{"id":"214577f6-b689-492b-a303-8d61a3ed9d32","arxiv_id":"2507.18425","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A 9-qubit parameterized quantum circuit predicts protein-ligand binding free energies with 2.37 kcal/mol RMSD and 0.65 Pearson r on the PDBbind Core set, while tolerating shot noise and light gate noise.","lead":"Researchers trained a nine-qubit quantum circuit to predict how well drug-like molecules bind to proteins, reaching 2.37 kcal/mol error on a standard benchmark. The result is a feasibility test for near-term quantum hardware, but it does not yet beat classical machine-learning baselines.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Test-set-based depth selection makes the headline 2.37 kcal/mol / 0.650 result an optimistically biased estimate, not an unbiased measure of generalization.","rationale":"The reader's weakest assumption focused on the single-qubit readout expressiveness, but the more load-bearing problem is that the headline test numbers appear to be selected after peeking at the test set. This is a standard selection-bias issue: the same Core set is used both to choose the depth and to report the accuracy, so the reported optimum is an optimistic upper bound. The paper has real strengths—the PyTorch state-vector implementation makes the training concrete, the code is linked, and the noise and sampling experiments are informative as a robustness check. However, those strengths do not fix the fact that the central numeric claim is not an unbiased estimate. My concern does not change the reader's verdict: the appropriate decision remains CONDITIONAL, requiring a validation-based selection protocol and ideally a classical baseline. I am not alleging anything improper; the manuscript simply does not document a leak-free model-selection rule, and as written the reported 'best at six units' language implies test-set comparison.","tokens_in":9657,"tokens_out":8398,"duration_ms":95251,"concrete_test":"Split the PDBbind Core set into a validation subset (e.g., 100 complexes) and a genuinely held-out test subset (the remaining 185). For each depth from 1 to 10 and each learning rate, train with several random seeds, select depth/learning rate by validation RMSD, then evaluate the selected model once on the held-out test set. Report mean and standard deviation over seeds, and compare against a classical baseline (e.g., linear regression or a 3D-CNN) trained on the same 512 occupancy features with the same split. If the held-out RMSD/PCC differs materially from 2.37/0.650, or if the classical baseline matches it within seed noise, the headline claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—six QMLunits yield RMSD 2.37 kcal/mol and PCC 0.650 on the PDBbind Core test set—is presented after inspecting test-set performance across depths 1–10. The Results section states that 'the test-set performance was with six QMLunits' and that 'beyond this depth, both metrics began to decline,' which is only knowable after scoring every depth on the Core set. If the depth was selected because it looked best on the test set, the reported number is a post-selection maximum, and the true generalization error is unknown and likely higher. The manuscript does not document a pre-registered or validation-based rule for choosing six units, and the model-selection sentence in Methods only says the lowest-training-RMSD run was retained. In addition, no variance across the three random initializations is reported for the headline metric, so 'six beats five' could be seed noise. Finally, there is no classical baseline trained on the same 512 occupancy features, so even an unbiased 2.37/0.650 would not establish that the quantum circuit, rather than the voxel features, is responsible for the predictive signal.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a parameterized quantum circuit model, built from repeated QMLunits, to predict protein–ligand binding free energies in structure-based virtual screening. Protein–ligand complexes from PDBbind v2020 are voxelized into 512 occupancy values that are encoded into a nine-qubit circuit; the model is trained with mean squared error and evaluated on the PDBbind Core set under full simulation, finite-shot sampling, and noisy simulation. The main claim is that with six QMLunits the model achieves an RMSD of 2.37 kcal/mol and a Pearson correlation of 0.650 on the test set, with relative ranking largely preserved under noise. The paper also proposes a parallel estimation scheme using ancillary qubits to process multiple complexes simultaneously.","tokens_in":9922,"tokens_out":4979,"duration_ms":51810,"significance":"If the central empirical result is reliable, the paper would demonstrate that a compact nine-qubit parameterized circuit can learn a transferable binding-affinity scoring function from structural features, which is a useful step toward NISQ-era virtual screening. The manuscript is relatively self-contained in its data-preparation and simulation methodology, and it ships a GitHub repository with the code. The strengths are the use of an external benchmark (PDBbind) with a standard train/test split and the evaluation under three realistic inference conditions. However, the significance is currently limited by an optimistic test-set-based model selection, lack of statistical error bars, absence of any classical baseline on the same features, an underspecified feature-to-qubit encoding, and a conceptually flawed parallel-estimation argument.","major_comments":[{"comment":"The headline result (six QMLunits, RMSD 2.37 kcal/mol, PCC 0.650) is selected after evaluating depths 1–10 on the test set, as shown in Fig. 4 and stated in the text ('The test-set performance was with six QMLunits'). The Methods, however, state that the retained model is the one with the lowest training-set RMSD, which does not define a rule for choosing the depth. Because the depth is chosen on the same data used to report the final metric, the reported numbers are post-selection maxima and are optimistically biased estimates of generalization. Please either pre-register a validation-based depth-selection rule (e.g., using a separate validation split) or report the test-set metrics for all depths along with the selection procedure, and discuss the resulting optimistic bias.","section":"Results, 'Estimation of ΔGbind Using Full Probability'; Fig. 4; Methods, 'Training Protocol'"},{"comment":"No statistical uncertainty is reported for any metric. The Methods state that three independent training runs are performed, but only the single run with the lowest training RMSD is retained, and no variance across seeds is shown. Consequently, differences between adjacent depths (e.g., five vs. six QMLunits) could be within seed noise. Please report means and standard deviations over the random initializations for the test-set RMSD and PCC, and justify that the reported depth is stable across seeds.","section":"Results (Fig. 3) and Methods, 'Training Protocol'"},{"comment":"The central claim of this section—that the outputs O^j for all 2^m inputs can be 'computed simultaneously' using a block-diagonal unitary—is not valid as a quantum algorithm. In quantum mechanics, applying a block-diagonal U to a superposition of index states entangles the index register with the data register; a single measurement collapses the state and reveals only one bitstring, not all 2^m outputs. To recover all outputs, one would need to repeat the experiment many times and incur measurement overhead, so the factorization in Eq. (5) is a classical vectorized matrix operation, not a demonstration of quantum parallel prediction. This directly affects the scalability motivation in the Introduction and Conclusions. Please either correct the claim to describe a classical batch-processing trick or provide a concrete quantum measurement scheme that avoids the collapse issue.","section":"Methods, 'Parallel Estimation of Multiple Protein–Ligand Complexes'; Eqs. (4)–(5); Fig. 5"},{"comment":"The manuscript never specifies how the 512 occupancy values are encoded into the initial state of the nine-qubit circuit. The text only says the 'normalized occupancy vector was encoded into a quantum state' and names three type qubits and six position qubits, but the mapping from 512 floats to the initial amplitudes or rotation angles is absent. This is a load-bearing technical detail: without it the architecture is not reproducible from the manuscript, and it is impossible to tell whether the predictive power comes from the circuit, the encoding, or the readout. Please state the encoding explicitly (e.g., angle encoding, amplitude encoding, or a learned classical preprocessing layer) or provide the precise circuit definition used in the code.","section":"Methods, 'Quantum Circuit Architecture' and 'Training Protocol'"},{"comment":"No classical baseline is trained or evaluated on the same 512 occupancy features. As a result, the reported RMSD of 2.37 kcal/mol and PCC of 0.650 cannot be attributed to the quantum circuit specifically; a classical model (e.g., ridge regression, random forest, or a feedforward network) on the same voxel features might achieve comparable or better accuracy. Adding such a baseline is necessary to support the paper's stated motivation that the quantum approach offers a practical alternative for virtual screening. Without it, the claims about quantum advantage or even quantum competitiveness are not yet established.","section":"Results and Discussion (throughout)"}],"minor_comments":[{"comment":"The Abstract states 'even when using 100,000 shots' while the Results section also reports 10,000 and 1,000,000 shots. Please align the abstract with the full set of shot counts or explicitly state that 100,000 is the representative setting.","section":"Abstract and Results"},{"comment":"The caption lists colors (black, red, orange, blue) but the legend in the figure itself is hard to distinguish in the printed version; consider adding markers or line styles and enlarge the font.","section":"Fig. 3 caption"},{"comment":"The notation 'U29∗29' is unclear; it should be written as a 2^9 × 2^9 unitary (or 512 × 512) with proper subscripts, and the dimensions of the identity matrices in Eq. (4) should be specified.","section":"Eq. (4) and surrounding text"},{"comment":"Reference 37 is cited for the Core subset, but the appropriate citation is the CASF benchmark papers or the original PDBbind follow-up descriptions; the current reference is about deep learning in virtual screening and does not describe the Core set.","section":"Methods, 'Dataset Preparation for Training and Validation'"},{"comment":"The statement 'adding just 200 more qubits could enable the simultaneous evaluation of up to 10^60 complexes' is overstated and depends on the parallel-estimation scheme criticized above; please temper this claim or remove it until a valid quantum measurement protocol is provided.","section":"Introduction, last paragraph"}],"recommendation":"major_revision","confidential_remarks":"The empirical part of the paper may be salvageable, but the parallel-estimation section is conceptually incorrect and should not be presented as a quantum advantage. The lack of a classical baseline and the test-set-based model selection are the main scientific weaknesses; both are fixable within a revision. I recommend major revision rather than rejection because the core idea—training a small PQC on voxelized structures for binding-affinity prediction—is plausible and the paper includes a reproducible code repository."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read: this is a legitimate application of an existing PQC building block (the QMLunit from Zhao et al. 2021) to the PDBbind binding-free-energy task, packaged with a clean noise/sampling robustness study and a GitHub code link. The interesting result is not the absolute numbers; it's that the ranking metric (PCC) stays roughly constant when you add shot noise or light depolarizing/amplitude-damping noise. That is a useful sanity check for anyone thinking about near-term QML scoring.\n\nWhat is new is modest: the architecture is borrowed, the voxel encoding is straight out of classical 3D-CNN scoring, and the task is standard. The contribution is the specific 9-qubit readout design and the three-mode evaluation. Fine as a feasibility note.\n\nThe soft spots are real but not fatal.\n\nFirst, the depth of six QMLunits is chosen after looking at the test-set curves. The text says 'the test-set performance was with six QMLunits' and that beyond it metrics declined. That makes the 2.37 kcal/mol / 0.650 a post-selection maximum, not an unbiased estimate. You need a validation split or nested CV, plus variance across the random initializations. Without that, 'six beats five' is possibly seed noise.\n\nSecond, there is no classical baseline trained on the same 512 occupancy features. A linear or MLP model on those same features would anchor whether the quantum circuit is adding anything. Without it, the headline could be the features doing all the work.\n\nThird, the parallel estimation section is not right. Putting m ancilla qubits around a 9-qubit circuit and writing a block-diagonal matrix gives you a superposition of outputs, not simultaneous classical answers; measuring collapses to one. Also the 'U29*29' notation looks like a typo for 2^9. This section should be corrected or removed.\n\nMinor: the intro's claim that 200 more qubits could evaluate 10^60 complexes is hand-waving. The single-qubit readout is a limitation worth noting but not disqualifying: one real number per complex is weak, but the experiment is what it is.\n\nWho is this for? People working on QML for drug screening who want a quick feasibility data point. It deserves serious peer review because the topic is timely and the code/data are promised, but it needs major revision before publication: baselines, a proper model-selection protocol, error bars, and a fixed parallelism section.","headline":"A plausible but under-validated QML feasibility study: the headline RMSD/PCC is selected on the test set, and without classical baselines the quantum component isn't actually shown to be doing the work.","tokens_in":10420,"tokens_out":2748,"would_cite":false,"duration_ms":30611,"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":"A nine-qubit parameterized quantum circuit, trained on voxelized protein–ligand structures, predicts binding free energies with an RMSD of 2.37 kcal/mol and Pearson correlation of 0.650 on the PDBbind Core set, and keeps ranking intact…","keywords":["quantum machine learning","binding free energy prediction","structure-based virtual screening","parameterized quantum circuits","protein-ligand binding","voxel encoding","PDBbind","NISQ robustness"],"falsifier":"Retrain the same six-unit circuit after randomly permuting the 512 voxel occupancies across the training complexes; if the held-out RMSD and Pearson correlation stay near 2.37 kcal/mol and 0.650, the reported accuracy does not come from the structural encoding.","tokens_in":9420,"feed_emoji":"⚛️","tokens_out":10083,"duration_ms":100604,"temperature":0.7,"pith_summary":"This paper proposes that a small parameterized quantum circuit can serve as a structure-based virtual screening scoring function: nine qubits encode a coarse three-dimensional occupancy grid of a protein–ligand complex, and the measurement of a single qubit is converted directly into a predicted binding free energy. The author trains this circuit on 19,158 complexes and reports that six circuit units predict the held-out Core set with an RMSD of 2.37 kcal/mol and a Pearson correlation of 0.650. The ranking of ligands remains stable under finite-shot sampling and under simulated amplitude-damping and depolarizing noise, which the paper reads as evidence that ranking-based screening can run on near-term quantum hardware. The paper also sketches how ancillary index qubits could let one circuit process many complexes in parallel, a route toward screening very large chemical libraries.","feed_headline":"Nine-qubit circuit predicts binding energies at 2.37 kcal/mol","feed_subtitle":"Quantum circuit trained on PDBbind keeps ligand rankings even under sampling noise and gate errors.","key_machinery":"The central object is the QCbind circuit built from repeated QMLunits, where each unit contains an Lpar layer of parameterized RX and RZ single-qubit rotations followed by an Lbreaker layer of CNOT gates with a distinct entanglement topology per qubit index. Nine qubits are used: three encode atom-type information and six encode spatial positions in a discretized 4×4×4 voxel grid derived from 512 occupancy values. Instead of measuring all qubits, only the 0th qubit is measured, and the predicted binding free energy is read out as $\\Delta G = 100(y_0-y_1)$, where $y_0$ and $y_1$ are the probabilities of measuring $|0\\rangle$ and $|1\\rangle$. Training minimizes the mean squared error between this readout and experimental free energies, and a parallel extension uses $m$ ancillary index qubits to process $2^m$ complexes through a block-diagonal unitary.","core_discovery":"The central claim is that a relatively shallow nine-qubit circuit—six stacked QMLunits, each combining single-qubit rotations with CNOT entanglement layers—can learn a transferable binding-energy scoring function from voxelized structures. On the PDBbind Core test set, the trained circuit reaches an RMSD of 2.37 kcal/mol and a Pearson correlation of 0.650, with accuracy improving up to six units and declining beyond. Sampling 100,000 measurement shots from a single qubit keeps predictions close to the ideal full-state results, with a maximum Pearson correlation of 0.632. Under simulated amplitude damping and depolarizing noise, the absolute error worsens to about 2.97 kcal/mol but the correlation holds at 0.653, which the paper interprets as preservation of ligand ranking rather than exact energy estimation.","pith_inferences":["The paper does not compare the quantum model against classical voxel-based scoring functions on the same split; a head-to-head benchmark would clarify whether the quantum circuit adds predictive power or merely matches an encoding-equivalent classical model.","Because the readout uses only one qubit, a natural testable extension is to measure additional qubits or learn the readout observable, which may improve absolute accuracy without requiring deeper circuits.","The noise robustness is demonstrated at two fixed low rates; scaling the damping and depolarizing rates would reveal how much noise the ranking guarantee can tolerate before it breaks.","The parallel-screening argument assumes ideal hardware and an efficient classical-to-quantum encoding, so practical advantage over classical screening remains a projection rather than a demonstrated outcome."],"forward_implications":["Five to six QMLunits are the useful circuit depth: predictive accuracy improves up to six units and declines afterward, so near-term quantum circuits do not need to be deep to be useful.","A single-qubit readout with about 100,000 measurement shots closely reproduces full-state predictions, substantially lowering the measurement overhead required for inference.","Under the tested noise levels, absolute prediction errors grow but the Pearson correlation remains around 0.65, meaning ligand ranking can still be trusted even when absolute energies are degraded.","Adding ancillary index qubits allows one circuit to evaluate many complexes at once through vectorized block-diagonal matrix operations, offering a concrete scaling path for high-throughput virtual screening."],"supporting_citations":[{"why":"Supplies the PDBbind v2020 dataset of 19,443 protein–ligand complexes with experimental pKd values used for training.","marker":"(36)"},{"why":"Provides the Core subset of 285 complexes used as the held-out test set for evaluation.","marker":"(37)"},{"why":"Supplies the molecular density occupancy encoding formalism, including the voxel occupancy function used to turn atomic coordinates into input features.","marker":"(13)"}],"fun_headline_variants":["Shallow quantum circuit predicts binding energies with 2.37 kcal/mol error","Six QML units keep ligand ranking under noise, 0.65 correlation","Quantum ML on nine qubits hits 0.65 correlation for affinity","Voxel-based QML: robust to noise, RMSD 2.37 kcal/mol on PDBbind"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that reading out one qubit's measurement probabilities after only nine qubits encode a 512-voxel occupancy vector leaves enough chemical information to reproduce the experimental binding free energy ($\\Delta G$).","fun_headline_variants_meta":{"raw":{"variants":["Shallow quantum circuit predicts binding energies with 2.37 kcal/mol error","Six QML units keep ligand ranking under noise, 0.65 correlation","Quantum ML on nine qubits hits 0.65 correlation for affinity","Voxel-based QML: robust to noise, RMSD 2.37 kcal/mol on PDBbind"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000643,"raw_usage":{"total_tokens":2943,"prompt_tokens":914,"completion_tokens":2029,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":1939}},"tokens_in":530,"tokens_out":2029,"duration_ms":14526,"temperature":1.0,"reasoning_tokens":1939,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:12:30.916027+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same six-unit circuit after randomly permuting the 512 voxel occupancies across the training complexes; if the held-out RMSD and Pearson correlation stay near 2.37 kcal/mol and 0.650, the reported accuracy does not come from the structural encoding.","supporting_citations":[],"review_version":2}