{"id":"5e31e3eb-0d20-4a66-a80c-e30cfd629127","arxiv_id":"2607.03513","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AquaGen is an all-atom explicit-solvent flow-matching model that yields AHFE estimates ~1 kcal/mol of MD at 4–10× lower GPU cost, with refinable samples and calibrated bootstrap uncertainties.","lead":"AquaGen generates all-atom, explicit-solvent molecular configurations that approximate MD Boltzmann samples, then estimates absolute hydration free energies via force-field energies and MBAR. If reliable at scale, it offers a faster, refinable, uncertainty-aware alternative to long alchemical MD for drug-relevant free energies.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Error cancellation and imperfect λ-conditioning can make AHFE AE look better than the underlying alchemical ensembles justify.","rationale":"The reader correctly isolates the softest load-bearing assumption: that λ-conditional flow matching yields alchemical ensembles whose adjacent-state overlaps support MBAR increments, not merely an AE that benefits from opposite-signed electrostatic vs VDW errors. The paper is unusually transparent about this (Fig. 4c, CAE definition, ablations where AE and CAE diverge, box bias). That honesty supports CONDITIONAL rather than REJECT: the method is still a genuine first at all-atom explicit-solvent PBC Boltzmann generation with end-to-end AHFE, scaling, refinement to <0.5 kcal/mol, and calibrated bootstrap uncertainties. Proprietary data and missing code/weights already limit independent verification; the cancellation/CAE issue is the main scientific reason not to treat the ~1 kcal/mol MD-surrogate result as fully settled physics. My concrete test is a direct operationalization of the reader’s weakest_assumption (AE vs CAE split + volume correction). No stronger internal inconsistency appears; the concern is about how much of the reported accuracy is physical fidelity versus cancellation and teacher matching. Verdict stays CONDITIONAL; agreement with the reader is full on the load-bearing point.","tokens_in":22252,"tokens_out":843,"duration_ms":6638,"concrete_test":"On the 218 held-out compounds (and FreeSolv/CombiSolv), recompute both AE and CAE (Eq. 12) for the 160M model with N=256 samples, and report the fraction of compounds where |AE| ≤ 1 kcal/mol while CAE > 2 kcal/mol (or electrostatic CAE > 1.5). Also re-run MBAR after a volume-correcting post-process that rescales each generated box to the MD mean length for that λ (or retrain with the auto-x0-variance prior of §4.5). If a large fraction of low-AE compounds still have high CAE, or if volume correction moves AE by >0.3 kcal/mol without lowering CAE, the “trustworthy MBAR without relying on cancellation” part of the claim is weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that AquaGen samples are close enough to MD Boltzmann ensembles that MBAR on force-field energies yields trustworthy AHFE (~1.22/0.93 kcal/mol AE) at 4–10× lower cost. That claim rests on adjacent-λ energy overlaps being faithful enough for free-energy increments to be physical, not just that the telescoping sum ΔĠ_1,K is close after cancellation. Paper §4.4–4.5 and Fig. 4c show systematic opposite-signed errors: negative ΔG errors on the electrostatic leg (k=1–5) and positive on the VDW leg (k=6–20), so AE can be much smaller than CAE (e.g. Base 40M: AE 2.21 vs overall CAE 4.76; several ablations improve AE while worsening CAE). The authors themselves attribute this to λ-conditioning that learns an “averaged” distribution, plus consistent ~0.1 Å box-length overestimation from the N(0,2) prior (§4.3, §4.5). Without refinement, public FreeSolv/CombiSolv errors are also larger (~3–4 kcal/mol). Thus the headline MD-comparable accuracy can be flattered by cancellation and teacher-force-field surrogate matching rather than uniformly faithful alchemical ensembles.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"AquaGen is presented as the first all-atom, explicit-solvent, PBC-aware generative model (flow matching on a GNN) that samples approximately Boltzmann-distributed configurations of solvated drug-like molecules (~10^3 atoms) conditioned on an alchemical parameter λ. Samples are used with force-field energy evaluations and MBAR to estimate absolute hydration free energies (AHFE) on held-out compounds, claiming ~1.22 kcal/mol mean / 0.93 kcal/mol median absolute error versus GPU MD at 4–10× lower cost, with further gains from model scaling, more samples, short MD refinement, calibrated bootstrap uncertainties, and gray-box interpretability relative to black-box regressors.","tokens_in":22662,"tokens_out":1275,"duration_ms":16302,"significance":"If the energetic fidelity and free-energy accuracy hold under broader scrutiny, this is a substantial advance: it closes a resolution gap relative to prior biomolecular generative models (Table 3) and shows that high-resolution ensemble generation can serve as a practical, refinable surrogate for alchemical MD free-energy workflows. The multi-layered evidence (energy histograms/decompositions, RDFs, tICA, H-bond trends, train/test-time scaling in Fig. 1c, uncertainty calibration, CAE ablations, and black-box baselines) and explicit gray-box advantages (refinability, inspectable samples) are genuine strengths that would matter for industrial free-energy tasks and for scaling toward ABFE or related properties.","major_comments":[{"comment":"§4.4–4.5, Fig. 4c and Table 2: The headline AHFE AE (~1.22/0.93 kcal/mol) is a telescoping sum of free-energy increments and can be substantially smaller than the cumulative absolute error (CAE) because of systematic opposite-signed errors (negative on the electrostatic leg k=1–5, positive on the VDW leg k=6–20). Several ablations improve AE while worsening CAE, and the authors themselves attribute this to imperfect λ-conditioning that learns an “averaged” distribution. For the central claim that generated ensembles are faithful enough for trustworthy MBAR free energies (not merely that the endpoint difference happens to match after cancellation), the paper needs either (i) stronger evidence that adjacent-state energy overlaps and per-leg increments are accurate, or (ii) primary reporting of CAE (or equivalent) alongside AE, with clearer caveats on when cancellation is acceptable.","section":"§4.4–4.5, Fig. 4c, Table 2"},{"comment":"§4.1, Table 1 and FreeSolv/CombiSolv columns: Without MD refinement the model’s mean AE on public experimental sets is ~3–4 kcal/mol (worse than the internal MD-matched numbers), while short refinement brings it below 1 kcal/mol. The abstract and main claims emphasize “comparable accuracy to standard GPU-based MD” and 4–10× speedup largely on the internal MD-vs-MD comparison under the same force field. The manuscript should more carefully separate surrogate fidelity to the teacher MD ensemble from transfer to experiment, and quantify how much of the claimed speedup remains after the refinement step that is needed for competitive experimental accuracy.","section":"§4.1, Table 1"},{"comment":"§4.3 and §4.5 (Gaussian prior / box-size results): There is a consistent ~0.1 Å overestimation of simulation box length attributed to the N(0,2) coordinate prior and cell prior. Because free energies and densities depend on volume, and because the cell is generated jointly, this systematic bias is load-bearing for the claim of energetically accurate Boltzmann samples; the paper should either correct it (e.g., better prior or post-hoc volume reweighting) or demonstrate that residual volume error does not materially affect the reported AHFE increments and energy distributions.","section":"§4.3, §4.5"}],"minor_comments":[{"comment":"Fig. 1c and related text: Clarify wall-clock vs GPU-time accounting when claiming 4–10× speedups, especially under parallel sampling across λ versus HREX communication overhead.","section":"Fig. 1c, §4.1"},{"comment":"§3.2 / §A.1: The water-compression scheme (O retained, H reconstructed via projection) is central to scaling; a short quantitative ablation of its effect on solvent bonded/nonbonded energies (beyond the qualitative note in §C.3) would help readers assess the approximation.","section":"§3.2, §A.1, §C.3"},{"comment":"Notation: τ is used for flow-matching time and t for physical time, but some figure captions and §C.1 still mix t/τ; unify for clarity.","section":"§3.2, §C.1"},{"comment":"Table 3 and related-work claims of “first”: The comparison is useful; a one-sentence qualification that “first at all-atom explicit-solvent Boltzmann sampling suitable for direct FF+MBAR” would avoid over-reading relative to concurrent trajectory or ensemble models.","section":"Table 3, §5"}],"recommendation":"major_revision","confidential_remarks":"The work is a strong systems contribution and likely of high interest to the journal’s audience. The main risk is over-selling endpoint AHFE AE relative to the documented cancellation and teacher-MD circularity; requiring the authors to elevate CAE and experimental-transfer discussion should fix this without changing the core technical story. Novelty relative to concurrent generative MD/ensemble papers should be checked carefully at acceptance time."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is the first generative model I have seen that keeps every atom, explicit TIP3P water, and PBCs, then produces configurations whose force-field energies are close enough to MD that you can run MBAR and get absolute hydration free energies. That is the real novelty. Prior work (AF3, BioEmu, Boltz-2, scalable flows, etc.) either drops solvent or never claims force-field-compatible Boltzmann samples. They train on >1B alchemical frames, show energy histograms and RDFs that overlap MD at λ=1, track tICA and H-bond trends across the path, and demonstrate clear gains from model size, sample count, and 40 ps MD refinement. Bootstrap uncertainties are calibrated. Against black-box regressors they are competitive, and with short refinement they beat them on FreeSolv/CombiSolv. The gray-box framing is honest and useful.\n\nSoft spots are real but not fatal. Adjacent-λ free-energy increments have opposite-signed errors (negative on electrostatics, positive on VDW), so AE can look better than CAE; the authors document this and introduce CAE themselves. Box lengths are systematically ~0.1 Å high because of the N(0,2) prior. Without refinement, public-set errors rise to ~3–4 kcal/mol. Training and primary evaluation are both against the same OpenFF/TIP3P MD teacher, so this is a high-quality surrogate, not independent experimental discovery. Proprietary data and no released code/weights limit verification. None of that erases the technical advance; it just bounds how far you can currently push the “replace MD for ABFE” claim.\n\nMath and citation pattern look solid; free parameters are ordinary generative-model knobs and are ablated. This is for people who care about free-energy methods and industrial MD workflows. I would bring it to reading group, cite it when discussing generative free-energy estimation, and send it to peer review. It deserves a serious referee.","headline":"First credible all-atom explicit-solvent Boltzmann generator at ~10³ atoms that actually feeds MBAR for AHFE; the ~1 kcal/mol headline is real but partly flattered by cancellation and teacher-force-field matching.","tokens_in":23312,"tokens_out":506,"would_cite":true,"duration_ms":4955,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"AquaGen generates all-atom, explicit-solvent molecular configurations from the Boltzmann distribution that match MD free-energy accuracy at a fraction of the cost.","keywords":["generative models","molecular dynamics","absolute hydration free energy","flow matching","explicit solvent","Boltzmann distribution","alchemical free energy","MBAR"],"falsifier":"On held-out compounds, compare cumulative absolute error (sum of absolute adjacent-λ free-energy increments) and box-length bias of AquaGen ensembles against long MD; if CAE stays large while AHFE absolute error looks small, or if short MD refinement from generated starts fails to approach MD free energies within the reported compute budgets, the claim that the generative ensembles are Boltzmann-faithful for free-energy work fails.","tokens_in":23111,"feed_emoji":"💧","tokens_out":988,"duration_ms":16223,"temperature":0.7,"pith_summary":"This paper claims that a flow-matching generative model can produce full all-atom configurations of drug-like molecules in explicit water under periodic boundaries that are close enough to true Boltzmann ensembles for free energies to be computed the same way as in molecular dynamics. Prior generative models drop solvent, use implicit solvent, or coarse-grain, so their frames are not plug-compatible with industrial force fields. By conditioning generation on an alchemical order parameter and evaluating OpenFF/TIP3P energies on the samples, AquaGen recovers absolute hydration free energies with roughly 1 kcal/mol mean error versus GPU MD on held-out compounds while using 4–10× less compute; short MD refinement from the generated starts can cut error further below 0.5 kcal/mol. A sympathetic reader cares because free energies ground decisions in drug and materials work, and a gray-box route—learned sampling plus white-box energies—gives inspectable, refinable predictions with calibrated uncertainty that pure black-box regression does not.","feed_headline":"Generative model matches MD free energies 4–10× faster","feed_subtitle":"All-atom explicit-solvent samples yield ~1 kcal/mol AHFE error and refinable gray-box predictions.","key_machinery":"λ-conditional flow matching: a graph neural network learns a velocity field that transports a Gaussian prior over atomic coordinates and a cubic simulation cell into alchemical Boltzmann configurations conditioned on the order parameter λ that gradually annihilates solute–solvent interactions; uncorrelated samples at each λ are scored with the same force field used in MD and pooled by the multistate Bennett acceptance ratio (MBAR) estimator to obtain free-energy differences.","core_discovery":"AquaGen is the first generative model to sample Boltzmann-distributed all-atom configurations that include explicit solvent and periodic boundaries for systems of thousands of atoms, with geometric and energetic fidelity high enough that MBAR free-energy estimates from force-field energies on those samples match molecular-dynamics absolute hydration free energies to about 1.22 kcal/mol mean (0.93 median) error on held-out compounds, at 4–10× lower GPU cost.","pith_inferences":["If adjacent-λ energy overlaps remain faithful at protein–ligand scale, the same pipeline could cut wall-clock cost of absolute binding free energy campaigns without abandoning force-field grounding.","Opposite-signed errors on the electrostatic versus van der Waals legs imply that better λ conditioning may improve reliability more than further model scaling alone.","Explicit-solvent generative ensembles may become a practical intermediate between structure predictors and full MD for any property that is an ensemble average of potential energies."],"forward_implications":["Absolute hydration free energy can be estimated 4–10× faster than GPU MD with comparable accuracy, and still faster when short MD refinement starts from generated frames.","Scaling model size and the number of generated samples predictably improves free-energy accuracy (train- and test-time compute).","Bootstrapped confidence intervals over generated samples give well-calibrated uncertainty without deep ensembles.","The same high-resolution ensemble approach is positioned to extend to lipophilicity, membrane permeability, and absolute binding free energy.","Gray-box predictions remain refinable and inspectable via force-field energies, unlike pure black-box regressors."],"fun_headline_variants":["AquaGen samples all-atom explicit-solvent configs matching MD AHFE 4–10× faster","First generative model for Boltzmann all-atom samples of thousands of atoms with solvent","Generative all-atom samples yield AHFE estimates comparable to MD at fraction of cost","AquaGen enables gray-box free-energy predictions from force-field energies on samples","All-atom generative model matches MD hydration free energies with refinable ensembles"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"Samples at neighboring alchemical windows must have energy overlaps faithful enough that free-energy increments from MBAR are trustworthy, not merely that opposite-signed errors along the electrostatic and van der Waals legs cancel to a small final number.","fun_headline_variants_meta":{"raw":{"variants":["AquaGen samples all-atom explicit-solvent configs matching MD AHFE 4–10× faster","First generative model for Boltzmann all-atom samples of thousands of atoms with solvent","Generative all-atom samples yield AHFE estimates comparable to MD at fraction of cost","AquaGen enables gray-box free-energy predictions from force-field energies on samples","All-atom generative model matches MD hydration free energies with refinable ensembles"]},"model":"grok-4.5","effort":"low","cost_usd":0.005848,"raw_usage":{"total_tokens":1590,"prompt_tokens":827,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":58480000,"prompt_tokens_details":{"text_tokens":827,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":671,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":827,"tokens_out":92,"duration_ms":5347,"temperature":1.0,"reasoning_tokens":671,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T01:57:27.513081+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On held-out compounds, compare cumulative absolute error (sum of absolute adjacent-λ free-energy increments) and box-length bias of AquaGen ensembles against long MD; if CAE stays large while AHFE absolute error looks small, or if short MD refinement from generated starts fails to approach MD free energies within the reported compute budgets, the claim that the generative ensembles are Boltzmann-faithful for free-energy work fails.","supporting_citations":[],"review_version":1}