{"id":"d1c0e4c0-8fbf-49d8-beb5-edb45b971bd5","arxiv_id":"2608.13355","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An electrostatic-embedding MLIP/MM scheme with machine-learned RESP charges improves relative binding free energy predictions on TYK2 but not on four other benchmark targets.","lead":"This paper tests whether letting a machine-learned potential update a ligand's electric charges during a protein-ligand binding simulation improves predicted binding free energies. On one of five benchmark targets it roughly halves the error; on the other four it matches existing methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"TYK2 gain is not yet attributable to electrostatic embedding: the electrostatically embedded AceFF-2-RESP-1 differs from the mechanical-embedding baseline AceFF-1.0 in both base potential and charge model; no same-potential mechanical-embedding control is run.","rationale":"Reader's weakest_assumption is the short-range-only PME correction. That is a genuine sensitivity, but it is secondary to the attribution problem: a systematic approximation in E_coul would affect all targets and does not need to be true for the paper's central TYK2 claim to be underdetermined. The distinguishing claim, that electrostatic embedding caused TYK2's improvement, requires comparing methods that differ only in embedding. The paper instead compares a new network plus new embedding against an old network with mechanical embedding. This is an omitted control, not a parameter-sensitivity issue, and it can be settled by one additional TYK2 simulation campaign. If the same AceFF-2-RESP-1 network with fixed charges reproduces the low RMSE, then the paper's mechanistic interpretation and the 'first electrostatic-embedding MLIP/MM' framing lose support; if it does not, the embedding claim is strengthened. The manuscript's honest Limitations section is a strength, but it misses this specific confound. The verdict should remain conditional: accept the benchmark as a useful evaluation, but require the mechanical-embedding control before treating the TYK2 result as evidence for electrostatic embedding. I partially agree with the reader because their rationale mentions this confound, but their stated weakest assumption is the PME approximation.","tokens_in":16549,"tokens_out":5163,"duration_ms":50687,"concrete_test":"Run TYK2 with AceFF-2-RESP-1 under mechanical embedding: predict charges with the same network but freeze them at a reference conformation (or use time-averaged charges), omit E_pol and E_distortion, and keep the standard PME treatment with fixed charges. Use the same lambda windows, soft-core parameters, simulation length, and random seeds as in the reported electrostatic-embedding runs. If the RMSE remains near 0.45 kcal/mol, the TYK2 gain does not depend on electrostatic coupling; if it returns to roughly 0.7-0.8 kcal/mol, the embedding is responsible. This single control isolates the variable the paper claims to evaluate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 2's TYK2 row is the empirical basis for the central claim: electrostatic embedding roughly halves DeltaDeltaG RMSE (0.86 -> 0.45 vs GAFF2, 0.77 -> 0.45 vs AceFF-1.0). But the comparison conflates electrostatic embedding with two other changes. AceFF-2-RESP-1 is a TensorNet2 model with 2 interaction layers and a 128-dimensional embedding, jointly trained on energies, forces, and RESP charges; the mechanical-embedding baseline AceFF-1.0 is a different, older network using fixed AM1-BCC charges. Thus E_MLIP in Eq. 2 is different, not only E_interaction in Eqs. 3-4. The Introduction states that holding systems, edges, protocol, and sampling fixed 'lets us separate the effect of the electrostatic coupling from the effect of the underlying potential,' but the underlying potential is not held fixed; the Methods only reuse the QuantumBind-RBFE protocol unchanged. The Limitations section lists untested embedding parameters and the short-range PME hypothesis, but not this missing control. Without a mechanical-embedding run of the same AceFF-2-RESP-1 network (fixed charges, no E_pol or E_distortion), the TYK2 improvement could come from AceFF-2's better potential energy surface or from switching to RESP charges, rather than from dynamic-charge electrostatic coupling. The target-dependence story ('TYK2 is rigid, static charges were limiting') is therefore not uniquely supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a TensorNet2-based machine learning potential, AceFF-2-RESP-1, that jointly predicts energies, forces, and conformation-dependent RESP partial charges, and embeds it electrostatically in an MLIP/MM scheme following Semelak et al. The method is tested in production alchemical RBFE calculations on five targets from the Wang et al. benchmark set, with three replicates per edge and the QuantumBind-RBFE protocol. The central results are that electrostatic embedding improves essentially every reported accuracy and correlation metric for TYK2 (ΔΔG RMSE 0.86→0.45 kcal/mol against GAFF2 and 0.77→0.45 against the mechanical-embedding AceFF-1.0 baseline) while giving neutral results for CDK2, thrombin, p38, and JNK1. The paper also reports single-molecule benchmarks (Wiggle150, Sellers torsion scans, Schrödinger ligand benchmark) and finds that static benchmarks do not predict the target-dependent RBFE outcome.","tokens_in":16897,"tokens_out":3006,"duration_ms":28998,"significance":"If the attribution to electrostatic embedding is accepted, this is the first production-scale alchemical RBFE campaign with electrostatic-embedding MLIP/MM, and the TYK2 result is a notable, target-dependent improvement with a plausible mechanistic story. The paper has concrete strengths: the target set and edges were fixed in advance by prior work, three independent replicates per edge were run, bootstrap confidence intervals are reported, the model weights and modified software are publicly released, and the Limitations section is unusually candid about untested embedding parameters and the short-range PME hypothesis. The main weakness is that the experimental design does not isolate the electrostatic embedding from changes in the base potential and the charge model, so the paper's headline causal claim is not uniquely supported by the data. The TYK2 gain remains a real empirical result for the combined AceFF-2-RESP-1 electrostatically embedded model, but the specific mechanism is not established by the current comparison.","major_comments":[{"comment":"The claim that fixing systems, edges, protocol, and sampling 'lets us separate the effect of the electrostatic coupling from the effect of the underlying potential' is not supported by the design, because AceFF-2-RESP-1 differs from the mechanical-embedding baseline AceFF-1.0 in three ways at once: the embedding (electrostatic vs mechanical), the base network architecture and training (TensorNet2 with 2 interaction layers and 128-dimensional embedding vs the older AceFF-1.0 model), and the charge model (RESP vs AM1-BCC). A mechanical-embedding run using the same AceFF-2-RESP-1 network with fixed charges is a necessary control to attribute the TYK2 improvement specifically to electrostatic coupling. Without it, the improvement could plausibly stem from the better base potential energy surface of AceFF-2 or from the switch to RESP charges. The paper should either add such a control or substantially soften the causal language throughout the abstract, results, and conclusion.","section":"Introduction; Table 2"},{"comment":"The approximation that only the short-range direct-space part of the PME sum uses the dynamic MLIP charges is justified by the hypothesis that 'long-range electrostatics vary negligibly between RESP and AM1-BCC parameterizations.' This hypothesis is load-bearing for the reported free energies, since the electrostatic correction in Eq. (6) is only a short-range difference term. The manuscript does not test the sensitivity of any RBFE result to this approximation, despite the Limitations section correctly flagging it. A concrete test would be to compute full-PME dynamic-charge energies (or at least the reciprocal-space contribution) for representative end-state and intermediate lambda configurations on one target, and report the magnitude of the neglected long-range term relative to the observed TYK2 free-energy changes.","section":"Methods: Short-Range PME Modification and Thole Damping; Limitations"},{"comment":"The Limitations paragraph lists untested embedding parameters (polarizabilities, epsilon, Thole exponent) and the short-range PME hypothesis, but does not list the missing same-potential mechanical-embedding control. As a result, the interpretation that 'static ligand charges were a primary limiting factor' for TYK2 and the conclusion that 'electrostatic embedding roughly halved the error' overstate the evidence. The authors should either run the missing control or reframe the main claim as an evaluation of the complete AceFF-2-RESP-1 electrostatically embedded scheme against two baselines, without asserting that the improvement is caused by electrostatic embedding per se.","section":"Limitations; Results: RBFE benchmarks"}],"minor_comments":[{"comment":"The dataset size '10 6' should be typeset as 10^6; the missing superscript appears in both the abstract and the Methods section.","section":"Abstract; Section: Quantum Chemical Dataset and Neural Network Training"},{"comment":"The model name appears inconsistently as AceFF-2-RESP-1 and AceFF-2-resp-1 (Table 1); please unify to the capitalization used in the title and abstract.","section":"Throughout"},{"comment":"The text contains typographical spacing errors in table captions ('T able 1', 'T able 2', 'T able 3'); these should be corrected to 'Table'.","section":"Tables 1-3"},{"comment":"The phrase 'hybrid MLIP/MM (or MLIP/MM) schemes' is redundant; the intended distinction between 'MLIP/MM' and 'NNP/MM' should be clarified or the parenthetical removed.","section":"Introduction"},{"comment":"Panel (b) shows a parity plot for predicted vs. reference RESP charges that the text calls 'tight, linear,' but no correlation coefficient or R^2 is reported; adding the numerical value would make the claim verifiable.","section":"Results: Figure 2"},{"comment":"The notation 'ˆ⃗ rij' is not defined before use; please define the unit vector symbol and the meaning of the over-hat arrow notation explicitly.","section":"Methods: Short-Range PME Modification and Thole Damping, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper reports a potentially important first production-scale electrostatic-embedding MLIP/MM RBFE campaign, with released code and weights and a commendably honest Limitations section. However, the central attribution claim is underdetermined by the experimental design: the comparison against AceFF-1.0 changes both the embedding and the underlying model, and the short-range-only PME approximation is untested. I would push the authors to add the same-potential mechanical-embedding control and a sensitivity check of the PME approximation; both are feasible within the scope of the manuscript. If the authors are unwilling or unable to run these controls, the paper should be reframed as an evaluation of a combined model rather than a causal study of electrostatic embedding."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know this paper is the first production alchemical RBFE assessment of electrostatic-embedding MLIP/MM against experimental affinities, on five Wang-set targets with three replicates per edge. That part is real and useful. The authors take Semelak's scheme, train a TensorNet2-based model on RESP charges (AceFF-2-RESP-1), and run it through a matched protocol. TYK2 improves across every reported metric (ΔΔG RMSE 0.86→0.45 vs GAFF2), while the other four targets are null. The null result is just as informative as the positive one, and the paper says so plainly.\n\nWhat it does well: the evaluation is careful, with bootstrap intervals, matched systems/edges/protocol, and code and weights released. The limitations section is unusually candid—it flags the untested embedding parameters, the RESP-vs-MBIS choice, and the short-range PME assumption. The static single-molecule benchmark analysis (Figure 5) is a nice negative result: standard energy/force/charge accuracy does not predict which targets benefit. Citation pattern looks fine; the AceFF-1.0 baseline is their own prior work but is the appropriate comparator.\n\nThe soft spot is the attribution of the TYK2 gain. AceFF-2-RESP-1 differs from the AceFF-1.0 baseline in both the embedding (electrostatic vs mechanical) and the underlying potential (new architecture, 2 layers, 128-dim embedding, RESP vs AM1-BCC charges). So the improvement could come from a better potential energy surface or from switching charge models, not from the dynamic electrostatic coupling itself. The Introduction claims that holding systems/edges/protocol fixed \"lets us separate the effect of the electrostatic coupling from the effect of the underlying potential,\" but the potential is not held fixed. A mechanical-embedding run of the same AceFF-2-RESP-1 network would settle this. The short-range PME approximation (only direct-space charges updated) is another untested hypothesis; if it is wrong, the TYK2 correction could be biased. These are acknowledged in Limitations, but the missing control is not. The authors' rigidity story is speculative with n=5, and they mostly say so.\n\nWho is this for? Anyone working on MLIP/MM free energy methods or choosing between embedding schemes for drug discovery. It deserves a serious referee. My recommendation: send it to peer review, but ask for a same-potential mechanical-embedding control, or at least a modest rewrite that stops claiming causal separation and instead reports the benchmark as an apples-to-oranges comparison with a large effect on one target. The paper is worth the extra round.","headline":"First production alchemical test of electrostatic-embedding MLIP/MM, with an honest target-dependent result and a confounded headline gain.","tokens_in":17396,"tokens_out":1481,"would_cite":true,"duration_ms":15609,"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":"This paper reports that electrostatic embedding of a machine-learned ligand potential halves the TYK2 relative binding free-energy error in a production alchemical benchmark, while leaving four other targets unchanged.","keywords":["relative binding free energy","electrostatic embedding","machine learning interatomic potential","MLIP/MM","RESP charges","alchemical free energy","protein-ligand binding","particle mesh Ewald"],"falsifier":"Recompute the TYK2 alchemical edges with the dynamic charges coupled into the full particle mesh Ewald sum, including reciprocal-space gradients, or alternatively with static RESP charges replacing AM1-BCC while keeping the original MLIP/MM scheme; if the $\\Delta\\Delta G$ RMSE reverts toward 0.86 kcal/mol, the reported improvement is an artifact of the short-range-only approximation or of the charge definition change, not of geometry-dependent electrostatics.","tokens_in":16370,"feed_emoji":"🧪","tokens_out":8401,"duration_ms":70420,"temperature":0.7,"pith_summary":"This paper tries to establish that electrostatic embedding of a machine-learned ligand potential inside a classical environment works in production alchemical relative binding free energy (RBFE) calculations, not just in single-point QM/MM-style tests. The authors train one neural network, AceFF-2-RESP-1, to simultaneously predict energies, forces, and RESP partial charges, and feed those charges into the short-range direct-space part of the particle mesh Ewald sum with Thole damping. Across five benchmark targets, this roughly halves the TYK2 $\\Delta\\Delta G$ root-mean-square error (from 0.86 to 0.45 kcal/mol against GAFF2 and from 0.77 to 0.45 kcal/mol against mechanical-embedding AceFF-1.0) and improves every accuracy and correlation metric on that target, while performing comparably to both baselines on CDK2, thrombin, p38, and JNK1. If true, this is the first demonstration that replacing static ligand charges with geometry-dependent ones can pay off in a production alchemical campaign, and that the benefit is target-dependent rather than universal.","feed_headline":"Electrostatic embedding halves TYK2 binding free-energy error","feed_subtitle":"First production alchemical test of dynamic ligand charges: TYK2 improves, four targets unchanged.","key_machinery":"The mechanism is a multi-task neural network, named AceFF-2-RESP-1, built on the TensorNet2 architecture and trained on $10^6$ conformations from the AceFF dataset to jointly predict potential energy, atomic forces, and RESP partial charges. The predicted charges replace static AM1-BCC charges in the ligand–environment Coulomb interaction, but only in the short-range, direct-space term of the particle mesh Ewald sum, implemented as an additive correction against the fixed baseline charges; a Thole-damping factor on the MM electric field prevents polarization catastrophes during alchemical transformations. This design keeps the calculation affordable while letting the ligand charges respond to geometry throughout the $\\lambda$ schedule.","core_discovery":"On the authors' own terms, the discovery is a demonstrated, reproducible target-dependent effect: electrostatic embedding improves RBFE accuracy on TYK2, where static charges appear to be the dominant error source, and is neutral on four more flexible or complex targets. The improvement is consistent across all eight metrics and both baselines, with the TYK2 $\\Delta\\Delta G$ RMSE falling by a factor of 1.7–2.2. The paper also shows that standard single-molecule energy, force, and charge benchmarks do not predict which targets benefit: TYK2 is the only clear winner and also has the lowest force error on the 650-molecule ligand benchmark, but that pattern does not hold for the other four targets. This is the first assessment of an electrostatic-embedding MLIP/MM scheme against experimental binding affinities on a standard congeneric-series benchmark.","pith_inferences":["Not claimed by the paper: the TYK2 gain could come from using RESP rather than AM1-BCC charges even without the charges being dynamic; a control experiment with static RESP charges in the classical baseline would separate these two variables.","If target rigidity is what gates the benefit, the natural testable extension is to run the same protocol on other rigid congeneric series; the authors' own explanation predicts gains where the pocket is rigid and static charges dominate.","The short-range-only PME approximation is a sensitivity that could be checked by reweighting the existing trajectories with full Ewald electrostatics, at low cost, before any new simulation is run.","The neutral results on four targets suggest pragmatic deployment: use electrostatic embedding after a rapid diagnostic of whether the ligand series has charge-sensitive interactions, rather than as a default replacement for mechanical embedding."],"forward_implications":["If the scheme is correct, replacing static AM1-BCC ligand charges with predicted conformation-dependent RESP charges reduces TYK2 $\\Delta\\Delta G$ RMSE from 0.86 to 0.45 kcal/mol against GAFF2 and from 0.77 to 0.45 kcal/mol against AceFF-1.0, improving all eight reported accuracy and correlation metrics.","On CDK2, thrombin, p38, and JNK1, electrostatic embedding matches both baselines within bootstrap intervals, so the method should not be expected to improve every target.","Joint training of energy, forces, and charges costs accuracy on standard benchmarks: AceFF-2-RESP-1 trails the single-task AceFF-2 on the Wiggle150 and torsion benchmarks while staying close on forces.","Standard single-molecule force, energy, and charge errors do not predict which targets will benefit, so model selection for RBFE cannot rely on those benchmarks alone.","The alchemical workflow is numerically stable across the full $\\lambda$ schedule when Thole damping is applied, meaning the scheme is deployable in production campaigns."],"supporting_citations":[{"why":"Supplies the electrostatic-embedding MLIP/MM formulation (energy partitioning, polarization, and distortion terms) that this paper adapts.","marker":"58"},{"why":"Provides the QuantumBind-RBFE protocol, the five benchmark targets and edges, and the mechanical-embedding AceFF-1.0 baseline reused unchanged.","marker":"43"},{"why":"Defines the RESP charge model the network is trained to predict, chosen for consistency with AMBER-family force fields.","marker":"14"},{"why":"Defines the AM1-BCC fixed-charge baseline that the dynamic RESP charges replace in the electrostatic coupling.","marker":"12"},{"why":"Supplies the AceFF dataset of QM conformations used for training and the single-task AceFF-2 benchmarks used for force-accuracy comparison.","marker":"61"},{"why":"Provides the TensorNet2 neural network implementation the joint energy-force-charge model is built on.","marker":"66"},{"why":"Provides the Alchemical Transfer Method framework the electrostatic embedding is implemented in, and the context that motivates Thole damping.","marker":"74"},{"why":"Sources the five benchmark protein targets whose experimental binding data the RBFE results are compared against.","marker":"1"}],"fun_headline_variants":["Electrostatic MLIP/MM slashes TYK2 RBFE error, others unchanged","First alchemical test: dynamic ligand charges cut TYK2 error in half","TYK2 binding error halved by electrostatic embedding, other targets flat","Electrostatic embedding wins on TYK2 only in RBFE production test","MLIP/MM electrostatic embedding: TYK2 improved, four targets neutral"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the long-range, reciprocal-space part of the electrostatic interaction differs negligibly between the dynamic RESP charges and the static AM1-BCC baseline, so correcting only the short-range direct-space term is enough; if that premise fails, every free-energy correction in the paper is systematically biased.","fun_headline_variants_meta":{"raw":{"variants":["Electrostatic MLIP/MM slashes TYK2 RBFE error, others unchanged","First alchemical test: dynamic ligand charges cut TYK2 error in half","TYK2 binding error halved by electrostatic embedding, other targets flat","Electrostatic embedding wins on TYK2 only in RBFE production test","MLIP/MM electrostatic embedding: TYK2 improved, four targets neutral"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000295,"raw_usage":{"total_tokens":1778,"prompt_tokens":1074,"completion_tokens":704,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":690,"completion_tokens_details":{"reasoning_tokens":599}},"tokens_in":690,"tokens_out":704,"duration_ms":6403,"temperature":1.0,"reasoning_tokens":599,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:01:08.777823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the TYK2 alchemical edges with the dynamic charges coupled into the full particle mesh Ewald sum, including reciprocal-space gradients, or alternatively with static RESP charges replacing AM1-BCC while keeping the original MLIP/MM scheme; if the $\\Delta\\Delta G$ RMSE reverts toward 0.86 kcal/mol, the reported improvement is an artifact of the short-range-only approximation or of the charge definition change, not of geometry-dependent electrostatics.","supporting_citations":[{"cited_title":"ChemRxiv , year =","cited_arxiv_id":null,"evidence_quote":"Supplies the AceFF dataset of QM conformations used for training and the single-task AceFF-2 benchmarks used for force-accuracy comparison."}],"review_version":1}