7264.464.410.810.8
79.0±4.2079.0\pm4.2075.2±0.475.2\pm0.43.843.84
73.1±1.8673.1\pm1.8655.655.617.517.5

Table 1. Calculated and rate-derived experimental barriers in kJ/mol, with calculated minus experimental differences obtained from unrounded values. DhlA uncertainties are a block-jackknife standard error and propagated rate uncertainty. Other uncertainties are reproduced from Gao et al. [66]; for poly(ethylene terephthalate) hydrolase, the calculated uncertainty is the half-width of the reported 95% interval. The nucleoside diphosphate kinase experimental estimate comes from a turnover rate and is an upper bound on the chemical barrier.

Beyond the barrier height, the shape of our free energy profile differs from the reference, with a sharp change in slope near the maximum (Figure 3A). To examine sampling near this feature, Figure 3B shows the evolution of the reaction coordinate and its production distribution in each umbrella. The change in slope coincides with reduced overlap between neighbouring umbrellas: the overlap between those centred at 0.125 and 0.175 Å is 0.34, compared with a median of 0.62 across neighbouring pairs. Here, overlap is calculated as Oij=∑bmin⁡(pib,pjb)O_{ij}=\sum_b\min(p_{ib},p_{jb}), where pibp_{ib} is the fraction of production samples from umbrella ii in a common bin of width 0.02 Å. This correlation suggests that insufficient overlap could introduce a systematic error in the reconstructed profile, although testing this possibility requires another set of umbrella sampling simulations.

Conclusions

We have presented GPU infrastructure for electrostatic machine learning/molecular mechanics simulations of enzyme reactions. This combines a JAX port of the pretrained MACE-POLAR-1-M model, an embedding scheme that allows the classical environment to polarize the learned region, and integration with JAX-MD and PLUMED-JAX for compiled enhanced sampling. Numerical comparisons quantify agreement with the original model, while timings demonstrate faster model evaluation and molecular dynamics for the tested systems.

Applied to nucleophilic substitution in DhlA, this framework enables replica exchange umbrella sampling with the entire protein and reactants in the learned region and most solvent described by molecular mechanics, at a total cost of approximately 200 H100 GPU hours. While the free energy estimates presented here are preliminary, their deviation from experiment is comparable in scale to discrepancies reported in other enzyme simulations using pretrained machine learning potentials [66, 57]. This computationally efficient infrastructure opens opportunities for systematic sampling and comparisons of how protein conformations and environmental polarization influence enzyme reactions.

Data and software availability

The data supporting this report will be made available at a later stage.

Acknowledgments

The author thanks members of the Initiative for Computational Catalysis and the Center for Computational Biology at the Flatiron Institute for insightful discussions. Computational resources were provided by the Flatiron Institute, a division of the Simons Foundation.

References

  1. [34]Senn, & Thiel. (2009). QM/MM methods for biomolecular systems.
  2. [35]Li, & Chan. (2025). Accurate QM/MM molecular dynamics for periodic systems in GPU4PySCF with applications to enzyme catalysis.
  3. [36]Batatia, Baldwin, Kuryla, Hart, Kasoar, Elena, Moore, Gawkowski, Shi, Kapil, Kourtis, Magdău, & Csányi. (2026). MACE-POLAR-1 : A polarisable electrostatic foundation model for molecular chemistry, 2026. https://arxiv.org/abs/2602.19411
  4. [37]Grisafi, & Ceriotti. (2019). Incorporating long-range physics in atomic-scale machine learning.
  5. [38]Schoenholz, & Cubuk. (2020). JAX MD : A framework for differentiable physics. https://proceedings.neurips.cc/paper/2020/hash/83d3d4b6c9579515e1679aca8cbc8033-Abstract.htmlproceedings.neurips.cc/paper/2020/hash/83d3d4b6c9579515e1679aca8cbc8033-Abstract.html
  6. [39]Venkat Kapil. PLUMED-JAX . \newblock Software repository, 2026. \newblock URL https://github.com/venkatkapil24/plumed-jax.github.com/venkatkapil24/plumed-jax
  7. [40]Tribello, Bonomi, Branduardi, Camilloni, & Bussi. (2014). PLUMED 2: New feathers for an old bird.
  8. [41]The PLUMED consortium. (2019). Promoting transparency and reproducibility in enhanced molecular simulations.
  9. [42]Bradbury, Frostig, Hawkins, Johnson, Katariya, Leary, Maclaurin, Necula, Paszke, VanderPlas, Wanderman-Milne, & Zhang. (2018). JAX : composable transformations of python+numpy programs. https://github.com/jax-ml/jaxgithub.com/jax-ml/jax
  10. [43]Levine, Shuaibi, Spotte-Smith, Taylor, Hasyim, Michel, Batatia, Csányi, Dzamba, Eastman, Frey, Fu, Gharakhanyan, Krishnapriyan, Rackers, Raja, Rizvi, Rosen, Ulissi, … Wood. (2025). The open molecules 2025 ( OMol25 ) dataset, evaluations, and models, 2025. https://arxiv.org/abs/2505.08762
  11. [44]Batatia, Kovács, Simm, Ortner, & Csányi. (2022). MACE : Higher order equivariant message passing neural networks for fast and accurate force fields. https://papers.neurips.cc/paper_files/paper/2022/hash/4a36c3c51af11ed9f34615b81edb5bbc-Abstract-Conference.htmlpapers.neurips.cc/paper_files/paper/2022/hash/4a36c3c51af11ed9f34615b81edb5bbc-Abstract-Conference.html
  12. [45]NVIDIA . cuEquivariance . \newblock Software repository, n.d. \newblock URL https://github.com/NVIDIA/cuEquivariance. \newblock Accessed 30 October 2026.github.com/NVIDIA/cuEquivariance
  13. [46]Paszke, Gross, Massa, Lerer, Bradbury, Chanan, Killeen, Lin, Gimelshein, Antiga, Desmaison, Köpf, Yang, DeVito, Raison, Tejani, Chilamkurthy, Steiner, Fang, … Chintala. (2019). PyTorch : An imperative style, high-performance deep learning library. https://proceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-Abstract.htmlproceedings.neurips.cc/paper/2019/hash/bdbca288fee7f92f2bfa9f7012727740-Abstract.html
  14. [47]Hickey, & Rowley. (2014). Benchmarking quantum chemical methods for the calculation of molecular dipole moments and polarizabilities. https://doi.org/10.1021/jp502475e}
  15. [48]Liang(n.d.). (2026). Gold-standard chemical database: input archive. https://github.com/JiashuLiang/GSCDBgithub.com/JiashuLiang/GSCDB
  16. [49]Larsen, Mortensen, Blomqvist, Castelli, Christensen, Dułak, Friis, Groves, Hammer, Hargus, Hermes, Jennings, Jensen, Kermode, Kitchin, Kolsbjerg, Kubal, Kaasbjerg, Lysgaard, … Jacobsen. (2017). The atomic simulation environment---a Python library for working with atoms.
  17. [50]Cohen, Riebesell, Goodall, Kolluru, Falletta, Krause, Colindres, Ceder, & Gangan. (2025). TorchSim : An efficient atomistic simulation engine in PyTorch.
  18. [51]Humphrey, Dalke, & Schulten. (1996). VMD : Visual molecular dynamics.
  19. [52]Sun, Li, Cui, & Xie. (2026). Evaluating mechanical-embedding ML/MM for predicting mutation effects in chorismate mutase catalysis.
  20. [53]Zinovjev. (2023). Electrostatic embedding of machine learning potentials.
  21. [54]Gradisteanu, Chan, Hedges, Malagarriga, David, de la Puente, Laage, Tuñón, van der Kamp, & Zinovjev. (2026). Simulating enzyme catalysis with electrostatically embedded machine learning potentials.
  22. [55]Morado, Zinovjev, Hedges, Cole, & Michel. (2025). Enhancing electrostatic embedding for ML/MM free energy calculations.
  23. [56]Lei, Yagi, & Sugita. (2025). Efficient training of neural network potentials for chemical and enzymatic reactions by continual learning.
  24. [57]Thürlemann, Pultar, Gordiy, Ruijsenaars, & Riniker. (2026). Multiscale neural network potential with anisotropic message passing for the fast and accurate simulation of protein dynamics and enzymatic reactions.
  25. [58]Essmann, Perera, Berkowitz, Darden, Lee, & Pedersen. (1995). A smooth particle mesh Ewald method.
  26. [59]Plotnikov. (2014). Computing the free energy barriers for less by sampling with a coarse reference potential while retaining accuracy of the target fine model.
  27. [60]Schanstra, Kingma, & Janssen. (1996). Specificity and kinetics of haloalkane dehalogenase.
  28. [61]Jorgensen, Chandrasekhar, Madura, Impey, & Klein. (1983). Comparison of simple potential functions for simulating liquid water.
  29. [62]Rowley, & Roux. (2012). The solvation structure of Na+ and K+ in liquid water determined from high level ab initio molecular dynamics simulations.
  30. [63]Miyamoto, & Kollman. (1992). SETTLE : An analytical version of the SHAKE and RATTLE algorithm for rigid water models.
  31. [64]Sugita, Kitao, & Okamoto. (2000). Multidimensional replica-exchange method for free-energy calculations.
  32. [65]Shirts, & Chodera. (2008). Statistically optimal analysis of samples from multiple equilibrium states.
  33. [66]Gao, Shayesteh Zadeh, Seal, Dasetty, Achar, Dzamba, Miller, Jacobson, Zitnick, Wood, Ulissi, Levine, & Ferguson. (2026). Quantum-accurate atomistic modeling of enzyme catalysis using a machine learned potential, 2026. https://arxiv.org/abs/2609.09293v2

Paper details

vlnc:W_zts6av87
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 15

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Codex

View statement
Looking for related papers…
Contents