{"id":"94c1a007-086c-424a-b730-d1fe5b908527","arxiv_id":"2607.06971","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Non-iterative amino-acid-level quantum sampling on IBM Heron R2 predicts 5–18-residue pocket peptides with 27–71% better RMSD than AI and VQE baselines while reconstructing approximate energy landscapes.","lead":"QSAD samples short peptide folds on IBM Heron R2 by evolving a coarse amino-acid Hamiltonian once, not by iterative VQE. On 101 binding-pocket peptides it reports lower RMSD than AlphaFold3 and other AI/quantum baselines, plus noise tolerance and a 27× runtime cut versus VQE.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The fixed lattice Hamiltonian may already place natives near low energy; classical same-H baselines then leave open how much multi-β quantum sampling (vs ranking) drives the 27–71% RMSD gains.","rationale":"The reader correctly flags the fixed lattice+implicit-side-chain Hamiltonian as the weakest assumption behind the strongest claim (hardware RMSD wins + noise/runtime advantages). Same-H classical baselines and flat post-processing RMSD (§5.7) already show sampling quality matters more than refinement, but they do not close the loop on whether H itself is a near-oracle for these pocket peptides. If crystal geometries are low-H by construction, classical ranking of a diverse valid ensemble can produce the reported medians and the quantum contribution shrinks to robust diversity generation—still useful on NISQ, but a narrower claim than “coarse-grained quantum sampling” as the practical path. The proposed native-under-H / H-minimizer check is a single, concrete diagnostic that either shores up or re-attributes the accuracy numbers without requiring new QPU runs. No contradiction of the reported Heron results is claimed; the issue is causal attribution of the 27–71% lift. That keeps the verdict CONDITIONAL (tighten Hamiltonian validation and baseline fairness) rather than REJECT, matching the reader’s posture with a sharper falsifiable test.","tokens_in":18252,"tokens_out":877,"duration_ms":9818,"concrete_test":"For all 101 sequences, map each crystal Cα trace to the nearest self-avoiding tetrahedral walk (or enumerate short-N exact folds), evaluate H, and report (i) energy percentile of that native-like walk among all valid walks and (ii) Cα RMSD of the global H-minimizer (exact for N≤8; SA/greedy with 10× budget for longer). If native-like walks are routinely in the bottom ~5% of H and the H-minimizer already has median RMSD ≲~2.7 Å, the accuracy claim is largely Hamiltonian fidelity + ranking, not multi-β quantum sampling.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim attributes near-native Cα geometry (median ~2.7 Å) and 27–71% gains over AI/VQE primarily to non-iterative multi-β Hamiltonian evolution on Heron R2 (Eqs. 3–6; §4.2), not classical post-processing. Section 5.7 shows stage-wise RMSD is flat after lattice decoding and that QSAD beats Random/Greedy/SA on the same H under matched pool size (median 2.51 vs 4.44–4.65 Å). That still leaves the weakest link: whether the fixed tetrahedral + implicit side-chain H (§4.1; α_pair/α_node/α_loc/α_ster from MJ-style tables, no per-sequence tuning) already scores crystal-like self-avoiding walks as low-energy for these 101 pocket peptides. If natives (or near-natives) systematically sit in the low-energy tail of this H, then energy-based ranking of any diverse valid pool can recover good Cα RMSD; the quantum circuit mainly supplies diversity under noise, and the headline “quantum sampling architecture” over-attributes the accuracy lift. The paper acknowledges a ~1.6 Å lattice ceiling and heuristic β={1,2,3,4}, but does not report the energy rank or RMSD of the crystal conformation under H, nor an oracle that returns the lowest-H valid lattice structure independent of sampling.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript introduces QSAD, a quantum–classical pipeline that encodes short binding-pocket peptides (5–18 residues) as a coarse-grained six-term Hamiltonian on a tetrahedral lattice with implicit side-chain coefficients, then samples via non-iterative multi-β Hamiltonian phase evolution (randomized EfficientSU2 ansatz + Trotterized e^{-iβH}) on IBM Heron R2. Classical decoding, diversity-preserving ranking, and all-atom reconstruction yield predicted Cα backbones and approximate energy landscapes. On 101 experimentally resolved peptides the method reports median Cα RMSD ~2.7 Å, pairwise wins against AF3, ColabFold variants, ESMFold, OmegaFold, OpenFold, and VQE (e.g. 95/101 vs AF3, 52/55 vs VQE), ~27× lower mean quantum runtime than VQE, and simulated noise resilience up to ε=2% where VQE degrades. Ablations claim that structural quality is fixed at the lattice-sampling stage and that stratified multi-β sampling beats random/greedy/SA pools on the same Hamiltonian.","tokens_in":18732,"tokens_out":1642,"duration_ms":24063,"significance":"If the attribution holds, the work is a concrete demonstration that utility-scale superconducting hardware can deliver competitive short-peptide structure prediction in a regime where MSA/language-model methods are known to be weak, with practical advantages in runtime predictability and noise tolerance over iterative VQE. Strengths that should be credited include: end-to-end execution of 101 cases on Heron R2 with reported telemetry; stage-wise RMSD flatness showing post-processing does not move Cα; matched-pool classical samplers on the same H; and explicit landscape reconstruction with funnel/basin diagnostics. These are rare for quantum biology claims and make the empirical package falsifiable and useful even if some design choices remain heuristic.","major_comments":[{"comment":"§4.1 and §5.7 (weakest load-bearing link): The central claim attributes near-native Cα geometry primarily to non-iterative multi-β sampling of H, not to classical ranking. Section 5.7 shows flat RMSD after lattice decoding and 94–95% wins vs Random/Greedy/SA under matched pool size, which is necessary but not sufficient. The manuscript never reports (i) the energy of the crystal (or best lattice-aligned crystal) conformation under H, (ii) its rank among valid self-avoiding walks, or (iii) an oracle that returns the lowest-H valid lattice structure independent of sampling. Without these, it remains possible that the fixed MJ-style coefficients already place near-natives in the low-energy tail for these 101 pocket peptides, so that any diverse valid pool plus energy ranking recovers good RMSD. Please add energy-rank / oracle-RMSD diagnostics (at least for N≤10 where enumeration or aggressi","section":"§4.1, §5.7"},{"comment":"§4.3.2 ranking vs Hamiltonian: Candidate ranking reuses Miyazawa–Jernigan contact energy, burial, and related physics terms that are closely related to H_pair / H_node. Combined with energy-based selection in the classical-sampler comparison (Fig. 17), this creates mild circularity between the model that generates samples and the score that picks the reported structure. Clarify which ranking terms are independent of H, report an ablation that ranks by sampling frequency alone (or by a held-out score), and state whether the reported median 2.51 Å end-to-end RMSD still holds under frequency-only selection.","section":"§4.3.2, Fig. 17"},{"comment":"§5.6 noise resilience: The claim of tolerating noise “3–5× beyond typical hardware error rates” rests on a single 6-residue (18-qubit) depolarizing simulation (Table 3, Fig. 11). Real-hardware VQE failure is shown for one protein (Fig. 2b), but QSAD’s multi-β ensemble is not re-characterized under calibrated Heron noise models or at N>6. Either extend the noise study to a few longer peptides with device-calibrated noise, or narrow the abstract/claim language to the simulated 6-residue setting and the qualitative real-hardware VQE contrast.","section":"§5.6, Table 3, Abstract"},{"comment":"Abstract / §5.3 “27–71% improvement”: The percentage range is not derived transparently from the reported medians/means (e.g. 2.7 vs 4.8 Å vs AF3 is ~44% relative reduction; other baselines differ). State the exact formula (relative median RMSD reduction per baseline? min–max over methods?) and which baseline endpoints produce 27% and 71%, so the headline number is reproducible from Table/Fig. 7.","section":"Abstract, §5.3, Fig. 7"}],"minor_comments":[{"comment":"Title and running headers appear concatenated without spaces (“QUANTUMSAMPLINGARCHITECTURE…”, “APREPRINT- JULY9, 2026”); fix typesetting.","section":"Title page"},{"comment":"Eq. (5): UH(β)=e^{-iβ H_D} e^{-iβ H} with H_D=∑X_j is a transverse-field driver; briefly note relation to QAOA-style mixing and whether the order of factors is intentional.","section":"§4.2.1, Eq. (5)"},{"comment":"Table 1 QPU times jump sharply with N; a short note on transpiled two-qubit gate counts (not only ansatz depth) would help readers assess scaling.","section":"Table 1"},{"comment":"§4.4 funnel score and basin count are useful but the 35th-percentile threshold and PCA variance (60–75%) should be sensitivity-checked or justified so landscape claims are not threshold-dependent.","section":"§4.4"},{"comment":"Discussion correctly notes the ~1.6 Å lattice ceiling and heuristic β={1,2,3,4}; consider moving a one-sentence statement of the ceiling into the abstract or evaluation so readers do not over-interpret sub-2 Å claims.","section":"Discussion, Abstract"},{"comment":"References include related prior work by overlapping authors (QDockBank, hybrid quantum-AI frameworks); ensure self-citation is balanced with independent lattice-folding and quantum-protein literature already cited (e.g. Robert et al., Boulebnane et al.).","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The empirical hardware campaign and same-H classical controls are stronger than typical quantum-biology submissions; the main risk is over-attribution of accuracy to “quantum sampling” without an oracle/energy-rank check of H. If the authors supply those diagnostics and the natives are not systematically low-energy under H, the paper becomes much more convincing. Scope fits a methods/quantum-computing venue; for a pure structural-biology journal the lattice ceiling and pocket-only Cα metric would need more emphasis. No integrity red flags from the text alone."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The headline result is real hardware work: 101 binding-pocket peptides (5–18 residues) run end-to-end on IBM Heron R2, median Cα RMSD ~2.7 Å, pairwise wins of 95/101 vs AF3, 101/101 vs ESMFold/OpenFold, 52/55 vs VQE, ~27× mean quantum time cut vs VQE, and noise resilience where iterative VQE collapses. That package is new as a system, not as a single equation.\n\nWhat they did well is the controls. Stage-wise RMSD is flat after lattice decoding, so classical reconstruction is not moving the backbone. Ablations (random ansatz → single-β → multi-seed multi-β) improve coverage and ESS. Same-Hamiltonian classical baselines (random, greedy, SA) under matched pool size lose badly on near-native yield and final RMSD. Noise is tested fairly on a simulator with explicit signal-retention mapping. Landscape reconstruction is secondary but coherent. Citations to lattice folding, MJ potentials, Robert et al., and their own QDockBank line are appropriate; this is incremental engineering on known pieces, not a rewrite of folding theory.\n\nSoft spots, in proportion. The load-bearing premise is that a fixed tetrahedral lattice plus literature-derived implicit side-chain coefficients already puts native-like walks in the low-energy tail for these pockets. They never report the crystal conformation’s energy rank or RMSD under H, nor an oracle lowest-H lattice structure. If natives sit low, energy ranking of any diverse valid pool can look good and the multi-β circuit mainly supplies diversity under noise. β={1,2,3,4}, ranking weights, and AI defaults are heuristic; no code/data release is stated. Lattice ceiling ~1.6 Å is acknowledged. These are real gaps for attribution, not contradictions of the measured RMSD tables.\n\nWho it is for: quantum applications people and pocket-focused computational chemists who care about NISQ-scale sampling, not general structure prediction. The math is standard Trotter + EfficientSU2; the data are the point. I would send it to peer review. Tighten baseline fairness, release bitstrings/telemetry or at least crystal-under-H ranks, and the claim becomes much cleaner. Worth engaging.","headline":"Real Heron R2 runs on 101 pocket peptides with clear RMSD wins over AF3/VQE and same-H classical samplers; the open question is how much the fixed lattice H already favors natives.","tokens_in":19392,"tokens_out":568,"would_cite":true,"duration_ms":6953,"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":"QSAD predicts short binding-pocket peptide structures on current quantum hardware by sampling a coarse-grained amino-acid Hamiltonian without iterative optimization, beating AI and VQE baselines.","keywords":["quantum sampling","protein structure prediction","binding-pocket peptides","coarse-grained Hamiltonian","non-iterative Hamiltonian evolution","NISQ hardware","energy landscape","tetrahedral lattice"],"falsifier":"On a held-out set of pocket peptides with crystal structures, if top-ranked or lowest-energy QSAD lattice conformations systematically exceed ~5 Å Cα RMSD from native while classical search (e.g. simulated annealing) on the identical Hamiltonian recovers near-native states, the claim that quantum sampling of this model is what delivers the accuracy would be falsified.","tokens_in":19103,"feed_emoji":"🧬","tokens_out":1004,"duration_ms":26964,"temperature":0.7,"pith_summary":"Short peptides in protein binding pockets are hard to predict because they lack the long-sequence patterns that data-driven fold predictors rely on, so the problem reduces to physics-based search of a complex energy landscape. This paper presents QSAD, which encodes that landscape as an amino-acid-level Hamiltonian on a tetrahedral lattice and samples it with non-iterative Hamiltonian evolution rather than variational optimization. Run end-to-end on IBM Heron R2 for 101 peptides of 5–18 residues, QSAD reports lower backbone error than AlphaFold3, language-model and MSA fold predictors, and VQE, while staying usable under noise levels that break iterative quantum methods and finishing far faster than VQE. The same measurement ensemble is used to sketch approximate energy landscapes. The claim is that coarse-grained, one-way quantum sampling is already a practical path for this regime on utility-scale superconducting processors.","feed_headline":"Quantum sampling beats AI on short pocket peptides","feed_subtitle":"Non-iterative lattice Hamiltonian runs on a 156-qubit chip cut error and runtime versus AlphaFold3 and VQE","key_machinery":"QSAD: an amino-acid-level coarse-grained Hamiltonian (tetrahedral-lattice backbone turns plus implicit side-chain contact, burial, local-propensity, and steric coefficients) sampled by randomized-ansatz state preparation followed by multi-β Trotterized Hamiltonian phase evolution and classical decoding/ranking—not iterative variational optimization.","core_discovery":"On 101 experimentally resolved binding-pocket peptides (5–18 residues), fully executed on IBM Heron R2, non-iterative multi-β sampling of a fixed amino-acid-level tetrahedral-lattice Hamiltonian yields median Cα RMSD near 2.7 Å and improves accuracy by 27–71% over evaluated AI and quantum baselines, while recovering ground-state energy under noise several times typical hardware rates and reducing mean quantum execution time by about 27× relative to VQE.","pith_inferences":["The same coarse-grained sampling recipe may transfer to other short interface or disordered peptides outside the binding-pocket benchmark once lattice validity is checked.","Because lattice discretization sets a precision floor near 1.6 Å, hybrid pipelines that lift lattice Cα traces into continuous force fields become the natural next bottleneck, not deeper quantum circuits alone.","Predictable batch (non-session) quantum job models may become preferred for production bioinformatics workloads where VQE’s long-lived optimizer sessions are hard to schedule.","The reported rise in coverage under depolarizing noise suggests controlled noise could be treated as a deliberate exploration knob rather than only a defect to mitigate."],"forward_implications":["Short peptides that lack useful MSA or evolutionary signal can be structure-predicted on existing superconducting processors without waiting for fault-tolerant machines.","Non-iterative quantum sampling becomes preferable to VQE for noise-sensitive structure tasks because errors do not compound through an optimizer loop.","A single sampling run can supply both a predicted backbone and an approximate multi-basin energy landscape for binding and mutational interpretation.","Amino-acid-level encoding keeps qubit demand inside current 156-qubit devices for peptides up to about 18 residues, unlike orbital-level formulations.","Fixed physics-encoded Hamiltonians can outperform purely learned predictors precisely when sequence signal is weak."],"fun_headline_variants":["Quantum sampling tops AI on short pocket peptides by 27-71%","Non-iterative lattice sampling hits 2.7 Å RMSD on 101 peptides","Heron R2 runs cut peptide error and VQE time by 27×","QSAD Hamiltonian sampling beats AI and VQE on pocket peptides","Coarse-grained quantum sampling reconstructs short peptide folds"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single fixed tetrahedral-lattice energy model with literature-derived contact and burial weights, never tuned per sequence, places true native backbones among the low-energy states that multi-β sampling can actually hit.","fun_headline_variants_meta":{"raw":{"variants":["Quantum sampling tops AI on short pocket peptides by 27-71%","Non-iterative lattice sampling hits 2.7 Å RMSD on 101 peptides","Heron R2 runs cut peptide error and VQE time by 27×","QSAD Hamiltonian sampling beats AI and VQE on pocket peptides","Coarse-grained quantum sampling reconstructs short peptide folds"]},"model":"grok-4.5","effort":"low","cost_usd":0.004728,"raw_usage":{"total_tokens":1354,"prompt_tokens":753,"num_sources_used":0,"completion_tokens":80,"cost_in_usd_ticks":47280000,"prompt_tokens_details":{"text_tokens":753,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":521,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":753,"tokens_out":80,"duration_ms":5266,"temperature":1.0,"reasoning_tokens":521,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T20:02:46.433240+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a held-out set of pocket peptides with crystal structures, if top-ranked or lowest-energy QSAD lattice conformations systematically exceed ~5 Å Cα RMSD from native while classical search (e.g. simulated annealing) on the identical Hamiltonian recovers near-native states, the claim that quantum sampling of this model is what delivers the accuracy would be falsified.","supporting_citations":[],"review_version":2}