Pith. sign in

REVIEW 3 major objections 5 minor 71 references

NepoIP/MM: Towards Accurate Biomolecular Simulation with a Machine Learning/Molecular Mechanics Model Incorporating Polarization Effects

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper shows that a machine-learned potential, NepoIP, which uses the external electrostatic potential as an input, can reproduce QM/MM electrostatic-embedding dynamics for alanine dipeptide in water.

desk verdict A credible step toward polarizable ML/MM with stable dipeptide MD, but the descriptor sufficiency and protein-transfer claims are shakier than the paper lets on. read the letter →

arxiv 2502.02801 v2 pith:UW45X4YT submitted 2025-02-05 physics.chem-ph

Ge Song , Weitao Yang This is my paper · ORCID
classification physics.chem-ph
keywords machinelearningforcefieldML/MMelectrostaticembeddingpolarizationexternalpotentialequivariantneuralnetworkNequIPalaninedipeptidemoleculardynamics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper develops a machine-learned force field, NepoIP, that adds polarization to a hybrid machine learning/molecular mechanics (ML/MM) model by feeding each quantum-region atom the electrostatic potential generated by the surrounding MM point charges. Trained on QM/MM energies with electrostatic embedding minus the MM energy, the model reproduces the reference potential-energy surface of solvated peptides and drives nanosecond-scale molecular dynamics that remains stable. The resulting Ramachandran distribution and secondary-structure populations for alanine dipeptide in water match direct QM/MM electrostatic-embedding simulations, and a single model transfers across different water models and between water and protein environments. The claim matters because it offers a route to QM/MM-level biomolecular dynamics at much lower cost, provided the electrostatic-potential descriptor captures the relevant polarization.

What carries the argument

The central object is NepoIP, a modification of the E(3)-equivariant NequIP graph neural network. Its polarization block takes the external electrostatic potential from MM point charges, one scalar per ML atom, embeds it like atomic numbers, and mixes it with neighboring atomic features through a tensor product; the resulting environment-dependent features let the network learn charge fluctuation beyond linear response. The companion machinery is a delta-machine-learning decomposition: the network predicts the difference between QM/MM electrostatic-embedding energy and the pure MM energy of the whole system, so the MM force field supplies large penalties that keep unphysical configurations from collapsing the simulation.

What would settle it

Compare NepoIP/MM energies and forces with QM/MM-EE reference on configurations where the on-atom electrostatic potentials are small but their spatial gradients are large, such as a solute inside a protein cavity with ordered dipoles or in a strong applied field gradient. If the reference polarization changes significantly while the scalar-potential input is near zero, the descriptor is incomplete and the model would fail; alternatively, running a model trained only on water potentials in a protein core outside the training ESP range should produce growing force error and unstable sampling.

Watch

Extended reading notes

Core claim

On its own terms, the paper establishes that the polarization of the QM region by the MM environment can be learned rather than derived: the external electrostatic potential per ML atom, computed from fixed MM partial charges, serves as a low-dimensional environment descriptor whose inclusion in an equivariant graph network brings ML/MM energies and forces into close agreement with QM/MM electrostatic embedding. With a delta-machine-learning decomposition, where the network predicts only the difference between QM/MM and MM energies, the total ML/MM potential stays stable in repeated 2 ns simulations and its converged conformational ensemble agrees with the QM/MM-EE reference. The same trained model also works when the water model is swapped and, after training on a merged water/protein dataset, covers extremely different electrostatic environments, supporting the prospect of a general polarizable ML/MM biomolecular force field.

Load-bearing premise

The load-bearing premise is that the scalar external electrostatic potential at each ML atom, from fixed MM point charges, carries enough information for the network to learn the environment's polarization effect; if polarization also depends on field gradients, charge transfer, or short-range exchange, the model could miss them and transferability across very different environments could break.

Editorial extensions

If this is right

  • NepoIP/MM reproduces QM/MM electrostatic-embedding energies for alanine dipeptide in water with test-set energy RMSE below 0.1 kcal/mol on the larger training datasets.
  • Nanosecond-scale NepoIP/MM simulations remain stable without collapse, and the converged ensemble-averaged 3J coupling (8.078 Hz) matches the QM/MM-EE reference (8.065 Hz).
  • A model trained with the TIP3P water model transfers directly to TIP3P-FB and OPC3 water models with essentially unchanged energy and force errors.
  • After training on a merged water-plus-protein dataset, one NepoIP model achieves errors comparable to environment-specific models in both the water and protein test sets, showing that coverage of the electrostatic-potential distribution is the key to transferability.
  • Because the MM background enters only through its electrostatic potential, the same ML model can be paired with different MM force fields without retraining.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The scalar-potential descriptor implies the model cannot distinguish different MM charge distributions that produce the same potential on the ML atoms; if polarization in larger biomolecules depends on field gradients or charge transfer, additional descriptors may be needed.
  • The merged-dataset result suggests a construction recipe for a general protein force field: fragment-based electrostatic-embedding datasets that deliberately span water and protein electrostatic-potential ranges, a step the paper only demonstrates for a single dipeptide.
  • The simulation bottleneck is the periodic electrostatic-potential computation, not the network evaluation, so adopting faster Ewald variants could make polarizable ML/MM-EE dynamics routine for much larger systems.
  • Beyond peptides, the same delta-learning-plus-electrostatic-potential design may apply to enzyme active sites, where solute polarization by the environment is chemically important.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The manuscript presents NepoIP, an E(3)-equivariant neural network potential based on NequIP that augments each ML atom with the external electrostatic potential from the MM environment as a scalar node attribute, and embeds it in a delta-machine-learning ML/MM scheme in which the network learns the difference between QM/MM-EE and MM energies. The authors train NepoIP on alanine dipeptide in water and in a designed protein core, report low held-out energy/force errors for the water system, stable 8x2 ns MD for solvated alanine dipeptide with Ramachandran and 3J(HNHA) distributions close to the QM/MM-EE reference, and demonstrate transferability across TIP3P, TIP3P-FB, and OPC3 water models and between water and protein environments. Code and datasets are made available on GitHub.

Significance. If the core claims hold, NepoIP/MM is a useful advance toward electrostatic-embedding ML/MM for biomolecular simulation: it shows that an equivariant network with a low-dimensional ESP descriptor can drive stable nanosecond MD reproducing a QM/MM-EE reference, and the open-source implementation is a concrete resource for the community. The paper's strengths include internally consistent force formulas derived from the energy model, held-out test sets, explicit transferability tests across water models, stable MD demonstrations, and public code/data. The main caveats are that the reference QM level is semi-empirical DFTB, the MD comparison is a self-consistency check against the labeling level rather than an external accuracy benchmark, and the protein-environment transferability evidence is considerably weaker than the water evidence.

major comments (3)
  1. [Section 2.1, Eq. (1); Section 3.4, Table 2] The network's only environmental descriptor is the scalar external electrostatic potential V_i evaluated at each ML nucleus. The target E_ML in Eq. (4) contains E_QM(QM) + E_QM/MM_elec, which is a functional of the full external potential over the QM electron density, not just its values at a finite set of nuclear positions. Two MM charge arrangements can produce identical {V_i} at all ML atoms while differing in electric field and field gradient, and therefore induce different polarization; a model that sees only {V_i} cannot represent those differences. This is not merely a formal concern: Table 2 shows that the merged-model protein energy RMSE is 1.377 kcal/mol versus 0.255 kcal/mol in water, and no protein-environment MD is reported. The dipeptide-in-water validation samples a narrow ESP range and cannot detect this descriptor incompleteness. The general-transferability claim in the abstract therefore rests on an assumption that is theoretically incomplete and empirically under-tested. I recommend adding ESP gradients or field components as descriptors, or providing a direct test with matched scalar ESP but different field/gradient distributions.
  2. [Section 3.4, Table 2] The merged model's training RMSE of 0.975 kcal/mol is far larger than the separate water model RMSEs, and its protein test RMSE of 1.377 kcal/mol is high in absolute terms for a potential intended for molecular dynamics. The conclusion that a single model is transferable to 'extremely different MM environments' is supported only by energy/force errors on one protein test set; there is no MD simulation in the protein environment and no test on other protein sites. Please provide either a protein-environment MD test that demonstrates stable sampling, or substantially relax the transferability claim and state a quantitative criterion for what 'transferable' means in terms of force or energy error thresholds.
  3. [Section 2.2, Eq. (7)] The ML correction forces on MM atoms are obtained by backpropagating through the ESP descriptor, but the model is not trained on these forces and no validation of MM-atom forces against QM/MM-EE is reported. Since the solvent and protein dynamics in NepoIP/MM depend on these forces, the quality of the total MM forces is a load-bearing unknown. Please report force RMSEs on MM atoms for at least the water test set, and preferably for protein atoms as well, or explain why training on energies and QM-region forces alone is sufficient to guarantee accurate ESP derivatives for the MM region.
minor comments (5)
  1. [Section 2.3] Please clarify the peptide-in-protein dataset construction: conformations are sampled from the solvated protein system, but the reference QM/MM energies are computed after removing all solvent and ions. If this mismatch is intentional, state the rationale and discuss how it affects the ESP distributions used for training and testing.
  2. [Section 3.2] The main text does not state which training dataset was used for the 8x2 ns MD simulations; it only becomes inferable from Fig. S2 that a 144k umbrella-sampling dataset was involved. Please state the exact model and training set in Section 2.4 or 3.2.
  3. [Section 3.2] The phrase 'nanoscale time period' should be 'nanosecond time scale'.
  4. [Section 3.4, Table 2 text] In the paragraph discussing Table 2, 'peptide-in-solvent system' appears to be a typo for 'peptide-in-protein system'.
  5. [Section 4.1] The claim of an 'overall energy RMSE of less than 0.1 kcal/mol' should specify the dataset size used for that comparison; Fig. 3 reports 0.179 kcal/mol for the 5k dataset, while Table S2 gives 0.0741 and 0.0595 for 50k and 100k, respectively.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the trained surrogate is validated on held-out conformations and independent MM environments, and the cited prior work is not load-bearing.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. NepoIP is a delta-machine-learning model trained on QM/MM-EE reference energies and forces (Eqs. 2-5), and the central validations are held-out energy/force tests, nanosecond MD stability, and Ramachandran/J-coupling agreement with direct QM/MM-EE simulations. Agreement with the training level of theory is a standard surrogate-model validation, not a statistically forced prediction, because the MD comparison involves new conformations sampled dynamically and the transferability tests use MM water models and a protein environment not seen in training. The self-citations to the Yang group's earlier work (refs. 37, 38, and 31) motivate the external-electrostatic-potential descriptor and the delta-learning strategy, but the paper's own comparisons against a non-polarizable NequIP baseline and against mechanical embedding provide independent empirical support for those choices. The reviewer concern that scalar ESP is an incomplete polarization descriptor is a correctness or transferability risk, not a circularity: it does not mean the energy prediction is equivalent to the input by construction. Overall, no load-bearing step relies on a self-citation chain or on a renamed input, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central fitted object is the neural network itself; the key modeling axiom is that the ESP is a sufficient descriptor. No physical entities are invented.

free parameters (3)
  • Neural network weights of NepoIP = not specified
    The central ML energy E_ML is a neural network trained on QM/MM reference energies and forces; all prediction results depend on this fit.
  • Ewald error tolerance epsilon = 0.005
    Chosen in Section 2.4 based on a trade-off between speed and force RMSE (Fig. S2); affects the accuracy of the external electrostatic potential input.
  • Direct space cutoff distance for Ewald summation = 9 Angstrom (MD), 10 Angstrom (Amber sampling)
    User-specified in OpenMM; used in computing the electrostatic potential descriptor.
assumptions (5)
  • domain assumption The external electrostatic potential from MM point charges at each ML atom is a sufficient descriptor for the polarization of the QM region.
    Central modeling hypothesis introduced in Section 2.1; the model only receives ESP values as environmental input, so any polarization effects not captured by atomic ESP (e.g., field gradients, charge transfer, exchange) are assumed negligible.
  • domain assumption DFTB with dispersion correction is an adequate reference QM/MM level for the alanine dipeptide systems studied.
    All training and reference MD uses SCC-DFTB as the QM theory (ref 51); the claimed 'quantum mechanical accuracy' is relative to this semi-empirical level.
  • standard math E(3)-equivariance is preserved when a scalar external potential is added as a node attribute.
    Stated in Section 2.1; standard property of equivariant networks when input is a scalar per-node feature.
  • standard math The delta-machine-learning energy decomposition (Eqs. 2-5) is an exact identity for electrostatic embedding when the MM energy functions are fixed.
    The decomposition follows from definitions; used to define the training target E_ML.
  • domain assumption MM force field fixed partial charges are adequate to generate the external electrostatic potential.
    The ESP descriptor is computed from fixed point charges of the MM force field; transferability is claimed across water models with different charges, but the model was only tested on these three.

how reviews work

0 comments
Cite this review

Pith. "Pith review of NepoIP/MM: Towards Accurate Biomolecular Simulation with a Machine Learning/Molecular Mechanics Model Incorporating Polarization Effects." pith.science (2026). https://pith.science/paper/UW45X4YT

@misc{pith2026250202801,
  author       = {Pith},
  title        = {Pith review of: NepoIP/MM: Towards Accurate Biomolecular Simulation with a Machine Learning/Molecular Mechanics Model Incorporating Polarization Effects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UW45X4YT}},
  note         = {Machine review of arXiv:2502.02801}
}
read the original abstract

Machine learning force fields offer the ability to simulate biomolecules with quantum mechanical accuracy while significantly reducing computational costs, attracting growing attention in biophysics. Meanwhile, leveraging the efficiency of molecular mechanics in modeling solvent molecules and long-range interactions, a hybrid machine learning/molecular mechanics (ML/MM) model offers a more realistic approach to describing complex biomolecular systems in solution. However, multiscale models with electrostatic embedding require accounting for the polarization of the ML region induced by the MM environment. To address this, we adapt the state-of-the-art NequIP architecture into a polarizable machine learning force field, NepoIP, enabling the modeling of polarization effects based on the external electrostatic potential. We found that the nanosecond MD simulations based on NepoIP/MM are stable for the periodic solvated dipeptide system and the converged sampling shows excellent agreement with the reference QM/MM level. Moreover, we show that a single NepoIP model can be transferable across different MM force fields, as well as extremely different MM environment of water and proteins, laying the foundation for developing a general machine learning biomolecular force field to be used in ML/MM with electrostatic embedding.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 68 canonical work pages

  1. [1]

    O.; Shaw, D

    Lindorff - Larsen, K.; Piana, S.; Dror, R. O.; Shaw, D. E., How Fast - Folding Proteins Fold. Science 2011 , 334, 517 - 520

    work page 2011

  2. [2]

    ML non - polarizable force fields trained to the QM/MM energy including solute - solvent interaction . Th is draws inspiration from the strategy in work (28) , where NequIP was trained against the pure MM energy of the solute and its interactions with the solvent, implicitly capturing solvation effects within a force field. In our case of QM/MM , the Nequ...

    work page

  3. [3]

    The reference QM theory was at the semi - empirical DFTB level with dispersion correction (51) , with TIP3P water model (52) for the MM region

    Results 3.1 Reproduction of the QM/MM Potential Energy We begin by comparing the accuracy of NepoIP to non - polarizable models in reproducing the QM /MM energy of peptide - in - water . The reference QM theory was at the semi - empirical DFTB level with dispersion correction (51) , with TIP3P water model (52) for the MM region . The reference energy incl...

    work page

  4. [4]

    The results shown in Fig

    The polarizable force field NepoIP , which explicitly treats both the sol ute - solvent interaction and polarization effects. The results shown in Fig. 3 indicate that the upper limit of ML/MM with mechanical embedding is even worse than implicitly learning all the solvation effects by NequIP. Meanwhile, NepoIP significantly reduced the errors of NequIP, ...

    work page

  5. [5]

    thus define our ∆ machine learning approach for ML/MM electrostatic embedding. Under this potential, the ML force on an ML atom is: ( ) , = ( , ) , + ( , ) , , , ∈ , ( 6 ) where represents the coordinates of the whole system, and represent the coordinates and electrostatic potentials on the ML atoms respectively. T he first term disregards the dependence ...

    work page

  6. [6]

    K.; Greenwood, J.; Romero, D

    Wang, L.; Wu, Y .; Deng, Y .; Kim, B.; Pierce, L.; Krilov, G.; Lupyan, D.; Robinson, S.; Dahlgren, M. K.; Greenwood, J.; Romero, D. L.; Masse, C.; Knight, J. L.; Steinbrecher, T.; Beuming, T.; Damm, W.; Harder, E.; Sherman, W.; Brewer, M.; Wester, R.; Mur cko, M.; Frye, L.; Farid, R.; Lin, T.; Mobley, D. L.; Jorgensen, W. L.; Berne, B. J.; Friesner, R. A....

    work page 2015

  7. [7]

    A., Molecular dynamics simulations of biomolecules

    Karplus, M.; McCammon, J. A., Molecular dynamics simulations of biomolecules. Nat. struct. biol. 2002 , 9, 646 - 652

    work page 2002

  8. [8]

    A.; McCammon, J

    Adcock, S. A.; McCammon, J. A., Molecular Dynamics: Survey of Methods for Simulating the Activity of Proteins. Chem. Rev. 2006 , 106, 1589 - 1615. 13

    work page 2006

Show all 71 references
  1. [9]

    M., Equilibrium Sampling in Biomolecular Simulations

    Zuckerman, D. M., Equilibrium Sampling in Biomolecular Simulations. Annu. Rev. Biophys. 2011 , 40, 41 - 62

    2011

  2. [10]

    A.; Dror, R

    Hollingsworth, S. A.; Dror, R. O., Molecular Dynamics Simulation for All. Neuron 2018 , 99, 1129 - 1143

    2018

  3. [11]

    F.; Daura, X.; Hansen, N.; Mark, A

    van Gunsteren, W. F.; Daura, X.; Hansen, N.; Mark, A. E.; Oostenbrink, C.; Riniker, S.; Smith, L. J., Validation of Molecular Simulation: An Overview of Issues. Angew. Chem. Int. Ed. 2018 , 57, 884 - 902

    2018

  4. [12]

    Nature 2024 , 627, 905 - 914

    Coshic, K.; Maffeo, C.; Winogradoff, D.; Aksimentiev, A., The structure and physical properties of a packaged bacteriophage particle. Nature 2024 , 627, 905 - 914

    2024

  5. [13]

    A.; Kanekiyo, M.; Amaro, R

    Casalino, L.; Seitz, C.; Lederhofer, J.; Tsybovsky, Y .; Wilson, I. A.; Kanekiyo, M.; Amaro, R. E., Breathing and Tilting: Mesoscale Simulations Illuminate Influenza Glycoprotein Vulnerabilities. ACS Cent. Sci 2022 , 8, 1646 - 1663

    2022

  6. [14]

    E.; Maragakis, P.; Lindorff - Larsen, K.; Piana, S.; Dror, R

    Shaw, D. E.; Maragakis, P.; Lindorff - Larsen, K.; Piana, S.; Dror, R. O.; Eastwood, M. P .; Bank, J. A.; Jumper, J. M.; Salmon, J. K.; Shan, Y .; Wriggers, W., Atomic - Level Characterization of the Structural Dynamics of Proteins. Science 2010 , 330, 341 - 346

    2010

  7. [15]

    P .; Dror, R

    Lindorff - Larsen, K.; Maragakis, P.; Piana, S.; Eastwood, M. P .; Dror, R. O.; Shaw, D. E., Systematic Validation of Protein Force Fields against Experimental Data. PLOS ONE 2012 , 7, e32131

    2012

  8. [16]

    S.; Head - Gordon, T., New developments in force fields for biomolecular simulations

    Nerenberg, P. S.; Head - Gordon, T., New developments in force fields for biomolecular simulations. Curr . Opin. Struct. Biol. 2018 , 49, 129 - 138

    2018

  9. [17]

    U.; Chen, H

    Song, G.; Zhong, B.; Zhang, B.; Rehman, A. U.; Chen, H. - F., Phosphorylation Modification Force Field FB18CMAP Improving Conformation Sampling of Phosphoproteins. J. Chem. Inf. Model. 2023 , 63, 1602 - 1614

    2023

  10. [18]

    - F., Comprehensive Comparison and Critical Assessment of RNA - Specific Force Fields

    Choi, T.; Li, Z.; Song, G.; Chen, H. - F., Comprehensive Comparison and Critical Assessment of RNA - Specific Force Fields. J. Chem. Theory Comput. 2024 , 20, 2676 - 2688

    2024

  11. [19]

    M.; Thiel, W., QM/MM methods for biomolecular systems

    Senn, H. M.; Thiel, W., QM/MM methods for biomolecular systems. Angew. Chem. Int. Ed. 2009 , 48, 1198 - 1229

    2009

  12. [20]

    Hu, H.; Yang, W., Free energies of chemical reactions in solution and in enzymes with ab initio quantum mechanics/molecular mechanics methods. Annu. Rev. Phys. Chem. 2008 , 59, 573 - 601

    2008

  13. [21]

    L., Advances in quantum and molecular mechanical (QM/MM) simulations for organic and enzymatic reactions

    Acevedo, O.; Jorgensen, W. L., Advances in quantum and molecular mechanical (QM/MM) simulations for organic and enzymatic reactions. Acc. Chem. Res. 2009 , 43, 142 - 151

    2009

  14. [22]

    Botu, V .; Batra, R.; Chapman, J.; Ramprasad, R., Machine Learning Force Fields: Construction, Validation, and Outlook. J. Phys. Chem. C 2017 , 121, 511 - 522

    2017

  15. [23]

    Cheng, Z.; Du, J.; Zhang, L.; Ma, J.; Li, W.; Li, S., Building quantum mechanics quality force fields of proteins with the generalized energy - based fragmentation approach and machine learning. Phys. Chem. Chem. Phys. 2022 , 24, 1326 - 1337

    2022

  16. [24]

    S.; Isayev, O.; Roitberg, A

    Smith, J. S.; Isayev, O.; Roitberg, A. E., ANI - 1: an extensible neural network potential with DFT accuracy at force field computational cost. Chem. Sci. 2017 , 8, 3192 - 3203

    2017

  17. [25]

    T.; Sauceda, H

    Schütt, K. T.; Sauceda, H. E.; Kindermans, P. J.; Tkatchenko, A.; Müller, K. R., SchNet – A deep learning architecture for molecules and materials. J. Chem. Phys. 2018 , 148, 241722

    2018

  18. [26]

    P.; Kornbluth, M.; Molinari, N.; Smidt, T

    Batzner, S.; Musaelian, A.; Sun, L.; Geiger, M.; Mailoa, J. P.; Kornbluth, M.; Molinari, N.; Smidt, T. E.; Kozinsky, B., E(3) - equivariant graph neural networks for data - efficient and accurate interatomic potentials. Nat. Commun 2022 , 13, 2453

    2022

  19. [27]

    Raja, S.; Amin, I.; Pedregosa, F.; Krishnapriyan, A. S. Stability - Aware Training of Neural Network Interatomic Potentials with Differentiable Boltzmann Estimators. arXiv:2402.13984, 2024

    2024 arXiv

  20. [28]

    - Y ., Enhancing geometric representations for molecules with equivariant vector - scalar interactive message passing

    Wang, Y .; Wang, T.; Li, S.; He, X.; Li, M.; Wang, Z.; Zheng, N.; Shao, B.; Liu, T. - Y ., Enhancing geometric representations for molecules with equivariant vector - scalar interactive message passing. Nat. Commun 2024 , 15, 313

    2024

  21. [29]

    Zhang, L.; Han, J.; Wang, H.; Car, R.; E, W., Deep Potential Molecular Dynamics: A Scalable Model with the Accuracy of Quantum Mechanics. Phys. Rev. Lett. 2018 , 120, 143001. 14

    2018

  22. [30]

    O., Quantum Chemistry in the Age of Machine Learning

    Dral, P . O., Quantum Chemistry in the Age of Machine Learning. J. Phys. Chem. Lett. 2020 , 11, 2336 - 2347

    2020

  23. [31]

    S.; Wieder, M.; Xu, Y .; Zhu, T.; Zhang, J

    Wang, Y .; Takaba, K.; Chen, M. S.; Wieder, M.; Xu, Y .; Zhu, T.; Zhang, J. Z. H.; Nagle, A.; Yu, K.; Wang, X.; Cole, D. J.; Rackers, J. A.; Cho, K.; Greener, J. G.; Eastman, P .; Martiniani, S.; Tuckerman, M. E. On the design space between molecular mech anics and machine lea...

    2024 arXiv

  24. [32]

    arXiv:2210.07237 2022

    Fu, X.; Wu, Z.; Wang, W.; Xie, T.; Keten, S.; Gomez - Bombarelli, R.; Jaakkola, T., Forces are not enough: Benchmark and critical evaluation for machine learning force fields with molecular simulations. arXiv:2210.07237 2022

    2022 arXiv

  25. [33]

    S.; Pickering, I.; Roitberg, A

    Grassano, J. S.; Pickering, I.; Roitberg, A. E.; González Lebrero, M. C.; Estrin, D. A.; Semelak, J. A., Assessment of Embedding Schemes in a Hybrid Machine Learning/Classical Potentials (ML/MM) Approach. J. Chem. Inf. Model. 2024 , 64, 4047 - 4058

    2024

  26. [34]

    E.; Chodera, J

    Galvelis, R.; Varela - Rial, A.; Doerr, S.; Fino, R.; Eastman, P.; Markland, T. E.; Chodera, J. D.; De Fabritiis, G., NNP/MM: Accelerating Molecular Dynamics Simulations with Machine Learning Potentials and Molecular Mechanics. J. Chem. Inf. Model. 2023 , 63, 5701 - 5708

    2023

  27. [35]

    Zhang, P .; Yang, W., Toward a general neural network force field for protein simulations: Refining the intramolecular interaction in protein. J. Chem. Phys. 2023 , 159

    2023

  28. [36]

    I.; Cieplak, P.; Cornell, W.; Kollman, P

    Bayly, C. I.; Cieplak, P.; Cornell, W.; Kollman, P. A., A well - behaved electrostatic potential based method using charge restraints for deriving atomic charges: the RESP model. J. Phys. Chem. 1993 , 97, 10269 - 10280

    1993

  29. [37]

    Bakowies, D.; Thiel, W., Hybrid Models for Combined Quantum Mechanical and Molecular Mechanical Approaches. J. Phys. Chem. 1996 , 100, 10580 - 10594

    1996

  30. [38]

    G., QM/MM: what have we learned, where are we, and where do we go from here? Theor

    Lin, H.; Truhlar, D. G., QM/MM: what have we learned, where are we, and where do we go from here? Theor. Chem. Acc. 2007 , 117, 185 - 199

    2007

  31. [39]

    J.; Ekesan, Ş.; York, D

    Zeng, J.; Giese, T. J.; Ekesan, Ş.; York, D. M., Development of Range - Corrected Deep Learning Potentials for Fast, Accurate Quantum Mechanical/Molecular Mechanical Simulations of Chemical Reactions in Solution. J. Chem. Theory Comput. 2021 , 17, 6993 - 7009

    2021

  32. [40]

    J.; Zeng, J.; Ekesan, Ş.; York, D

    Giese, T. J.; Zeng, J.; Ekesan, Ş.; York, D. M., Combined QM/MM, Machine Learning Path Integral Approach to Compute Free Energy Profiles and Kinetic Isotope Effects in RNA Cleavage Reactions. J. Chem. Theory Comput. 2022 , 18, 4304 - 4317

    2022

  33. [41]

    Shen, L.; Yang, W., Molecular Dynamics Simulations with Quantum Mechanics/Molecular Mechanics and Adaptive Neural Networks. J. Chem. Theory Comput. 2018 , 14, 1442 - 1455

    2018

  34. [42]

    Shen, L.; Wu, J.; Yang, W., Multiscale Quantum Mechanics/Molecular Mechanics Simulations with Neural Networks. J. Chem. Theory Comput. 2016 , 12, 4934 - 4946

    2016

  35. [43]

    T.; Müller, K

    Gastegger, M.; Schütt, K. T.; Müller, K. - R., Machine learning of solvent effects on molecular spectra and reactions. Chem. Sci. 2021 , 12, 11473 - 11483

    2021

  36. [44]

    Pan, X.; Yang, J.; Van, R.; Epifanovsky, E.; Ho, J.; Huang, J.; Pu, J.; Mei, Y .; Nam, K.; Shao, Y ., Machine - Learning - Assisted Free Energy Simulation of Solution - Phase and Enzyme Reactions. J. Chem. Theory Comput. 2021 , 17, 5745 - 5758

    2021

  37. [45]

    Neural Network Potential with Multi - Resolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution

    Pultar, F.; Thuerlemann, M.; Gordiy, I.; Doloszeski, E.; Riniker, S. Neural Network Potential with Multi - Resolution Approach Enables Accurate Prediction of Reaction Free Energies in Solution. arXiv:2411.19728, 2024 (accessed November 01, 2024)

    2024 arXiv

  38. [46]

    Haghiri, S.; Viquez Rojas, C.; Bhat, S.; Isayev, O.; Slipchenko, L., ANI/EFP: Modeling Long - Range Interactions in ANI Neural Network with Effective Fragment Potentials. J. Chem. Theory Comput. 2024 , 20, 9138 - 9147

    2024

  39. [47]

    W., emle - engine: A Flexible Electrostatic Machine Learning Embedding Package for Multiscale Molecular Dynamics Simulations

    Zinovjev, K.; Hedges, L.; Montagud Andreu, R.; Woods, C.; Tuñón, I.; van der Kamp, M. W., emle - engine: A Flexible Electrostatic Machine Learning Embedding Package for Multiscale Molecular Dynamics Simulations. J. 15 Chem. Theory Comput. 2024 , 20, 4514 - 4522

    2024

  40. [48]

    A.; Ignacio, P.; Huddleston, K

    Semelak, J. A.; Ignacio, P.; Huddleston, K. K.; Olmos, J.; Grassano, J. S.; Clemente, C.; Drusin, S. I.; Marti, M.; Gonzalez Lebrero, M. C.; Roitberg, A. E.; Estrin, D. A., ANI Neural Networks Meet Electrostatics: A ML/MM Implementation in Amber. ChemRxiv 2024

    2024

  41. [49]

    Thürlemann, M.; Riniker, S., Hybrid classical/machine - learning force fields for the accurate description of molecular condensed - phase systems. Chem. Sci. 2023 , 14, 12661 - 12675

    2023

  42. [50]

    P .; Abreu, C

    Eastman, P .; Galvelis, R.; Peláez, R. P .; Abreu, C. R. A.; Farr, S. E.; Gallicchio, E.; Gorenko, A.; Henry, M. M.; Hu, F.; Huang, J.; Krämer, A.; Michel, J.; Mitchell, J. A.; Pande, V . S.; Rodrigues, J. P.; Rodriguez - Guerra, J.; Simmonett, A. C.; Singh , S.; Swails, J.; T...

    2024

  43. [51]

    F.; Wu, Y .; Lemmin, T.; Maxwell, A

    Polizzi, N. F.; Wu, Y .; Lemmin, T.; Maxwell, A. M.; Zhang, S. - Q.; Rawson, J.; Beratan, D. N.; Therien, M. J.; DeGrado, W. F., De novo design of a hyperstable non - natural protein – ligand complex with sub - Å accuracy. Nat. Chem. 2017 , 9, 1157 - 1164

    2017

  44. [52]

    D.; Cieplak, P.; Bayly, C

    Cornell, W. D.; Cieplak, P.; Bayly, C. I.; Gould, I. R.; Merz, K. M.; Ferguson, D. M.; Spellmeyer, D. C.; Fox, T.; Caldwell, J. W.; Kollman, P. A., A second generation force field for the simulation of proteins, nucleic acids, and organic molecules. J. Am. Chem. Soc 1995 , 117...

    1995

  45. [53]

    L.; Grubmuller, H.; MacKerell, A

    Huang, J.; Rauscher, S.; Nawrocki, G.; Ran, T.; Feig, M.; de Groot, B. L.; Grubmuller, H.; MacKerell, A. D., Jr., CHARMM36m: an improved force field for folded and intrinsically disordered proteins. Nat Methods 2017 , 14, 71 - 73

    2017

  46. [54]

    - S.; Bax, A., Determination of φ and χ1 Angles in Proteins from 13C−13C Three - Bond J Couplings Measured by Three - Dimensional Heteronuclear NMR

    Hu, J. - S.; Bax, A., Determination of φ and χ1 Angles in Proteins from 13C−13C Three - Bond J Couplings Measured by Three - Dimensional Heteronuclear NMR. How Planar Is the Peptide Bond? J. Am. Chem. Soc 1997 , 119, 6360 - 6368

    1997

  47. [55]

    Elstner, M.; Porezag, D.; Jungnickel, G.; Elsner, J.; Haugk, M.; Frauenheim, T.; Suhai, S.; Seifert, G., Self - consistent - charge density - functional tight - binding method for simulations of complex materials properties. Phys. Rev. B 1998 , 58, 7260 - 7268

    1998

  48. [56]

    L.; Chandrasekhar, J.; Madura, J

    Jorgensen, W. L.; Chandrasekhar, J.; Madura, J. D.; Impey, R. W.; Klein, M. L., Comparison of simple potential functions for simulating liquid water. J. Chem. Phys. 1983 , 79, 926 - 935

    1983

  49. [57]

    Darden, T.; York, D.; Pedersen, L., Particle mesh Ewald: An N ⋅ log(N) method for Ewald sums in large systems. J. Chem. Phys. 1993 , 98, 10089 - 10092

    1993

  50. [58]

    Liang, J.; Tan, P.; Hong, L.; Jin, S.; Xu, Z.; Li, L., A random batch Ewald method for charged particles in the isothermal – isobaric ensemble. J. Chem. Phys. 2022 , 157, 144102

    2022

  51. [59]

    constants

    Schutz, C. N.; Warshel, A., What are the dielectric “constants” of proteins and how to validate electrostatic models? PROTEINS 2001 , 44, 400 - 417

    2001

  52. [60]

    T., Molecular polarizabilities calculated with a modified dipole interaction

    Thole, B. T., Molecular polarizabilities calculated with a modified dipole interaction. Chem. Phys. 1981 , 59, 341 - 350

    1981

  53. [61]

    - D.; Head - Gordon, M., Long - range corrected hybrid density functionals with damped atom – atom dispersion corrections

    Chai, J. - D.; Head - Gordon, M., Long - range corrected hybrid density functionals with damped atom – atom dispersion corrections. Phys. Chem. Chem. Phys. 2008 , 10, 6615 - 6620

    2008

  54. [62]

    T.; Unke, O

    Frank, J. T.; Unke, O. T.; Müller, K. - R.; Chmiela, S., A Euclidean transformer for fast and stable machine learned force fields. Nat. Commun 2024 , 15, 6539

    2024

  55. [63]

    T.; Dou, S

    Kabylda, A.; Frank, J. T.; Dou, S. S.; Khabibrakhmanov, A.; Sandonas, L. M.; Unke, O. T.; Chmiela, S.; Muller, K. - R.; Tkatchenko, A., Molecular Simulations with a Pretrained Neural Network and Universal Pairwise Force Fields. ChemRxiv 2024

    2024

  56. [64]

    T.; Stöhr, M.; Ganscha, S.; Unterthiner, T.; Maennel, H.; Kashubin, S.; Ahlin, D.; Gastegger, M.; Medrano Sandonas, L.; Berryman, J

    Unke, O. T.; Stöhr, M.; Ganscha, S.; Unterthiner, T.; Maennel, H.; Kashubin, S.; Ahlin, D.; Gastegger, M.; Medrano Sandonas, L.; Berryman, J. T.; Tkatchenko, A.; Müller, K. - R., Biomolecular dynamics with machine - 16 learned quantum - mechanical force fields trained on diver...

  57. [65]

    Wang, H.; Yang, W., Toward Building Protein Force Fields by Residue - Based Systematic Molecular Fragmentation and Neural Network. J. Chem. Theory Comput. 2019 , 15, 1409 - 1417

    2019

  58. [66]

    - Y ., Ab initio characterization of protein molecular dynamics with AI2BMD

    Wang, T.; He, X.; Li, M.; Li, Y .; Bi, R.; Wang, Y .; Cheng, C.; Shen, X.; Meng, J.; Zhang, H.; Liu, H.; Wang, Z.; Li, S.; Shao, B.; Liu, T. - Y ., Ab initio characterization of protein molecular dynamics with AI2BMD. Nature 2024 , 635, 1019 - 1027

    2024

  59. [67]

    W.; Ren, P., Polarizable Atomic Multipole - Based AMOEBA Force Field for Proteins

    Shi, Y .; Xia, Z.; Zhang, J.; Best, R.; Wu, C.; Ponder, J. W.; Ren, P., Polarizable Atomic Multipole - Based AMOEBA Force Field for Proteins. J. Chem. Theory Comput. 2013 , 9, 4046 - 4063

    2013

  60. [68]

    Warshel, A.; Levitt, M., Theoretical studies of enzymic reactions: Dielectric, electrostatic and steric stabilization of the carbonium ion in the reaction of lysozyme. J. Mol. Biol. 1976 , 103, 227 - 249

    1976

  61. [69]

    J.; Bash, P

    Field, M. J.; Bash, P. A.; Karplus, M., A combined quantum mechanical and molecular mechanical potential for molecular dynamics simulations. J. Comput. Chem. 1990 , 11, 700 - 733

    1990

  62. [70]

    Schaefer, P.; Riccardi, D.; Cui, Q., Reliable treatment of electrostatics in combined QM/MM simulation of macromolecules. J. Chem. Phys. 2005 , 123, 014905. Supporting information 16 A1. Ewald - summation of external electrostatic potential in periodic ML/MM systems. For perio...

    2005

  63. [71]

    This computation of external electrostatic potential in periodic ML/MM system s is implemented in the OpenMM - Torch API using PyTorch as the programming framework

    The self - energy term: = − 2 √ 2 . This computation of external electrostatic potential in periodic ML/MM system s is implemented in the OpenMM - Torch API using PyTorch as the programming framework. The intensive computation of the reciprocal term is parallelized through com...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.