{"id":"1f64f8a4-f502-44a0-bc20-60c3e5ccc504","arxiv_id":"2501.04869","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A five-gene expression signature, replicated in two cohorts, is associated with improved overall survival in ovarian cancer patients who received bevacizumab, though the formal interaction tests were not significant after multiple-testing correction.","lead":"This paper reports a gene expression signature that may identify ovarian cancer patients who live longer when bevacizumab is added to chemotherapy, based on a new 181-patient transcriptome cohort and a 377-patient validation cohort. It matters because oncologists currently lack a reliable biomarker for choosing bevacizumab maintenance therapy.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"External validation is circular: DASL was used to filter biclusters and to compute feature importances that prioritized bicluster 84, so its interaction p-values do not provide independent confirmation.","rationale":"The reader's weakest assumption was confounding by indication in the UKE discovery cohort. That is a real problem, but the UKE is only the discovery set; the paper's main support for the predictive claim is the DASL external cohort. The most load-bearing issue is that DASL is not external in the statistical sense: it was used to filter the candidate biclusters and to compute the feature importances that selected bicluster 84. Consequently, the DASL interaction p-values are in-sample after a data-driven selection, so they overstate the evidence. The two-gene overlap further weakens the claim that the same biological signature was validated. These flaws jointly mean the central claim — improved OS with bevacizumab in both cohorts — is not supported by the analysis as presented. The paper is transparent and the underlying data could be reanalyzed with a proper holdout, so the appropriate assessment is that the current claim should not be accepted; hence REJECT rather than CONDITIONAL, while acknowledging that a corrected pipeline could yield a promising candidate.","tokens_in":30412,"tokens_out":4307,"duration_ms":42888,"concrete_test":"Hold DASL out completely until the signature is fixed: define bicluster 84 from UKE alone (or via UKE-only cross-validation), then fit the Methods Cox interaction model in DASL for this fixed signature. Separately, compute the same treatment-interaction test in DASL for all 234 UKE biclusters and report the rank and Benjamini-Hochberg adjusted p-value of bicluster 84 across all 234. If bicluster 84 is not significant after adjustment (or its rank is not extreme), the reported DASL validation is an artifact of the selection step.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim requires that the DASL cohort independently validates the bicluster 84 signature. The manuscript's own pipeline violates this independence. In 'Unsupervised patient stratification', the 23 candidate biclusters are defined as those 'presenting in both UKE and DASL cohorts' — i.e., DASL is used to filter the 234 UnPaSt biclusters. In 'Random forest prioritizes predictive bicluster candidates', the permutation-based feature importances in Table 3 are explicitly computed 'when the corresponding feature column in the validation dataset was permuted', so DASL survival outcomes directly determine which bicluster (84) is prioritized. The Cox interaction test in Table 4 is then run on the same DASL data. The reported DASL adj.p-value 0.07 therefore does not account for selection among 23 candidates, let alone 234 biclusters. Additionally, DASL contains only two of the five bicluster 84 genes (SLCO6A1, PTH2R), so the 'replication' is at best a two-gene proxy. The UKE cohort alone cannot support the causal predictive claim because treatment was not randomized and the groups differ in prognostic factors. Thus the strongest claim — benefit in both cohorts — rests on an in-sample DASL analysis, not a true external validation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Olga Zolotareva and colleagues generate a new RNA-seq cohort of ovarian cancer patients (UKE, n=181 with survival data) and combine it with the published DASL cohort (n=377) to search for transcriptomic signatures that identify patients who benefit from bevacizumab added to platinum-based chemotherapy. Using UnPaSt biclustering, they identify 234 biclusters in the UKE cohort, retain 23 that are detectable in DASL, and use random survival forest permutation importance to prioritize candidate signatures. One signature, bicluster 84 (five genes in UKE, two measurable in DASL), shows a large survival benefit from bevacizumab in signature-positive patients in both cohorts (UKE HR=0.41, DASL HR=0.51) but no significant benefit in signature-negative patients. The authors also show that the signature is associated with a set of 14 differentially expressed genes across three cohorts and propose a CTCFL-driven stemness mechanism. The central claim is that bicluster 84 is a predictive biomarker of bevacizumab response.","tokens_in":30654,"tokens_out":6380,"duration_ms":56932,"significance":"A validated predictive biomarker for bevacizumab in ovarian cancer would be of substantial clinical value, because current guidelines do not specify which patients should receive maintenance bevacizumab and the drug has meaningful toxicity. The paper contributes a novel, publicly available RNA-seq dataset and a reproducible computational pipeline with code, and it is commendable that the authors report the non-significant interaction p-values in Table 4 and acknowledge the limited cohort size. However, the study as presented is hypothesis-generating rather than confirmatory: the interaction between treatment and signature, which is the appropriate test of predictive value, does not reach significance in the discovery cohort and is only borderline in the validation cohort after adjustment. Moreover, as detailed below, the DASL cohort is not an independent validation set because it was used to filter biclusters and to compute feature importances that prioritized bicluster 84. If the authors reframe the claims accordingly and provide a non-circular validation strategy, the dataset and signature would be a useful contribution to the biomarker literature.","major_comments":[{"comment":"The manuscript's central claim that bicluster 84 is validated in the independent DASL cohort is undermined by the fact that DASL data were used twice before the validation test. In 'Unsupervised patient stratification', the 23 candidate biclusters are defined as those 'presenting in both UKE and DASL cohorts' (i.e., DASL is used to filter the 234 UnPaSt biclusters). In 'Random forest prioritizes predictive bicluster candidates', permutation-based feature importances in Table 3 are computed 'when the corresponding feature column in the validation dataset was permuted', so DASL survival outcomes directly influence which bicluster (84) is selected. The Cox interaction p-value for DASL in Table 4 (adj.p=0.07) is therefore an in-sample statistic, not an independent replication. No adjustment is made for selection among 23 candidates, let alone the original 234 biclusters. The authors should either pre-specify the signature using UKE data alone and then test it in DASL without any DASL-informed filtering, or clearly label the DASL analysis as a second discovery/prioritization step and seek a genuinely independent cohort for validation.","section":"Random forest prioritizes predictive bicluster candidates; Methods, Replication of biclusters"},{"comment":"The appropriate statistical test for a predictive biomarker is the treatment-by-biomarker interaction in a Cox model. In the UKE discovery cohort this interaction is not significant even before multiple-testing correction (p=0.069, adjusted p=0.23), and in the DASL cohort it is only borderline after adjustment (adj.p=0.07). The headline hazard ratios (UKE HR=0.41, DASL HR=0.51) are obtained from stratified analyses restricted to signature-positive patients. Such subgroup comparisons are not a sufficient basis for a predictive claim, because the null hypothesis of no interaction is not rejected; the observed effect in one stratum could be due to chance or to confounding. The manuscript should report the interaction test as the pre-specified primary analysis, include the full cross-tabulation of events by treatment and signature, and temper the conclusion to say that the signature is a candidate that requires confirmatory testing.","section":"Table 4; Methods, Statistical analysis"},{"comment":"The UKE cohort is non-randomized, and Table 1 shows imbalances in prognostic factors: the bevacizumab-treated group contains more FIGO stage III/IV tumors, more high-grade tumors, and more patients with residual tumor ≥1cm. Although the Cox models adjust for age, tumor stage, and surgery outcome, these adjustments cannot fully eliminate confounding by indication; unmeasured factors (e.g., performance status, comorbidities, center-specific treatment policies) could explain part or all of the observed survival difference between treatment groups in signature-positive patients. The authors should explicitly acknowledge this limitation and, if possible, perform sensitivity analyses such as E-values or propensity-score weighting to assess the robustness of the stratified HR to unmeasured confounding.","section":"Table 1; Results, The UKE ovarian cancer RNA-Seq cohort"},{"comment":"The DASL array measures only two of the five bicluster 84 genes (SLCO6A1 and PTH2R); the other three members are absent. The DASL 'replication' therefore tests a two-gene proxy, not the full five-gene signature. Stratifying patients by a k-means split on two genes may not recapitulate the biological state defined by the UKE signature. This discrepancy should be stated as a major caveat in the interpretation of the DASL results, and the authors should avoid the phrase 'replicates across datasets and platforms' when the signature content is not identical.","section":"Bicluster 84 signature replicates across datasets and platforms"}],"minor_comments":[{"comment":"The GEO accession number for the UKE dataset is given as 'XXXXXX'; this must be replaced with a valid accession before publication.","section":"Methods, Expression data"},{"comment":"The term 'SRN' appears to be a typo for 'SNR' (signal-to-noise ratio); also, 'rare patients subgroups' should be 'rare patient subgroups'.","section":"Methods, Biclustering for unsupervised patient stratification"},{"comment":"Please standardize the notation for adjusted p-values: both 'adj.p-value' and 'adjusted p-value' are used in the abstract, main text, and tables.","section":"Throughout"},{"comment":"The conclusion that higher UKE–TCGA-OV concordance is 'likely due to both cohorts being profiled using RNA-seq technology' is speculative, since cohort composition and platform are confounded; consider softening this claim.","section":"Impact of platform on reproducibility of expression signatures"},{"comment":"The caption reports that all adjusted p-values exceeded 0.93 for UKE and 0.09 for DASL; it would be helpful to also state the number of genes tested and the minimum unadjusted p-value for context.","section":"Supplementary Figure S1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper would be better received if the authors re-frame it as a hypothesis-generating study. The DASL circularity is a serious issue that could be perceived as undermining the central claim; the authors should be given the opportunity to either reanalyze with a pre-specified UKE-only discovery and DASL validation, or to soften the claims to be consistent with the interaction results. The dataset is a valuable resource, and with appropriate changes the manuscript could be suitable for a bioinformatics-focused journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is an honest attempt to find a bevacizumab-response biomarker in ovarian cancer using a new RNA-seq cohort (UKE, n=181) and a previously published randomized trial cohort (DASL, n=377). The novel piece is real: the UKE cohort is new, the analysis pipeline is transparent, code is on GitHub, and the bicluster 84 five-gene signature was not in the earlier literature. The authors also report their negative results plainly—the interaction p-values in Table 4 are not significant after adjustment (UKE adj.p=0.23, DASL adj.p=0.07), and they frame the signature as a candidate rather than a validated biomarker.\n\nThe soft spot is the validation. The DASL cohort is not truly independent: it was used to filter the 234 UKE biclusters down to 23 that \"present in both cohorts,\" and then the random survival forest feature importances were computed by permuting features in the DASL validation data. That step uses DASL survival outcomes to rank which bicluster matters. Running the Cox interaction test on the same DASL data after selecting bicluster 84 gives an adjusted p-value that doesn't account for the selection. On top of that, DASL only measures two of the five bicluster 84 genes, so the replication is a two-gene proxy. The UKE cohort itself is non-randomized, and the bevacizumab group has more advanced tumors, though the Cox models adjust for the measured confounders.\n\nI'd still send this to peer review rather than desk reject. The new cohort and the candidate signature are worth publishing as a hypothesis-generating study, and the methods are clear enough that a referee can demand a genuinely independent validation (or a re-analysis that treats DASL as exploratory) plus the missing data accession. The biology narrative (CTCFL/stemness) is post hoc, but it's presented as a hypothesis. The abstract oversells the claim by highlighting the subgroup HRs without the interaction caveat; the conclusion is more measured.\n\nThis is a paper for biomarker researchers and computational oncologists who want to see an honest example of the perils of validation set contamination. A serious referee can help fix the framing.","headline":"New cohort and a plausible candidate signature, but the DASL 'validation' is not independent and the interaction tests don't clear the bar; the paper is a useful hypothesis-generation report, not a validated biomarker.","tokens_in":31252,"tokens_out":3459,"would_cite":false,"duration_ms":31874,"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 14-gene tumor expression signature could identify which ovarian cancer patients gain a survival benefit from bevacizumab.","keywords":["ovarian cancer","transcriptome","expression signature","bevacizumab response","bicluster 84","CTCFL","cancer stemness","overall survival"],"falsifier":"A randomized trial with tumor expression profiling would settle it: if patients randomized to bevacizumab who overexpress bicluster 84 do not show a materially lower hazard of death than signature-positive control patients, or if the treatment-by-signature interaction is null, the central claim fails.","tokens_in":30197,"feed_emoji":"🧬","tokens_out":7833,"duration_ms":71724,"temperature":0.7,"pith_summary":"This paper tries to establish that a gene-expression pattern in ovarian tumor tissue can tell which patients will live longer when bevacizumab is added to standard platinum-based chemotherapy. In a new RNA-seq cohort of 181 patients and an earlier microarray cohort of 377 patients, those whose tumors overexpressed the pattern, called the bicluster 84 signature, showed improved overall survival with bevacizumab, with hazard ratios of 0.41 and 0.51, while patients without the signature showed no significant benefit. The authors interpret the signature as a marker of cancer stemness driven by activation of the transcription factor CTCFL, which regulates VEGF-A, the molecule bevacizumab blocks. If true, the signature would give clinicians a test for deciding who should receive bevacizumab and who can be spared its side effects.","feed_headline":"Tumor RNA test flags ovarian cancer patients helped by bevacizumab","feed_subtitle":"In two independent cohorts, patients with the signature lived longer on the drug; patients without it did not.","key_machinery":"The argument runs on a differentially expressed bicluster: a submatrix of the expression data in which a set of genes is concordantly over- or under-expressed in a subset of samples, splitting all tumors into two groups. The bicluster 84 signature was found by unsupervised biclustering in the discovery cohort, replicated in two independent cohorts, prioritized by random-survival-forest importance, and tested in proportional-hazards models adjusted for age, tumor stage, and surgery outcome. Its biological anchor is CTCFL/BORIS, a master transcription factor that maintains cancer stemness, regulates VEGF-A, and appears among the genes whose expression defines the signature.","core_discovery":"The central discovery is a reproducible transcriptomic signature, bicluster 84, that identifies ovarian cancer patients whose overall survival improves when bevacizumab is added to standard chemotherapy. The signature is a set of genes concordantly overexpressed in one tumor subgroup; in patients with that overexpression, bevacizumab treatment was associated with a hazard ratio of 0.41 (95% CI 0.23–0.74) in the discovery RNA-seq cohort and 0.51 (95% CI 0.34–0.75) in the validation microarray cohort, with no significant treatment-related survival difference in signature-negative patients. The signature does not line up with the four established ovarian cancer molecular subtypes, and the authors connect it to CTCFL/BORIS-driven acquisition of cancer stemness, a state linked to angiogenesis and treatment resistance.","pith_inferences":["Editorial extension: a randomized trial with archived tumor RNA would be the decisive test, since the non-randomized comparison in the discovery cohort leaves confounding by indication as an alternative explanation.","Editorial extension: if the signature marks CTCFL-driven stemness, it may predict benefit from other anti-angiogenic agents or from therapies aimed at cancer stem cells, not only bevacizumab.","Editorial extension: because the validation cohort was microarray-based and lacked three of the five bicluster genes originally identified, a dedicated RNA-seq validation could either strengthen the observed effect or reveal platform-specific artifacts.","Editorial extension: combining bicluster 84 with other reproducible signatures, such as the angiogenesis-related one, might yield a more complete predictor than any single signature."],"forward_implications":["In patients whose tumors overexpress bicluster 84, adding bevacizumab to standard chemotherapy is associated with roughly half the risk of death during follow-up in both cohorts (HR 0.41 and 0.51).","Signature-negative patients show no significant overall-survival difference between treatment groups, so the test could identify patients unlikely to benefit and avoid unnecessary bevacizumab exposure.","The signature is independent of the four established ovarian cancer molecular subtypes, so it adds information beyond current tumor classification.","If the CTCFL/stemness interpretation holds, the same biology could guide other anti-angiogenic or stemness-directed therapies, and CTCFL itself becomes a potential therapy target.","Several other reproducible expression biclusters (70, 90, 109, 130) are reported as candidate biomarkers awaiting validation in additional RNA-seq cohorts."],"supporting_citations":[{"why":"Supplies the independent validation cohort with treatment and survival data used to confirm the signature's predictive effect.","marker":"17"},{"why":"Describes the unsupervised biclustering method used to discover differentially expressed biclusters such as bicluster 84.","marker":"25"},{"why":"Provides the consensus molecular-subtype classification used to show that bicluster 84 is not a surrogate for known ovarian cancer subtypes.","marker":"19"},{"why":"Documents that CTCFL/BORIS is frequently expressed or amplified in ovarian and other cancers, anchoring the proposed biology.","marker":"41"},{"why":"Shows that CTCFL regulates VEGF-A expression, connecting the signature to bevacizumab's target pathway.","marker":"42"},{"why":"Establishes that CTCFL/BORIS maintains a treatment-resistant, stem-like phenotype in cancer cells.","marker":"43"},{"why":"Shows that ovarian cancer stem cells can contribute to blood vessel formation and that bevacizumab can inhibit this process.","marker":"53"}],"fun_headline_variants":["RNA signature IDs ovarian cancer patients who live longer on bevacizumab","Bevacizumab benefit in ovarian cancer predicted by tumor RNA signature","Ovarian cancer RNA signature reveals who gains from bevacizumab","Signature finds ovarian cancer patients who get a survival benefit from bevacizumab"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the statistical adjustments for age, tumor stage, and residual tumor fully remove the differences between patients who did and did not receive bevacizumab, so the observed survival benefit in signature-positive patients is caused by the drug rather than by unmeasured clinical factors.","fun_headline_variants_meta":{"raw":{"variants":["RNA signature IDs ovarian cancer patients who live longer on bevacizumab","Bevacizumab benefit in ovarian cancer predicted by tumor RNA signature","Ovarian cancer RNA signature reveals who gains from bevacizumab","Signature finds ovarian cancer patients who get a survival benefit from bevacizumab"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000925,"raw_usage":{"total_tokens":3966,"prompt_tokens":950,"completion_tokens":3016,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":2940}},"tokens_in":566,"tokens_out":3016,"duration_ms":22033,"temperature":1.0,"reasoning_tokens":2940,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:22:23.498396+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A randomized trial with tumor expression profiling would settle it: if patients randomized to bevacizumab who overexpress bicluster 84 do not show a materially lower hazard of death than signature-positive control patients, or if the treatment-by-signature interaction is null, the central claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the independent validation cohort with treatment and survival data used to confirm the signature's predictive effect."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Describes the unsupervised biclustering method used to discover differentially expressed biclusters such as bicluster 84."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the consensus molecular-subtype classification used to show that bicluster 84 is not a surrogate for known ovarian cancer subtypes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents that CTCFL/BORIS is frequently expressed or amplified in ovarian and other cancers, anchoring the proposed biology."},{"cited_title":"& Soto-Reyes, E","cited_arxiv_id":null,"evidence_quote":"Shows that CTCFL regulates VEGF-A expression, connecting the signature to bevacizumab's target pathway."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that CTCFL/BORIS maintains a treatment-resistant, stem-like phenotype in cancer cells."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Shows that ovarian cancer stem cells can contribute to blood vessel formation and that bevacizumab can inhibit this process."}],"review_version":1}