{"id":"f76f3677-7876-4ff4-8ea3-b898cb2938a8","arxiv_id":"2506.20825","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new interface connects AMBER/QUICK with FCI and quantum-centric SQD solvers, enabling alchemical free energy corrections and the first use of SQD nuclear gradients.","lead":"The paper builds a software bridge that lets a standard molecular dynamics program call quantum chemistry solvers, including a real quantum computer, to compute small corrections to hydration free energies of three small molecules. It is a proof-of-concept for using quantum hardware inside drug-discovery-style free energy calculations.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The CI-stride design invalidates MBAR: production trajectories alternate between HF and FCI/SQD potentials every 10 steps, so no sample is drawn from a Boltzmann distribution of any state in Eq. 7, making the reported free-energy corrections uncontrolled.","rationale":"I read the paper as establishing an interface/toolchain (QUICK→PySCF/Qiskit-SQD inside sander) and using three solutes as proof-of-concept. The interface contribution is credible and independent of the numerical claim. The accuracy claim, however, depends entirely on the book-ending corrections. The reader's weakest assumption identifies the every-10th-step CI activation as the most fragile link, and I agree: it is not merely a convergence or basis-set issue but a violation of the MBAR sampling assumption. A stride=1 control is the direct, feasible test. I do not see a more load-bearing concern; the ammonia result and STO-3G limitation are acknowledged limitations, not internal inconsistencies, and the SQD-vs-FCI comparisons are internally consistent given the stride protocol. Therefore the reader's CONDITIONAL verdict stands.","tokens_in":15431,"tokens_out":5092,"duration_ms":62241,"concrete_test":"Run the identical book-ending protocol for ammonia, methane, and water with CI stride set to 1 (FCI energy and gradient every MD step) for the same six λ-windows, 1 ps production, and three replicates, and compare the resulting MBAR corrections with the stride=10 values in Table S2. The FCI active spaces are tiny (441–15,876 determinants), so this is feasible with PySCF. If the stride=10 corrections agree with stride=1 within the reported ± uncertainties, the mixed-ensemble bias is negligible; if they differ beyond those uncertainties, the MBAR analysis must be redone with samples generated under a single well-defined Hamiltonian (e.g., pure CI or pure HF with one-shot CI energy reweighting).","verdict_should_be":"UNCHANGED","load_bearing_attack":"Methods says 'The external CI solvers were activated only during the production runs, at every 10th step', and Results say FCI/SQD 'energies and gradients were computed and incorporated in the dynamics at every 10th step'. The MBAR analysis (Eq. 7) is derived under the assumption that each configuration is an equilibrium sample from one of the K λ-states with a well-defined Hamiltonian U_k. In this protocol the production potential is a deterministic alternation between HF and CI/SQD potentials; its stationary distribution is not Boltzmann for the HF state, for the CI state, or for any λ-interpolated state in the MBAR set. Thus the 'reproducible and consistent' corrections in Table S2 are not unbiased estimates of the MM→QM/MM free energy difference they claim to compute. The HF+FCI vs HF+SQD differences then conflate the SQD method error with an unquantified stride-induced ensemble bias. This is load-bearing: the entire numerical demonstration of 'enhanced accuracy' rests on these MBAR corrections, and the paper provides no control simulation with stride=1 or any demonstration that stride=10 reproduces the pure-CI ensemble.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a hybrid quantum-classical workflow that integrates full configuration interaction (FCI) and sample-based quantum diagonalization (SQD) calculations into the AMBER/QUICK QM/MM book-ending correction scheme for alchemical free energy calculations. The authors compute Hartree-Fock (HF), HF+FCI, and HF+SQD book-ending corrections to the hydration free energies of ammonia, methane, and water, and compare the results with the MNSol database. The main claims are that this is the first demonstration of quantum hardware inside an alchemical workflow, that the new interface is modular and extensible, and that the inclusion of CI-based corrections improves the accuracy of free energy predictions relative to classical MM results.","tokens_in":15714,"tokens_out":3070,"duration_ms":35994,"significance":"If the central method were validated, this would be a notable step: it is the first embedding of real quantum hardware in an alchemical free energy workflow, the first use of SQD nuclear gradients in molecular dynamics, and it provides a reusable interface between AMBER/QUICK and external CI solvers. The paper is commendably explicit about its limitations, including the failure of all quantum corrections for ammonia, and it provides code and workflow links that aid reproducibility. However, the numerical demonstration of 'enhanced accuracy' is currently not supported because the production protocol uses a time-alternating mixture of HF and CI potentials and the paper lacks a validation that this gives the same ensemble as a fully CI-driven trajectory.","major_comments":[{"comment":"The production protocol alternates between HF and FCI/SQD potentials every 10th MD step ('The external CI solvers were activated only during the production runs, at every 10th step'), and the Results repeat that FCI/SQD 'energies and gradients were computed and incorporated in the dynamics at every 10th step'. MBAR, as written in Eq. (7), assumes that each configuration is an equilibrium sample from one of the K λ-states with a well-defined Hamiltonian U_k. With a deterministic alternation between HF and CI potentials, the trajectory's stationary distribution is not Boltzmann for the HF state, the CI state, or any λ-interpolated state in the MBAR set. Consequently, the corrections in Table S2 are not unbiased estimates of the MM→QM/MM free energy difference as defined by Eq. (7). Because all numerical claims rest on these corrections, the authors must either run a control with CI stride = 1 (full CI-driven production) or provide evidence that stride = 10 reproduces the pure-CI ensemble to within the reported statistical errors.","section":"Methods, 'Book-ending simulations' and 'MBAR analysis'"},{"comment":"The claim that the workflow 'enhance[s] the accuracy of free energy predictions' is not supported by the ammonia results in Table 1: the classical MM value (-3.87 kcal/mol) underestimates MNSol (-4.29 kcal/mol) by 0.42 kcal/mol, while all three quantum-corrected values (-1.94, -2.35, and -2.20 kcal/mol) move further from the reference. The paper later narrows this to 'particularly for methane and water', but the abstract and the opening of the Results overstate the accuracy gain. The authors should either temper the abstract/conclusion wording or provide a quantitative accuracy claim that acknowledges the negative ammonia result.","section":"Abstract and Conclusions"},{"comment":"The paper does not assess whether the differences between the HF, HF+FCI, and HF+SQD corrections are statistically significant. For ammonia, Table S2 reports 1.93±0.30 (HF), 1.52±0.36 (HF+FCI), and 1.67±0.33 (HF+SQD); the pairwise differences are smaller than the combined standard errors. The statement that the corrections are 'reproducible and consistent' would be strengthened by reporting confidence intervals for the differences or a proper statistical test, especially given the small production length (1 ps per λ-window) and the fact that only every 10th step carries the CI correction.","section":"Results, Table 1 and Table S2"}],"minor_comments":[{"comment":"Eqs. (4)-(7) refer to 'Poisson-Boltzmann distribution' and 'Poisson-Boltzmann constant'; these should be 'Boltzmann distribution' and 'Boltzmann constant'.","section":"Methods, MBAR analysis"},{"comment":"The index structure in Eq. (7) is ambiguous: the numerator sums over j and i, and the denominator uses k, but the relationship between these indices and the sample count N_k is not defined precisely. Please rewrite with explicit summation indices and state what quantities are known versus iterated.","section":"Methods, MBAR analysis, Eq. (7)"},{"comment":"The phrase '2-qubits gate depth' is likely a typo for 'two-qubit gate depth'.","section":"Supplementary Information, Figure S3 caption"},{"comment":"The database name is inconsistently written as both 'MNsol' and 'MNSol'; please standardize.","section":"Throughout"},{"comment":"The production runs are only 1 ps per λ-window with a 1 fs timestep, so with a stride of 10 the CI solver is invoked only about 100 times per window. Please state explicitly the total number of CI evaluations per window and whether the reported error bars reflect these 100 samples or the full 1000 MD steps.","section":"Methods, Book-ending simulations"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope and the authors have been transparent about code and limitations. The main obstacle is the missing validation of the stride-10 mixed-Hamiltonian protocol against a fully CI-driven reference, which is necessary before the numerical free energy corrections can be interpreted. I would encourage the editor to request a revision with stride-1 control simulations or a formal reweighting justification, rather than reject outright, since the interface itself appears novel and useful."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the plumbing: the authors extended AMBER/QUICK's file-based interface so that at a user-set stride the QM calculation hands off to PySCF for FCI or to Qiskit's SQD stack for a quantum-centric CI, then feeds energies and analytic gradients back into sander. They actually ran it on ibm_strasbourg for ammonia, methane, and water, and got reproducible corrections with error bars. That is a legitimate engineering contribution and, as far as I know, the first use of real quantum hardware inside an alchemical free energy calculation. The SQD gradients are also a real step beyond prior SQD energy-only work. I want to give credit for that.\n\nThe soft spot is load-bearing, and it's the CI stride. Methods says the CI solvers fire every 10th MD step; the rest of production runs on HF. So the trajectory feels a time-varying potential that is neither the HF state, nor the CI state, nor any lambda-interpolated Hamiltonian in the MBAR set. MBAR assumes each sample is drawn from an equilibrium distribution of one of the K states with a fixed U_k. That condition is violated, so the Book-ending corrections in Table S2 are biased in an unquantified way. The paper presents no control with stride=1 or any check that stride=10 reproduces a pure-CI ensemble. This isn't a minor technicality; the entire numerical demonstration of \"enhanced accuracy\" rests on these MBAR numbers. Also, the accuracy claim is overstated: for ammonia every quantum correction moves the HFE further from MNSol, which the paper does admit, but then the conclusions still talk about moving results closer to reference. The code for the interface itself isn't released, only the external packages, so others can't easily reproduce the AMBER-to-PySCF/SQD bridge.\n\nThere are also a couple of textbook-level slips in the MBAR equations (Poisson-Boltzmann constant, a garbled Eq. 7) that suggest a hasty write-up, though they don't change the method.\n\nWho is this for? Computational chemists working on QM/MM free energy and on quantum-centric methods. The interface is a useful prototype, and the paper honestly reports its limitations. But the numerical results should not be taken at face value until the stride problem is addressed.\n\nRecommendation: send it to peer review, because the interface and the first SQD-in-MD demonstration deserve referee time and the field should have this on record. But the reviewers should push for a stride=1 control or a proper reweighting scheme, and the accuracy claims should be scaled back.","headline":"A real first: quantum hardware inside an alchemical free energy workflow—but the CI-stride scheme undermines the MBAR corrections, so the claimed numbers need a control.","tokens_in":16220,"tokens_out":1778,"would_cite":false,"duration_ms":23108,"reading_group":"maybe","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 demonstrates the first alchemical free energy calculation that uses real quantum hardware, with FCI and SQD corrections injected every tenth MD step through a book-ending MBAR workflow.","keywords":["alchemical free energy","book-ending correction","quantum-centric simulation","sample-based quantum diagonalization","QM/MM","full configuration interaction","hydration free energy","MBAR"],"falsifier":"Run the book-ending protocol with the CI stride set to one instead of ten and compare the MBAR corrections for the same three solutes; if the every-step and every-tenth-step results differ by more than the reported statistical uncertainties, the mixed Hartree-Fock and CI trajectory is not equivalent to a fully correlated one.","tokens_in":15244,"feed_emoji":"⚛️","tokens_out":10017,"duration_ms":103225,"temperature":0.7,"pith_summary":"The paper sets out to show that quantum hardware can be plugged into an alchemical free energy calculation—one that estimates a free energy difference by mutating a molecule along a coupling parameter $\\lambda$—without rebuilding the classical machinery. It uses the book-ending correction: compute a force-field-level free energy first, then add free energy differences between the MM and QM/MM descriptions at the end states, estimated with MBAR over the $\\lambda$ path. The new step is that the QM/MM description, normally Hartree-Fock or DFT, can be replaced at intervals by a full configuration interaction (FCI) solver or by a quantum-centric sample-based quantum diagonalization (SQD) workflow run on real quantum hardware. The authors benchmark this on hydration free energies of ammonia, methane, and water, and report that the CI corrections move methane and water closer to reference solvation values while ammonia moves away. If the approach holds, it opens a practical route to correlated electronic-structure corrections in binding free energy predictions for drug-like systems.","feed_headline":"First alchemical free energy run on real quantum hardware","feed_subtitle":"Book-ending MBAR corrections now mix Hartree-Fock trajectories with FCI and quantum SQD energies every ten steps.","key_machinery":"The load-bearing mechanism is the book-ending correction computed with MBAR: six $\\lambda$ windows interpolate the potential from pure MM at $\\lambda=0$ to QM/MM at $\\lambda=1$, and the free energy difference between the two descriptions is added to the classical result. The enabling mechanical piece is a CI-stride control in the MD driver's interface to the quantum engine: normally the engine evaluates Hartree-Fock energies and gradients every step, but at user-set intervals the engine writes a molecular-orbital file and hands control to an external CI solver, whose energies and gradients are read back into the trajectory. The quantum-centric route is sample-based quantum diagonalization (SQD), in which a low-depth local unitary cluster Jastrow (LUCJ) ansatz is sampled on quantum hardware, the noisy bitstrings are cleaned by a configuration-recovery loop, and the resulting determinant subspaces are diagonalized classically; the paper's contribution is extending that loop to compute nuclear gradients, so the correlated results can drive the dynamics.","core_discovery":"On the paper's own terms, the discovery is a working hybrid quantum-classical free energy protocol: the first alchemical free energy calculation in which real quantum hardware contributes correlated electronic-structure corrections, and the first time the SQD method supplies nuclear gradients inside a molecular dynamics trajectory. Production runs propagate the QM/MM trajectory with Hartree-Fock energies and gradients, and at every tenth step the calculation is redirected to either an FCI solver or an SQD backend; the returned energies and gradients re-enter the dynamics, and MBAR over the six $\\lambda$ windows turns the accumulated energies into a correction added to the classical hydration free energy. With the minimal STO-3G basis, the FCI and SQD corrections agree with each other closely and improve the reference agreement for methane and water, but overshoot for ammonia; the authors therefore present the numbers as a baseline benchmark of the interface rather than a final accuracy statement.","pith_inferences":["If the every-tenth-step injection truly reproduces a fully correlated trajectory, the cost of correlated QM/MM free energy sampling drops by roughly an order of magnitude, making the approach practical for far larger solutes than the three tested.","The same interface could be used as a method scanner: on a fixed set of sampled configurations, one could compare MBAR corrections from FCI, SQD, and cheaper approximate CI solvers before choosing where to spend sampling effort.","The strongest stress test would be a solute with genuine static correlation, where FCI and SQD corrections should diverge from Hartree-Fock and DFT; ammonia, methane, and water are too weakly correlated to reveal such a difference.","Because SQD gradients are computed only on the final configuration-recovery iteration, the trajectory's sensitivity to the stride length and gradient frequency remains untested; a stride sweep would show whether the approximation biases the sampled ensemble."],"forward_implications":["Book-ending corrections are no longer limited to Hartree-Fock or DFT; any CI backend that can return energies and gradients can be swapped into the same MD workflow.","Because SQD now supplies nuclear gradients, correlated quantum chemistry can shape molecular dynamics trajectories rather than only correcting single-point energies.","For weakly correlated solutes, FCI and SQD corrections track each other within a few tenths of a kcal/mol, indicating the quantum-centric route can stand in for exact diagonalization in this regime.","The independence of the $\\lambda$ windows means the correction phase can be parallelized across multiple quantum devices, which is the paper's stated route to scaling the workflow.","The remaining error in ammonia is attributed in the paper to the minimal STO-3G basis and to the force-field Lennard-Jones parameters, so both are the natural targets for improving accuracy next."],"supporting_citations":[{"why":"Defines the book-ending correction scheme that the workflow implements and extends.","marker":"[30]"},{"why":"Provides the MBAR estimator used to compute free energy differences between MM and QM/MM states.","marker":"[56]"},{"why":"Introduces the quantum-centric SQD method and the configuration recovery loop adapted here.","marker":"[37]"},{"why":"Supplies the LUCJ ansatz whose sampled bitstrings seed the SQD subspaces.","marker":"[61]"},{"why":"Provides the full configuration interaction solver and the integral and orbital machinery shared by both CI backends.","marker":"[62]"},{"why":"Supplies the molecular dynamics engine and QM/MM driver that the custom CI interface extends.","marker":"[45]"},{"why":"Supplies the quantum engine whose file-based interface is extended to redirect to external CI solvers.","marker":"[54]"},{"why":"Provides the experimental solvation free energies used as the benchmark for the corrections.","marker":"[64]"}],"fun_headline_variants":["Alchemical free energy meets quantum hardware","First alchemical free energy with quantum SQD","Quantum hardware joins alchemical free energy","Hybrid quantum-classical alchemical free energy","Real quantum hardware for alchemical free energy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The calculation assumes that using FCI or SQD energies and gradients only every tenth MD step, with Hartree-Fock in between, gives the same free energy difference as a trajectory driven by the correlated method at every step.","fun_headline_variants_meta":{"raw":{"variants":["Alchemical free energy meets quantum hardware","First alchemical free energy with quantum SQD","Quantum hardware joins alchemical free energy","Hybrid quantum-classical alchemical free energy","Real quantum hardware for alchemical free energy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000484,"raw_usage":{"total_tokens":2450,"prompt_tokens":1063,"completion_tokens":1387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":679,"completion_tokens_details":{"reasoning_tokens":1321}},"tokens_in":679,"tokens_out":1387,"duration_ms":13259,"temperature":1.0,"reasoning_tokens":1321,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:40:45.446037+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the book-ending protocol with the CI stride set to one instead of ten and compare the MBAR corrections for the same three solutes; if the every-step and every-tenth-step results differ by more than the reported statistical uncertainties, the mixed Hartree-Fock and CI trajectory is not equivalent to a fully correlated one.","supporting_citations":[{"cited_title":"J.; Mori-Sánchez, P.; Yang, W","cited_arxiv_id":null,"evidence_quote":"Defines the book-ending correction scheme that the workflow implements and extends."},{"cited_title":"K.; Bello, L.; Ben-Haim, Y.; Bucher, D.; Cabrera-Hernández, F","cited_arxiv_id":null,"evidence_quote":"Provides the MBAR estimator used to compute free energy differences between MM and QM/MM states."},{"cited_title":"C.; Blunt, N","cited_arxiv_id":null,"evidence_quote":"Supplies the LUCJ ansatz whose sampled bitstrings seed the SQD subspaces."},{"cited_title":"S.; Bogdanov, N","cited_arxiv_id":null,"evidence_quote":"Provides the full configuration interaction solver and the integral and orbital machinery shared by both CI backends."},{"cited_title":"H.; Yang, W.; Lee, T.-S.; Wang, J","cited_arxiv_id":null,"evidence_quote":"Supplies the molecular dynamics engine and QM/MM driver that the custom CI interface extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quantum engine whose file-based interface is extended to redirect to external CI solvers."},{"cited_title":"https://github.com/qiskit-community/ffsim, 2024; Accessed: 2024-09-01","cited_arxiv_id":null,"evidence_quote":"Provides the experimental solvation free energies used as the benchmark for the corrections."}],"review_version":1}