{"id":"26447fed-ab23-466e-84db-3dac289e5039","arxiv_id":"2411.19728","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"An anisotropic message-passing neural network, embedded electrostatically in a classical MM environment, predicts solution-phase reaction free energies in explicit solvent with near-chemical accuracy in three benchmark applications.","lead":"The authors replace the expensive quantum-mechanical part of QM/MM simulations with a neural network potential (AMP), letting molecular dynamics simulations in explicit solvent run much longer and cheaper. They compute solution free energies for three test systems, including charged pyridine and quinoline dimers, and report close agreement with experiment.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on an unvalidated DFT-plus-point-charges proxy for experimental free energies: the NNP can only be as accurate as its reference labels in explicit solvent, and no explicit-solvent CCSD(T)-level check is provided.","rationale":"The reader's weakest_assumption correctly identifies the central vulnerability: the NNP is trained to reproduce DFT energies with fixed MM point charges, and no explicit-solvent CCSD(T)-level benchmark is provided. I agree with that assessment. The paper has real independent support: deposited datasets, a public implementation, stable prospective MD simulations, consistent generalization experiments, and chemically sensible free-energy surfaces for alanine dipeptide. The central concern is not internal inconsistency or fraud; it is that the claim 'same accuracy' is anchored to a reference electronic-structure method whose explicit-solvent accuracy is unmeasured. The proposed test directly attacks this anchor by replacing the DFT labels with a higher-level wavefunction reference for a representative system and measuring the shift in the derived free energy. The nickel results reinforce the need for cautious wording, but they are a secondary issue; the core conditionality of the verdict is the unvalidated DFT proxy. Therefore the reader's CONDITIONAL verdict stands, with no adjustment needed.","tokens_in":32460,"tokens_out":3750,"duration_ms":40728,"concrete_test":"For one small charged dimer (system 5a) and the methyl outlier (5b), take the existing umbrella-sampling configurations from Section 3.2 (or regenerate 1000 frames per window) and recompute the QM energy with the same MM point charges at DLPNO-CCSD(T)/CBS instead of omegaB97M-D4/def2-TZVPP. Train the 600k-parameter AMP model on these corrected labels for the two systems and rerun the 42-window umbrella-sampling calculation. If the resulting dissociation free energies shift by more than about 4.184 kJ mol-1 relative to the current values, the DFT reference is the dominant source of error and the 'excellent agreement' claim must be softened. If they shift by less than 1 kJ mol-1, this concern is resolved and the reported agreement is robust to the choice of reference electronic-structure method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims that substituting a QM Hamiltonian by the AMP NNP gives 'the same accuracy' and yields 'excellent agreement with experimental data.' For this central claim to hold, the training labels must faithfully represent the true solvated potential energy surface. In Section 5.3, all labels are DFT energies evaluated in the field of fixed MM point charges: B2-PLYP/def2-QZVPP for alanine dipeptide and omegaB97M-D4/def2-TZVPP for the nickel and pyridine/quinoline systems. The NNP is trained to reproduce exactly these labels, so any systematic error in the DFT functional, in the fixed-charge electrostatic embedding, or in the QM/MM cutoff treatment is inherited by the NNP and cannot be corrected by enhanced sampling. Footnote 4 explicitly concedes that an explicit-solvent double-hybrid DFT ground truth was not feasible, and no independent explicit-solvent CCSD(T)-level reference is reported anywhere in the manuscript. Consequently, the agreement for pyridine/quinoline dimers in Sections 3.1-3.2 validates the pipeline 'DFT labels + explicit solvent + sampling', not specifically the NNP replacement. The comparison against static QM with implicit solvent (up to ten times larger deviations) is not a substitute: it differs in both the solvent model and the sampling, so it cannot isolate the quality of the DFT reference. The nickel application further strains the 'excellent agreement' wording: two of ten ligation states are misclassified and the Pt-Bu3 value is hand-assigned. The load-bearing vulnerability, however, is the unvalidated DFT ground truth: if omegaB97M-D4/def2-TZVPP with fixed MM charges is biased by more than chemical accuracy for these charged dimers, the central claim needs to be restated as 'agreement conditioned on the DFT reference' rather than 'same accuracy as experiment.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an extended version of the anisotropic message passing (AMP) neural network architecture for ML/MM simulations with electrostatic embedding, and applies it to compute free energies for three systems: alanine dipeptide conformational sampling, ligation states of nickel phosphine complexes, and dissociation free energies of charged pyridine/quinoline dimers. The model is trained on DFT energies and gradients evaluated in the field of fixed MM point charges, then used as the QM Hamiltonian in umbrella-sampling MD with explicit solvent. The authors report training errors below chemical accuracy, stable trajectories in prospective simulations, and computed free energies that are compared with experimental data and with static DFT/implicit-solvent and GFN2-xTB QM/MM calculations.","tokens_in":32782,"tokens_out":6558,"duration_ms":58308,"significance":"If the claims are validated, the approach is a significant methodological advance: it demonstrates electrostatic-embedding ML/MM simulations of explicitly solvated systems with hundreds of QM atoms over hundreds of nanoseconds, including quantitative free-energy predictions for challenging charged dimers and transition-metal complexes. The paper's strengths include independent experimental benchmarks (NMR ligation states in Ref. [102], experimental dissociation free energies in Ref. [119]) rather than validation only against training labels, explicit discussion of limitations (footnote 4 on the absence of a double-hybrid explicit-solvent ground truth; the LJ-parameter hypothesis for 5b), and releases of training data and code. The experimental agreement provides an independent grounding of the central claim, although the nickel application shows partial misclassifications that temper the 'excellent agreement' wording.","major_comments":[{"comment":"The abstract's claim of 'excellent agreement with experimental data' is overstated for the nickel phosphine application: two of ten complexes are misclassified by the AMP models (CataCXium A in both entries, CyJohnPhos in entry 7), and the value for Pt-Bu3 is set arbitrarily to 41.84 kJ/mol because the bisligated complex dissociated in simulation. This post hoc assignment directly influences the classification success rate and should not be counted as a quantitative prediction. Please report classification accuracy and confidence intervals on ΔGdiss, treat Pt-Bu3 as a qualitative dissociation event, and either exclude it from numerical error statistics or justify why the arbitrary value is not biasing the conclusion.","section":"Section 3.2 (Nickel complex application), Figure 7"},{"comment":"All training labels are DFT energies evaluated in a fixed-charge QM/MM embedding (B2-PLYP/def2-QZVPP for alanine dipeptide; ωB97M-D4/def2-TZVPP for the other systems), with no explicit-solvent CCSD(T)- or double-hybrid-level reference. The experimental benchmarks therefore validate the combined pipeline (DFT reference + explicit solvent + NNP sampling), not the NNP substitution in isolation. This limitation is acknowledged in footnote 4, but the abstract's statement that the NNP has 'the same accuracy' as the QM Hamiltonian should be qualified. A small explicit-solvent high-level benchmark (e.g., a few pyridine dimers at DLPNO-CCSD(T) level) would directly quantify the systematic error of the reference labels and materially strengthen the central claim.","section":"Section 5.3 and footnote 4"},{"comment":"The expression for the MM-induced multipoles contains a sum over all QM atoms i on the right-hand side for a quantity indexed by i on the left-hand side. As written, the equation suggests a nonlocal polarization term that redistributes multipoles across the entire QM zone, which is inconsistent with the per-atom polarizability description in the text. Please clarify the summation limits (presumably only over MM atoms j for each fixed i) or explain the intended global contribution. This is a technical description that directly affects the reproducibility of the method.","section":"Section 2.3.2, Eq. (16)"}],"minor_comments":[{"comment":"The comparison between AMP free-energy minima and static B2-PLYP minima mixes two different definitions of 'minimum': free-energy surface local minima from umbrella sampling versus optimized geometries with quasi-RRHO corrections. A direct comparison using the same state definition (e.g., same integration regions on the free-energy surface) would make the quantitative agreement for alanine dipeptide more transparent.","section":"Section 3.1, Table 2"},{"comment":"The phrase 'up to ten times higher deviation' for static QM methods is an overstatement; the reported medAE values are about 5-8 times higher than the AMP medAE. Please rephrase to 'roughly five to eight times higher' or provide the maximum ratio explicitly.","section":"Section 3.2, Figure 10"},{"comment":"The arbitrary assignment of the Pt-Bu3 dissociation free energy is mentioned in a note in the text but should be highlighted in the main body of the results because it is a central caveat for interpreting the nickel classification numbers.","section":"Section 3.2, discussion after Eq. (19)"},{"comment":"The scaling claim O(N^1.24) is based on CPU inference steps per second, while the GPU scaling is described only qualitatively. Please report error bars or confidence intervals for the fitted scaling exponents and specify the system sizes used for each point on the GPU curve.","section":"Section 2.3.3, Figure 2"},{"comment":"There are several formatting artifacts (e.g., 'bracehtipupleft' in Eq. (6), the limits in Eq. (14), and the placement of summation indices in Eq. (15)). Please ensure the equations are typeset cleanly so that the summation ranges and tensor indices are unambiguous.","section":"Throughout (Eqs. (6), (14), (15))"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and makes a genuinely interesting methodological contribution. The main technical concern is that the 'excellent agreement' claim is not uniformly supported by the nickel results, and the arbitrary Pt-Bu3 assignment should be handled more carefully. The absence of an explicit-solvent high-level reference is a limitation but not a fatal flaw, since the experimental benchmarks provide independent validation of the overall pipeline. I would encourage the editor to invite a revision addressing the major comments above."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a solid, useful ML/MM paper that delivers on its main promise — explicit-solvent free energies at near-chemical accuracy for challenging charged systems — and it ships data and code. The central caveat is that the NNP is trained on DFT-in-a-field-of-point-charges labels, and there is no explicit-solvent CCSD(T) check, so the 'same accuracy as QM' claim really means 'same accuracy as the DFT reference plus sampling.' The experimental agreement for the pyridine/quinoline dimers is the independent evidence that the pipeline holds, but that should be framed more carefully than 'excellent agreement' in the abstract.\n\nWhat's new: the next-gen AMP architecture with multiple multipole channels, the added intra-QM Coulomb term, and the first quantitative ML/MM free-energy applications for charged dimers and nickel phosphine ligation states. The generalization tests — training on subregions of the Ramachandran plot or on small pyridines only — are well designed and show real transferability. The scaling data on CPU and GPU is believable, and the stability of prospective MD is a genuine plus. Self-citation to the prior AMP paper is appropriate; the new content is clearly separated from that work.\n\nSoft spots, in proportion. The DFT-proxy issue is the biggest. All labels are B2-PLYP or omegaB97M-D4 with fixed MM charges; footnote 4 admits an explicit-solvent double-hybrid ground truth wasn't feasible. If that DFT model is biased by more than chemical accuracy for these charged dimers, the NNP inherits the bias. But note this is not circular: the benchmarks are against experimental data, not training labels. The paper validates the full pipeline — DFT labels plus explicit solvent plus sampling — and that is the honest claim.\n\nThe nickel application is the weakest part: two of ten complexes misclassified (CataCXium A in both models, CyJohnPhos in one), and the PtBu3 value is hand-assigned because the bisligated state dissociated. That is a real blemish but the overall ranking is correct and the method still beats static DFT with implicit solvent on classification. The abstract's 'excellent agreement' is too strong for this section.\n\nMinor: main-text figures lack per-point error bars, and the sensitivity to LJ parameters for nickel and to standard-state corrections is discussed only as hand-waving. The authors list plausible causes for the misclassifications but do not test them.\n\nWho this is for: computational chemists working on ML potentials or QM/MM free energy. It deserves serious referee time. I'd send it to review with a request to soften the abstract, report error bars, and add a sensitivity analysis for the nickel LJ parameters and the DFT reference (e.g., one CCSD(T) or double-hybrid benchmark for a couple of dimers in explicit solvent, even at a reduced basis).","headline":"A practical ML/MM workflow with real experimental grounding; the DFT-proxy caveat and two Ni misclassifications keep it from being a clean sweep, but it deserves serious referee time.","tokens_in":33413,"tokens_out":3440,"would_cite":true,"duration_ms":28020,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A trained neural network potential can replace the expensive quantum-mechanical Hamiltonian inside an explicit-solvent simulation and reproduce experimental reaction free energies within a few kJ/mol across very different chemical systems.","keywords":["neural network potential","QM/MM","anisotropic message passing","electrostatic embedding","free energy","umbrella sampling","explicit solvent","reaction free energy"],"falsifier":"Compute the dissociation free energy of one challenging dimer, such as 7f or 6l, using explicit-solvent QM/MM umbrella sampling directly at the omegaB97M-D4/def2-TZVPP level of theory without any neural network, and compare the result with the AMP/MM prediction; if the directly sampled value differs from the AMP/MM value by more than the reported few kJ/mol, the network is not faithfully replacing the QM Hamiltonian for that system.","tokens_in":1809,"feed_emoji":"🧪","tokens_out":2152,"duration_ms":59663,"temperature":0.7,"pith_summary":"The paper tries to establish that a neural network potential can stand in for the expensive quantum-mechanical Hamiltonian inside a QM/MM simulation without losing accuracy, so that reaction free energies in explicit solvent can be computed by sampling hundreds of nanoseconds rather than by static high-level calculations. It reports that the anisotropic message passing architecture reproduces DFT energies and forces below chemical accuracy for three demanding test beds: alanine dipeptide, nickel phosphine complexes, and charged pyridine and quinoline dimers. It then shows that free energies from ML/MM umbrella sampling agree with experimental values within a few kJ/mol, while static DFT with implicit solvent and QM/MM simulations using a semi-empirical Hamiltonian deviate by up to ten times more and can even rank the wrong species. If the central claim is correct, the result would remove the main sampling bottleneck of QM/MM for condensed-phase reactivity studies.","feed_headline":"Neural net reproduces solution reaction free energies","feed_subtitle":"ML/MM umbrella sampling with anisotropic message passing matches experiment where static DFT and semi-empirical QM fall short.","key_machinery":"The central object is the AMP (anisotropic message passing) neural network potential, an equivariant graph neural network that places atomic multipoles (monopole, dipole, quadrupole) on each atom and uses multipole-interaction coefficients to make the messages directionally sensitive. The model is embedded in a QM/MM-style total energy $V_{\\text{total}} = V_{\\text{QM}} + V_{\\text{QM-MM}} + V_{\\text{MM}}$, with a Coulomb monopole-monopole term inside the QM zone for interactions beyond the graph cutoff, Lennard-Jones plus multipole-electrostatic coupling to MM point charges, and an MM-charge-induced polarization term added to the QM multipoles. This machinery does two jobs at once: it keeps the description of long-range and directional electrostatics physical enough to stabilize charged and metal-containing solutes in explicit solvent, and it is cheap enough to run umbrella sampling with tens of thousands of solvent atoms for hundreds of nanoseconds.","core_discovery":"The central claim is that substituting a QM Hamiltonian with the AMP neural network potential in an electrostatic-embedding ML/MM scheme preserves the accuracy of the underlying DFT reference while making long enhanced-sampling molecular dynamics tractable, and that the resulting free energies match experiment. Evidence includes a two-dimensional alanine dipeptide free-energy surface whose minima match DFT and NMR-derived expectations even when the network was trained on only a narrow slice of the torsional space; dissociation free energies for nickel phosphine complexes that rank experimentally mono- and bisligated states correctly, including large complexes absent from the training set; and dissociation free energies for a series of charged pyridine and quinoline dimers with a median absolute error of 2.63 kJ/mol, with static DFT plus implicit solvent off by seven to eleven times more. The paper further shows the approach scales to systems with more than 350 solute atoms embedded in tens of thousands of solvent atoms, and that inference is fast enough to run on a single CPU for small systems.","pith_inferences":["If the central claim is right, a substantial share of the error in static QM free energies comes from missing sampling and explicit solvent rather than from the DFT functional itself, which suggests that other ML/MM potentials could inherit this accuracy improvement without architectural changes.","The paper hints that a globally trained foundational potential fine-tuned on a few dozen examples per target could replace per-system training; that is a natural next step the authors describe but do not demonstrate.","Because the solvent is represented by fixed MM point charges, the method's accuracy is ultimately capped by that point-charge description of the environment; testing a polarizable MM model or a larger QM zone would show whether remaining outliers are due to the embedding or to Lennard-Jones parameters.","The authors note that a larger network did not meaningfully improve free-energy predictions, which implies that for these observables the DFT label error, not network capacity, may dominate; retraining the same architecture on a different high-level functional would test that interpretation directly."],"forward_implications":["For reaction free energies in solution, long enhanced-sampling MD becomes feasible at near-DFT accuracy because the neural network Hamiltonian is cheap enough to evaluate for hundreds of nanoseconds with explicit solvent.","The same trained model can be applied to chemically related molecules that were absent from the training set, so one training dataset can cover a family of ligands or substituted dimers without per-molecule retraining.","Static DFT with implicit solvent and semi-empirical QM/MM can be wrong by up to an order of magnitude in free energy and sometimes rank the wrong state; replacing the QM Hamiltonian with the trained network removes most of that error.","Free-energy differences between bound and unbound states can be computed with median absolute errors around 2-3 kJ/mol, at or below the usual chemical-accuracy threshold of 4.184 kJ/mol.","The approach extends to transition-metal complexes, charged species, and large ligands where analytical Hessian calculations become impractical, so mechanistic questions that static methods cannot address become accessible."],"supporting_citations":[{"why":"Introduces the AMP anisotropic message passing architecture that the paper extends and applies as the neural network Hamiltonian.","marker":"[14]"},{"why":"Defines the semi-empirical GFN2-xTB Hamiltonian used to generate training configurations and serves as the main comparison baseline in prospective simulations.","marker":"[15, 16]"},{"why":"Supplies the umbrella sampling methodology used for all free-energy calculations.","marker":"[80, 81]"},{"why":"Defines the B2-PLYP/def2-QZVPP(D3BJ) reference level used to label alanine dipeptide training data.","marker":"[82, 83, 84, 85]"},{"why":"Defines the omegaB97M-D4/def2-TZVPP reference level used to label nickel and pyridine dimer training data.","marker":"[110, 111, 112, 113]"},{"why":"Provides the experimental nickel phosphine ligation-state assignments used as ground truth for the catalysis benchmark.","marker":"[102]"},{"why":"Provides the experimental dissociation free energies of charged pyridine and quinoline dimers used as ground truth for the quantitative benchmark.","marker":"[119]"}],"fun_headline_variants":["Neural net potential accelerates accurate solution free energies","Anisotropic message passing ML/MM matches experiment in free energies","Graph neural net replaces QM for fast, accurate solvation free energies","Machine learned potential scales to >350 solute, 40k solvent atoms","MM with neural net beats static DFT in reaction free energies"],"cache_read_input_tokens":35328,"weakest_assumption_plain":"The central premise is that the DFT reference used for training (B2-PLYP for alanine dipeptide and omegaB97M-D4 for the nickel and pyridine systems), evaluated with fixed MM point charges, is accurate enough to stand in for the experimentally measured free energies once explicit solvent and sampling are included; if that DFT-plus-charges model carries systematic error, the trained network inherits it and so do all three benchmark predictions.","fun_headline_variants_meta":{"raw":{"variants":["Neural net potential accelerates accurate solution free energies","Anisotropic message passing ML/MM matches experiment in free energies","Graph neural net replaces QM for fast, accurate solvation free energies","Machine learned potential scales to >350 solute, 40k solvent atoms","MM with neural net beats static DFT in reaction free energies"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000225,"raw_usage":{"total_tokens":1491,"prompt_tokens":999,"completion_tokens":492,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":405}},"tokens_in":615,"tokens_out":492,"duration_ms":4632,"temperature":1.0,"reasoning_tokens":405,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:55:04.972222+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the dissociation free energy of one challenging dimer, such as 7f or 6l, using explicit-solvent QM/MM umbrella sampling directly at the omegaB97M-D4/def2-TZVPP level of theory without any neural network, and compare the result with the AMP/MM prediction; if the directly sampled value differs from the AMP/MM value by more than the reported few kJ/mol, the network is not faithfully replacing the QM Hamiltonian for that system.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the experimental nickel phosphine ligation-state assignments used as ground truth for the catalysis benchmark."},{"cited_title":"Pollice, M","cited_arxiv_id":null,"evidence_quote":"Provides the experimental dissociation free energies of charged pyridine and quinoline dimers used as ground truth for the quantitative benchmark."}],"review_version":1}