{"id":"9418dcc4-2d53-4818-975a-a8111ddcb321","arxiv_id":"2505.00037","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"Applying a simulated quantum support vector machine to proteomic and metabolomic COVID-19 data gives classification performance comparable to a classical SVM in selected settings, but the paper's biomarker-identification claims are not established.","lead":"Simulated quantum support vector machines were tested on blood proteomics and metabolomics data to classify COVID-19 patients and long COVID subgroups. The paper finds that some quantum kernels match or slightly beat classical SVMs, but the claim that QSVM identifies key biomarkers is not supported by the design.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ridge-regression feature ranking is performed before cross-validation, so the Group 1 vs Group 4 biomarker-importance result is in-sample and unsupported.","rationale":"The reader's weakest_assumption identifies the same core issue: the ridge ranking is not nested in the cross-validation, and Cleveland repeated measures are treated as independent. My reading of the full appendix supports this as the decisive weakness; the highlighted tables in the main text are a small selection from a large set of results, many of which are chance-level, so 'consistently' is inflated. The central claim as stated is therefore not supported. A revision with nested feature selection and subject-level blocking could still show QSVM competitive on some tasks, but the current design does not establish that QSVM preserves biomarker importance order or consistently outperforms CSVM. The verdict should remain REJECT.","tokens_in":43599,"tokens_out":5093,"duration_ms":54246,"concrete_test":"Re-run the pipeline with the ridge-regression ranking nested inside the cross-validation loop: for each training fold, fit ridge on that fold only, select Group 1 and Group 4, then evaluate CSVM and QSVM on the held-out fold; for the Cleveland Clinic data, block folds by subject so repeated samples from one patient never appear in both train and test. If the average Group 1 minus Group 4 AUC gap shrinks toward zero or reverses, the biomarker-importance claim fails. Also tabulate how many of the 96 conditions actually show QSVM AUC >= CSVM AUC across all appendix entries, rather than only highlighted rows.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that QSVM reflects the importance rankings by ridge regression rests on Group 1 (top-ranked biomarkers) yielding higher AUC than Group 4 (bottom-ranked biomarkers). This result is not out-of-sample because the ridge regression in Section 2.3 ranks biomarkers on the full dataset before the k-fold splits in Section 2.5, and Section 3.2 selects stable common biomarkers from that ranking. The test labels are therefore visible during feature-group selection, so the Group 1 versus Group 4 AUC gap can reflect overfitting to idiosyncratic noise rather than genuine predictive importance. The Group 4 comparison is meant to be a negative control, but a negative control is only valid when the ranking is computed independently of the evaluation folds. In the Cleveland Clinic data this leakage is compounded by splitting repeated samples from the same patient across folds, an issue the Discussion acknowledges but does not correct. The appendix also shows many QSVM conditions at or near chance (for example amplitude encoding in Tables B.3 through B.8), while the main text highlights only conditions where QSVM matched or exceeded CSVM, so the consistency claim is not supported by the full record. The most load-bearing defect, however, is the un-nested feature ranking: without nested validation, the paper's signature biomarker-ranking result is indistinguishable from in-sample selection.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript investigates whether a simulated quantum support vector machine (QSVM) can classify COVID-19 status or long-COVID subtypes from proteomic and metabolomic biomarker panels, and whether the quantum model preserves a ridge-regression-based ranking of biomarker importance. Two datasets are used: the Su et al. INCOV cohort and a Cleveland Clinic cohort. Biomarkers are ranked by ridge-regression coefficients, divided into high- and low-importance groups, and used as inputs to classical SVM and to QSVM with amplitude, angle, ZZ-feature-map, and projected quantum kernels. Performance is measured by AUC under k-fold cross-validation. The authors report that QSVM consistently matched or exceeded CSVM and that high-importance groups outperform low-importance groups.","tokens_in":43871,"tokens_out":6323,"duration_ms":65992,"significance":"If the results were fully supported, the paper would provide a useful empirical data point on quantum-kernel SVM for small multi-omics biomedical datasets, with a biologically grounded biomarker-selection workflow. The study covers multiple quantum embeddings, two independent cohorts, and makes code available at a public repository. The main contribution, however, is an evaluation claim about QSVM viability and biomarker-importance consistency; both are compromised by the evaluation protocol and by selective reporting of the full result tables, as detailed below.","major_comments":[{"comment":"The biomarker-importance claim rests on a feature-selection procedure that is not nested inside the cross-validation loop. Ridge regression is applied to each full classification task in Section 3.2 to rank biomarkers and define Group 1 and Group 4; the same groups are then used in Section 2.5 to train and evaluate classifiers under k-fold cross-validation. Because the ridge fit uses the labels of all samples before the folds are formed, the Group 1 versus Group 4 comparison is an in-sample separation rather than an out-of-sample test. The negative control is only valid if the ranking is computed independently of the evaluation folds. This issue is load-bearing for the claim that QSVM 'reflects the importance rankings by ridge regression.' The Cleveland Clinic analysis is further affected by splitting repeated samples from the same patient across folds, as acknowledged in the Discussion but not corrected.","section":"Section 3.2 and Section 2.5"},{"comment":"The abstract's 'consistently' claim is contradicted by the full results in Appendix B. Many QSVM conditions produce AUC at or near chance, and some below chance: for example, Table B.3 amplitude encoding gives 0.4430, 0.3618, and 0.3175 for Type2 vs. Type1; Table B.4 gives 0.3401 and 0.3011; and Tables B.5–B.8 contain numerous exactly-0.5000 entries (e.g., Tables B.5 and B.6). The main text and Tables 6–11 highlight only the conditions in which QSVM matched or exceeded CSVM, so the reader cannot infer consistent comparable-or-better performance. The authors should report the full distribution of AUC values, a summary statistic across conditions, and a comparison that includes all quantum kernels, not only the favorable bold entries.","section":"Appendix B and Section 3.4"},{"comment":"No uncertainty or significance testing accompanies the performance comparisons. Reported AUC values are averaged over folds without confidence intervals or significance tests, and many apparent QSVM improvements are tiny: for example, Table 6 rows 1–2 show 0.987 vs. 0.987 and 0.985 vs. 0.984, and Table 7 row 1 shows 0.990 vs. 0.989. Given small sample sizes and repeated measures, these differences are within noise. The claim that QSVM 'exceeded' CSVM requires either error bars or a statistical test; at minimum, the comparisons should be described as descriptive rather than as evidence of superiority.","section":"Tables 6–11 and Section 3.4"}],"minor_comments":[{"comment":"The sentence beginning 'However, Despite the use of QML algorithms' contains a capitalization error and is grammatically incomplete; it should be revised to a complete sentence, e.g., 'However, the use of QML algorithms for analyzing multi-omics datasets and identifying COVID-19-related biomarkers remains unexplored.'","section":"Section 1"},{"comment":"The entry for the number of long-COVID individuals appears as '1271', which is likely a typo for 127; the table should be checked and the notation 'meta2/proteo3' clarified.","section":"Table 1.1"},{"comment":"The statement that long-COVID subgroup classification 'reaching a maximum AUC of 87% (See Table 6 and Table 7)' cites the wrong tables: Tables 6 and 7 report Healthy vs. T1 comparisons, not T3 subgroup comparisons.","section":"Section 4.1"},{"comment":"The sentence 'angle encoding suppressed CSVM in the proteomic domain' appears to be a typo; 'suppressed' should presumably read 'surpassed'.","section":"Section 4.1"},{"comment":"The displayed quantum kernel formula 'STφ(x)Sφ(x)US' is garbled and should be written as an inner product such as ⟨φ(x)|φ(x')⟩.","section":"Section 2.4(b)"},{"comment":"Class-label names are not harmonized across tables: for example, Table A.5 uses 'Naive vs. Inter' and 'Type2 vs. Inter' while Table 4 uses 'Intermediate vs. Naive' and 'Intermediate vs. Type2'; consistent naming would improve reproducibility.","section":"Appendix A"}],"recommendation":"reject","confidential_remarks":"The paper has a clear experimental structure and a public code link, but the combination of un-nested feature selection and the selective highlighting of favorable QSVM results makes both central claims unsupported as written. A re-analysis with nested feature selection and full reporting of all experimental conditions would be required before the claims could be evaluated fairly; as it stands, the manuscript's main conclusions do not survive the full appendix record."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a straightforward application-level benchmark of QSVM kernels against classical SVM on two COVID-19 multi-omics datasets. The numerical comparison is fine as far as it goes, but the paper's headline claim — that QSVM consistently matches or beats CSVM while 'reflecting' ridge-regression importance rankings — overstates what the design supports. The feature-ranking step is done on the full data before cross-validation, so the Group 1 vs Group 4 comparison is in-sample selection, not out-of-sample validation. The authors acknowledge repeated measures in Cleveland Clinic but don't correct for it. And the appendix shows many QSVM conditions at or near chance (amplitude encoding especially), while the main text highlights the best cells. So the consistency claim isn't supported by the full record.\n\nWhat's genuinely good: the paper is transparent about simulation-only results, releases code on GitHub, and reports full AUC tables in the appendix — that's more than most QML papers do. Testing four quantum kernels (amplitude, angle, ZZ, projected) across two independent cohorts is a legitimate contribution to the QML-for-omics benchmarking conversation, even if no new algorithm or kernel is introduced. The discussion of selected biomarkers (LGALS9, MMP7, quinolinate, etc.) is reasonable and connects to prior biology.\n\nSoft spots, in order of severity. First and load-bearing: the ridge regression ranks biomarkers on the full dataset, then stable groups are selected, then classifiers are trained on those same groups. Labels leak into feature-group construction, so the finding that Group 1 outperforms Group 4 is not a valid negative-control demonstration. This is fixable: put the ranking inside each cross-validation fold (nested feature selection). Second: the Cleveland Clinic dataset includes repeated samples from the same patients; the Discussion notes this but the analysis doesn't split by subject, so CV AUCs are optimistic. Third: no confidence intervals or significance tests; many reported AUC differences are within noise (e.g., 0.709 vs 0.711). Fourth: the main text selects highlighted conditions while the appendix contains many chance-level QSVM results; the abstract's 'consistently' is not accurate. The QSVM-vs-CSVM comparison itself is not circular, and the paper never claims quantum advantage — that's good, because none is shown.\n\nWho it's for: readers interested in whether quantum kernels can be dropped into small biomedical classification problems, and referees wanting a concrete example of how feature-selection leakage can invalidate an otherwise honest benchmark. It deserves a serious referee — the data and code are real, the flaw is identifiable and fixable, and a revised version with nested validation and subject-level splitting could be a solid application note. I'd send it to review, with a request for major revision.","headline":"A clean QSVM-vs-CSVM benchmark on two COVID-19 omics datasets, undermined by an un-nested feature selection step that makes the biomarker-importance claim circular.","tokens_in":44414,"tokens_out":1635,"would_cite":false,"duration_ms":17484,"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":"Simulated quantum SVM matches classical SVM on COVID-19 biomarker classification while preserving biomarker importance rankings.","keywords":["quantum support vector machine","COVID-19 biomarkers","multi-omics","ridge regression","quantum kernel","projected quantum kernel","biomarker ranking","long COVID classification"],"falsifier":"Rerun the same pipeline with the ridge regression ranking computed only on the training fold (or on a held-out ranking set), and with cross-validation folds defined by individual patient rather than by sample; if the Group-1-versus-Group-4 AUC gap and the QSVM superiority shrink or disappear, the central biomarker-ranking claim would not survive.","tokens_in":43417,"feed_emoji":"🧬","tokens_out":3843,"duration_ms":37227,"temperature":0.7,"pith_summary":"This paper tests whether a quantum support vector machine trained on ridge-regression-selected proteomic and metabolomic biomarkers can classify COVID-19 status and long-COVID subtypes. Using two independent cohorts, the authors rank hundreds of biomarkers by ridge-regression coefficient magnitude, split them into high- and low-importance groups, and compare classical SVM with several QSVM quantum kernels after PCA reduction to 2–16 dimensions. Across the 16 classification tasks, simulated QSVM models matched or exceeded CSVM in most settings where the comparison was highlighted, and the high-importance biomarker group consistently outperformed the low-importance group in both algorithms. The paper argues this shows QSVM is a viable alternative for small multi-omics classification and preserves the biological signal captured by ridge regression.","feed_headline":"Simulated quantum SVM matches classical SVM on COVID-19 biomarkers","feed_subtitle":"Multi-omics tests show quantum kernels reaching AUC about 0.99 while preserving ridge-regression biomarker rankings.","key_machinery":"The argument is carried by the kernel-based SVM trained on PCA-reduced biomarker groups. Biomarker importance is defined by the absolute coefficients of a ridge regression with an L2 penalty, which ranks all features and divides them into fixed-size groups. The QSVM replaces the classical RBF or polynomial kernel with four quantum kernels: amplitude encoding, angle encoding, the ZZ feature map, and the projected quantum kernel (PQK), each evaluated at 2, 4, 8, and 16 principal components or qubits. The load-bearing comparison is the AUC difference between Group 1 (top-ranked biomarkers) and Group 4 (bottom-ranked biomarkers) under k-fold cross-validation, with class weights applied to address imbalance.","core_discovery":"The central claim is that a QSVM with an appropriate quantum kernel can classify COVID-19 cases from proteomic and metabolomic profiles at least as well as a classical SVM, while preserving the feature-importance ordering produced by ridge regression. On the INCOV healthy-versus-T1 task, angle-encoding and angle-with-PQK kernels produced AUC values of about 0.99, matching or slightly exceeding the best RBF-based CSVM; in long-COVID subtype comparisons and in the Cleveland Clinic dataset, several QSVM configurations also showed higher AUC than CSVM. The authors interpret the consistent Group-1-over-Group-4 gap as evidence that the quantum models retain the biologically meaningful biomarker ranking rather than fitting noise.","pith_inferences":["The reported biomarker-ranking result likely overstates generalization because the ridge regression ranks features on the full cohort before the train/test split; a nested cross-validation or a separate ranking-only training set would give more honest out-of-sample AUC for the Group-1 group.","Repeated samples from the same Cleveland Clinic patients are treated as independent cross-validation units, so the healthy-versus-COVID AUC differences may be inflated by intra-subject correlation; a leave-one-individual-out evaluation would test this.","If the same experiments were rerun with nested CV and patient-level grouping, the QSVM-versus-CSVM margin might shrink, but the paper's qualitative claim of comparability would likely survive; the strongest testable prediction is that angle-encoding kernels retain an edge over RBF at low PCA dimensions.","The choice of 2–16 principal components means the 'quantum' part operates on a heavily compressed input; an extension would be to compare QSVM with raw biomarker input or with a classical kernel on the same compressed features to isolate what the quantum kernel adds."],"forward_implications":["If the central claim holds, QSVM is a drop-in alternative to classical SVM for small multi-omics classification tasks, with no loss in discrimination when the kernel is chosen appropriately.","The angle-encoding and angle-with-PQK kernels appear most consistently competitive, suggesting that feature encoding choice matters more than raw qubit count for these datasets.","The Group-1-versus-Group-4 AUC gap gives a template for validating quantum classifiers against a classical feature-ranking reference before deploying them on new omics data.","Numerical parity in noise-free simulation is a prerequisite for, not evidence of, quantum advantage; the next step implied by the paper is testing on noisy simulators and real hardware.","Classification of long-COVID subtypes is harder (AUC around 0.7) than acute infection detection (AUC above 0.9), so the method's practical value is greater for early diagnosis than for subtype stratification."],"supporting_citations":[{"why":"Supplies the longitudinal INCOV multi-omics dataset with four long-COVID subtype labels used for most classification tasks.","marker":"[26]"},{"why":"Introduces the ZZ feature map and supervised learning with quantum-enhanced feature spaces, one of the tested quantum kernels.","marker":"[35]"},{"why":"Introduces the projected quantum kernel (PQK) and demonstrates its expressibility, the basis for the PQK variants tested here.","marker":"[36]"},{"why":"Provides the ridge regression formulation whose coefficients define the biomarker importance groups.","marker":"[28]"},{"why":"Establishes the quantum support vector machine concept that the paper adapts for kernel-based classification.","marker":"[16]"},{"why":"Supports the choice of SVM for small datasets, motivating the QSVM comparison.","marker":"[29]"},{"why":"Provides encoding patterns for quantum algorithms, including the amplitude and angle encoding schemes used in the study.","marker":"[33]"}],"fun_headline_variants":["Simulated QSVM matches classical SVM on COVID-19 biomarkers","Quantum kernels match classical for COVID-19 biomarker ranking","QSVM hits 0.99 AUC on COVID-19 multi-omics test","Quantum SVM rivals classical for COVID-19 biomarker tests","QSVM preserves biomarker importance in COVID-19 omics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that ridge regression's biomarker importance ranking, computed on the full cohort, can be used to define training groups for cross-validated classifiers without leaking information; the paper also assumes repeated samples from the same individual are independent cross-validation units.","fun_headline_variants_meta":{"raw":{"variants":["Simulated QSVM matches classical SVM on COVID-19 biomarkers","Quantum kernels match classical for COVID-19 biomarker ranking","QSVM hits 0.99 AUC on COVID-19 multi-omics test","Quantum SVM rivals classical for COVID-19 biomarker tests","QSVM preserves biomarker importance in COVID-19 omics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000535,"raw_usage":{"total_tokens":2538,"prompt_tokens":875,"completion_tokens":1663,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1577}},"tokens_in":491,"tokens_out":1663,"duration_ms":12974,"temperature":1.0,"reasoning_tokens":1577,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:16:30.026372+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the same pipeline with the ridge regression ranking computed only on the training fold (or on a held-out ranking set), and with cross-validation folds defined by individual patient rather than by sample; if the Group-1-versus-Group-4 AUC gap and the QSVM superiority shrink or disappear, the central biomarker-ranking claim would not survive.","supporting_citations":[],"review_version":1}