{"id":"2e29dcb0-7d4e-4537-9a25-a2278f9ed2ed","arxiv_id":"2501.08205","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Under simulated NISQ noise, QSVC is the most robust classifier for genomic data, while Peg-QSVC and QNN are sensitive and PauliFeatureMap is the most fragile encoding.","lead":"This paper simulates six types of quantum noise on four quantum machine learning classifiers and three feature-encoding circuits, using a reduced genomic sequence dataset. It reports that the quantum support vector classifier stays accurate under noise while pegasos and neural-network variants degrade, and that the Pauli feature map is the most fragile.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central robustness ranking rests on qualitative figure summaries with no reported numeric accuracies, error bars, or experimental parameters, and the paper concedes the feature maps ignore sequence order, leaving the claim unsubstantiated.","rationale":"The reader's weakest assumption concerned representativeness: that a small PCA-reduced subset with default hyperparameters transfers to real genomic classification. I agree this is a serious limitation, but the more immediate, load-bearing issue is that the manuscript provides no quantitative evidence at all for the central ranking. Figures 9–14 are described only qualitatively; no accuracies, error bars, or run counts appear in the text. This means the claim is not merely non-representative—it is currently unverifiable even for the reported experiment. The paper's own concession in Section IV.A that the feature maps cannot preserve sequential information further undermines the genomic relevance, because the experiment reduces 200-bp sequences to four PCA components and discards order. My proposed test settles both issues at once: if the ranking fails to reproduce with error bars or disappears without PCA, the conclusion is an artifact of the setup. The reader's verdict of CONDITIONAL remains appropriate; the paper is plausible but the evidence as written is insufficient, so I do not change the verdict. A separate minor issue is the apparent swap of p0 and p1 in the thermal relaxation channel equations (Section II.D, around the evolved density matrix), which does not directly affect the empirical claim if Qiskit's built-in noise channel was used, but should be corrected.","tokens_in":17727,"tokens_out":8279,"duration_ms":78836,"concrete_test":"Obtain the authors' released code and data (or re-implement from Sections III and IV) and re-run the full comparison for depolarizing and amplitude-damping noise at p=0.1: all four algorithms times three feature maps, using 10 random seeds and 1000 shots per circuit, reporting mean test accuracy plus or minus one standard deviation. Include a control without PCA (e.g., using raw k-mer frequencies or one-hot 200-bp sequences) and a classical RBF-SVM baseline. If QSVC's error bars overlap those of Peg-QSVC or QNN, or if the ranking changes in the no-PCA control, the robustness ordering is not established by this manuscript.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim—QSVC is notably robust, Peg-QSVC and QNN are more sensitive, and PauliFeatureMap is especially vulnerable—is supported only by narrative descriptions of Figures 9–14 in Section IV.B. No numeric accuracy, standard deviation, number of runs, or statistical test is given in the text, so it is impossible to judge whether the observed ordering exceeds shot noise or seed-to-seed variation. The experimental setup is also underspecified: the subset size of the 100,000-sequence Genomic Benchmarks dataset, the train/test split, the shot count, and the noise placement (which gates, what error rates per gate) are not reported, despite the abstract claiming a 'systematic evaluation.' Section IV.A explicitly concedes that the feature maps 'lack the ability to preserve the sequential information of the genomics data,' meaning the task is not genuine genomic sequence classification after PCA to four dimensions. Consequently, the apparent robustness ordering could be an artifact of PCA preprocessing and default Qiskit hyperparameters (the paper states 'all parameters set to their default values'), rather than an intrinsic property of the algorithms. If the ranking is not statistically significant even in the reported setup, or if it vanishes when raw sequence features are used, the conclusion that feature-map selection and noise mitigation are 'critical' for genomic classification is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies how six single-qubit quantum noise channels (dephasing, amplitude damping, depolarizing, thermal relaxation, bit-flip, and phase-flip) affect four QML algorithms (QSVC, Peg-QSVC, QNN, VQC) and three feature maps (ZFeatureMap, ZZFeatureMap, PauliFeatureMap) on a binary genomic classification task. Data from the Genomic Benchmarks 'democoding vs intergenic' dataset are PCA-reduced to four dimensions, encoded into four qubits, and simulated in Qiskit AerSimulator at error rates 0.01-0.3. The paper reports qualitative observations that QSVC is notably robust, Peg-QSVC and QNN are more sensitive especially to depolarizing and amplitude-damping noise, and PauliFeatureMap is especially vulnerable, and it concludes that feature-map selection and noise mitigation are critical for genomic classification on NISQ devices.","tokens_in":17960,"tokens_out":4634,"duration_ms":47391,"significance":"If the robustness ordering were quantitatively established, this would be a useful empirical data point for QML benchmarking on noisy simulators, and the paper's noise inventory is reasonably complete. The strengths are the systematic scope of six noise models, the consistent comparison across four algorithms and three feature maps, and the use of a real genomic benchmark. However, as presented the central ordering is supported only by narrative descriptions of figures, with no numeric accuracies, error bars, or statistical tests; the experimental protocol is underspecified; and the PCA preprocessing makes the connection to genomic sequence classification indirect. The central claim is therefore plausible but not demonstrated.","major_comments":[{"comment":"The central ordering (QSVC robust, Peg-QSVC/QNN sensitive, PauliFeatureMap vulnerable) is supported only by qualitative statements such as 'shows the most stability' and 'testing accuracy varies significantly'; no numeric accuracies, standard deviations, number of independent runs, or statistical tests are reported. Without these, the reader cannot distinguish a genuine effect from shot noise or seed-to-seed variation. Please report mean and standard deviation (or confidence intervals) for each model/feature-map/noise combination, and apply a paired comparison or equivalent test to support the claimed ranking.","section":"Section IV.B, Figs. 9-14"},{"comment":"The experimental protocol is underspecified. The text says 'We used a subset' but gives no subset size, train/test ratio, stratification, or sequence-selection rule; Section IV.B mentions 'varying shots' without listing shot counts; and the noise application is not stated precisely, including which gates are affected, whether single- and two-qubit gates have separate error rates, and whether the error rate is per gate or per circuit. This prevents replication and makes the Abstract's 'systematic evaluation' claim unverifiable. Please add a full configuration table with dataset subset, split, shots, and noise placement.","section":"Section IV, experimental setup"},{"comment":"Because each 200-base-pair sequence is reduced to four PCA components before encoding, and the paper itself concedes that the feature maps 'lack the ability to preserve the sequential information of the genomics data,' the experiments do not actually classify genomic sequences in the usual sense. The observed robustness ranking could be an artifact of PCA preprocessing and the four-dimensional numerical encoding. Please either test a sequence-appropriate encoding (for example, k-mer or positional encodings) or explicitly reframe the conclusions as applying to PCA-reduced feature vectors rather than to genomic sequence classification.","section":"Section IV.A"},{"comment":"The paper does not report noiseless test accuracies or a classical baseline. 'Robust' is a comparative statement; without the noise-free test accuracy of each model and without a classical SVM or other baseline on the same PCA features, a small drop under noise could still mean a model that is already inaccurate when noiseless. Please include noiseless test accuracies and a classical baseline to calibrate the robustness claims.","section":"Section IV.B, baseline comparison"}],"minor_comments":[{"comment":"The phrase 'we have get' should be corrected to 'we get'.","section":"Section II.D"},{"comment":"The word 'disccused' in 'as disccused above' is a typo and should read 'discussed'.","section":"Section III.C"},{"comment":"The caption 'dephase damping noise' should read 'dephasing (phase damping) noise' for consistency with the main text.","section":"Fig. 11 caption"},{"comment":"The table entry 'Changes diagonal elements based on flip probability' is inconsistent with Section II.F, where phase-flip noise leaves populations |α|^2 and |β|^2 unchanged and instead multiplies the off-diagonal terms by (1-2p); the table should be corrected.","section":"Table I, Phase-Flip row"},{"comment":"The symbol ρ'_i is used both for the noisy input state and for the post-circuit state ('ρ′_i = U(θ)ρ′_iU†(θ)'); using distinct symbols for the encoded noisy state and the transformed state would avoid confusion.","section":"Section III.D"},{"comment":"Please add a reproducibility statement with Qiskit/Aer versions and, if possible, a link to the code and exact hyperparameter defaults used.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The reader's stress-test is well-founded: the manuscript is a qualitative benchmark report whose central ordering is not yet substantiated by numerical evidence. I do not see circularity, but the underspecified experimental setup is the main risk. If the authors can provide the configuration details, numerical results with uncertainty quantification, and a proper baseline, a publishable version is feasible within the scope of the paper."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a readable parameter scan over standard Qiskit feature maps, standard noise channels, and four QML algorithms on a PCA-reduced slice of one genomic dataset. The headline ordering—QSVC robust, Peg-QSVC and QNN sensitive, PauliFeatureMap fragile—is plausible and matches general expectations, but the manuscript does not give you the numbers to check it. All accuracy claims are qualitative descriptions of Figures 9–14. No test accuracy values, no error bars, no shot counts, no train/test split size, no per-gate error placement. The dataset subsection is one sentence: \"We used a subset of the dataset.\" That is not a systematic evaluation.\n\nWhat's genuinely useful: the breadth of the sweep—six noise models × three feature maps × four algorithms—is a reasonable map of behavior on a small, simulated NISQ setting. The noise-model review is clear and correct as far as it goes. And the authors openly concede in Section IV.A that the feature maps cannot preserve sequential information, which is the right kind of honesty, even though it undercuts their own genomics framing.\n\nSoft spots: the central claim is load-bearing and unsupported. If the ordering fluctuates with shot noise or seed, the conclusion changes. Also, PCA to four dimensions means the task is not genomic sequence classification; it's classifying four PCA components. The paper admits that. The conclusion that feature-map selection and noise mitigation are \"critical\" for genomic classification is an overreach. Minor issues: some equations are sloppy (VQC section has a state-transformation typo), and the paper leans on one Genomic Benchmarks dataset, so transfer is speculative.\n\nWho this is for: someone starting QML noise benchmarks who wants a quick sense of which combinations degrade; a pilot study, not a definitive result. I'd bring it to a reading group as a reproducibility case study, not as a source of numbers.\n\nRecommendation: I'd send it to peer review with a heavy-revision expectation. The design is coherent, the missing details are fixable in a few weeks of work, and the checklist of required additions is obvious: numeric accuracy tables with error bars, dataset split and shot schedule, noise insertion details, classical baselines, and code/data release. If the authors can't or won't add those, the paper should not be accepted.","headline":"A broad but numerically thin noise sweep; the plausible robustness ranking is not actually supported by the reported evidence.","tokens_in":18462,"tokens_out":2887,"would_cite":false,"duration_ms":29151,"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":"Using six simulated noise models on a genomic classification task, this paper finds that the Quantum Support Vector Classifier stays robust while the PauliFeatureMap is the most fragile encoding.","keywords":["quantum machine learning","quantum noise models","feature maps","genome sequence classification","quantum support vector classifier","PauliFeatureMap","NISQ devices"],"falsifier":"Rerun the same four algorithms and three feature maps on the full genomic benchmark with a sequence-preserving encoding, or with PCA dimension varied from 4 to, say, 32, under hardware-calibrated T1/T2 noise; if QSVC's lead or PauliFeatureMap's vulnerability inverts, the paper's central robustness ranking is an artifact of the preprocessing and default settings.","tokens_in":1233,"feed_emoji":"🧬","tokens_out":3050,"duration_ms":71624,"temperature":0.7,"pith_summary":"This paper tries to establish that, for classifying genomic sequences on near-term noisy quantum hardware, choosing the quantum feature map matters as much as choosing the algorithm. Simulating six noise models on a genomic sequence dataset, the authors find that the Quantum Support Vector Classifier (QSVC) is consistently more robust than Pegasos-QSVC, quantum neural networks, and variational quantum classifiers, with the PauliFeatureMap degrading the most. The practical upshot is a proposed robustness ranking: prefer QSVC with the ZFeatureMap for noise resilience, and treat expressive maps like the PauliFeatureMap as needing mitigation. The paper also cautions that the studied feature maps do not preserve the sequential structure of genomic data.","feed_headline":"QSVC stays robust under quantum noise; PauliFeatureMap doesn't","feed_subtitle":"Six simulated noise types rank quantum-ML pipelines for genome classification, so the choices matter.","key_machinery":"The argument runs on the pairing of Kraus-operator noise channels with three Qiskit feature maps—ZFeatureMap, ZZFeatureMap, and PauliFeatureMap—and four classifier families: QSVC, Pegasos-QSVC, QNN, and VQC. Each noise type is modeled as a completely positive, trace-preserving map on the qubit density matrix, and the same noisy state preparations feed both kernel-based and circuit-based classifiers. The feature maps are the load-bearing variable: ZZFeatureMap and PauliFeatureMap add entanglement and multi-axis rotations for expressiveness, and that added complexity is what makes them more fragile under noise.","core_discovery":"The central claim is a robustness ordering of QML pipelines under simulated NISQ noise on a protein-coding versus non-coding genome classification task. QSVC remains notably robust under all six noise models, Pegasos-QSVC and QNN are more sensitive, particularly to depolarizing and amplitude-damping noise, and the PauliFeatureMap is especially vulnerable. The authors also find that training accuracy stays relatively stable while testing accuracy drops as noise increases, indicating that noise damages generalization more than fitting. They conclude that feature-map selection and noise mitigation are critical for genomic classification on NISQ devices.","pith_inferences":["The QSVC advantage may come from kernel smoothing averaging out noise-induced kernel perturbations, which would predict that its robustness grows with dataset size; this could be tested by scaling the subset.","On real IBM hardware the six isolated noise channels act jointly with calibration-dependent T1/T2 times, so the ordering suggests depolarizing and amplitude-damping terms will dominate observed genomic classification errors.","A natural next experiment is to replace PCA with a sequence-preserving embedding, such as k-mer or positional encoding, and see whether PauliFeatureMap's expressiveness recovers enough accuracy to offset its noise sensitivity."],"forward_implications":["On noisy near-term hardware, QSVC with the ZFeatureMap is the safer default for genomic classification.","Depolarizing and amplitude-damping noise should be the first targets for mitigation because they cause the largest drops in testing accuracy.","Expressive encodings such as the PauliFeatureMap should not be paired with Pegasos-QSVC or QNN without error mitigation.","The train/test accuracy gap implies that noise hurts generalization more than model fitting, so noise-aware validation is necessary.","Sequence-aware feature maps are needed before QML can exploit genomic order; the maps studied here lose sequential information."],"supporting_citations":[{"why":"Supplies the democoding versus intergenomic genomic benchmark dataset, including the 100,000 sequences of length 200 base pairs used for classification.","marker":"[15]"},{"why":"Defines quantum feature maps and quantum kernel methods, including the ZZFeatureMap and PauliFeatureMap that the experiments compare.","marker":"[32]"},{"why":"Provides the circuit-centric encoding underlying the ZFeatureMap and the characteristic single-qubit Z-axis rotation.","marker":"[33]"},{"why":"Gives the quantum support vector machine formulation that QSVC extends to quantum kernels.","marker":"[35]"},{"why":"Introduces the Pegasos stochastic sub-gradient solver that defines Pegasos-QSVC.","marker":"[38]"},{"why":"Supplies the quantum neural network model and its expressiveness analysis that the QNN implementation is based on.","marker":"[39]"},{"why":"Provides the quantum toolkit and simulator used to implement the feature maps, algorithms, and noisy channels with default configurations.","marker":"[40]"}],"fun_headline_variants":["Quantum noise: QSVC holds, PauliFeatureMap fails","Under noise, QSVC wins, PauliFeatureMap loses","Noise harms test accuracy more than training in QML","Genomic QML: PauliFeatureMap weak against noise"],"cache_read_input_tokens":20736,"weakest_assumption_plain":"The central ordering is assumed to transfer from a small, dimension-reduced subset of one genomic benchmark, with default hyperparameters and uniform simulated noise rates, to real genomic classification on NISQ hardware.","fun_headline_variants_meta":{"raw":{"variants":["Quantum noise: QSVC holds, PauliFeatureMap fails","Under noise, QSVC wins, PauliFeatureMap loses","Noise harms test accuracy more than training in QML","Genomic QML: PauliFeatureMap weak against noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000157,"raw_usage":{"total_tokens":1156,"prompt_tokens":814,"completion_tokens":342,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":270}},"tokens_in":430,"tokens_out":342,"duration_ms":3489,"temperature":1.0,"reasoning_tokens":270,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:28:57.264957+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same four algorithms and three feature maps on the full genomic benchmark with a sequence-preserving encoding, or with PCA dimension varied from 4 to, say, 32, under hardware-calibrated T1/T2 noise; if QSVC's lead or PauliFeatureMap's vulnerability inverts, the paper's central robustness ranking is an artifact of the preprocessing and default settings.","supporting_citations":[{"cited_title":"Genomic benchmarks: a collection of datasets for genomic sequence classification,","cited_arxiv_id":null,"evidence_quote":"Supplies the democoding versus intergenomic genomic benchmark dataset, including the 100,000 sequences of length 200 base pairs used for classification."},{"cited_title":"Pegasos: Primal estimated sub-gradient solver for svm,","cited_arxiv_id":null,"evidence_quote":"Introduces the Pegasos stochastic sub-gradient solver that defines Pegasos-QSVC."}],"review_version":1}