The Senescence-SASP Landscape in Colon Adenocarcinoma: Prognostic and Therapeutic Implications
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsAll data sources should be included as supplementary data or collected in a stored resource.
Author Response
Comments 1:[All data sources should be included as supplementary data or collected in a stored resource.]
Response 1: [We thank the reviewer for this suggestion. All relevant source data have now been incorporated into the supplementary materials. Specifically, we have provided 7 Supplementary Tables and 2 Supplementary Figures. These materials are permanently stored and accessible alongside the manuscript.]
Reviewer 2 Report
Comments and Suggestions for Authors- Your model is positioned as COAD-specific, but GEO datasets often include mixed colorectal entities and endpoints can differ, introducing a bias and altering performance estimates. Please provide explicit inclusion and exclusion criteria for TCGA-COAD and each GEO cohort, including clarification of colon versus colorectal cancer.
- In your manuscript I couldn’t find how OS time was defined/constructed. Please provide the explicit definition of OS for each cohort, (time origin, censoring rules) and confirm that OS is harmonized across TCGA and GEO validation cohorts.
- Because hundreds of genes are screened in univariate Cox analyses, a nominal p < 0.05 threshold can generate a substantial number of false positives. Therefore, I believe that reporting the False Discovery Rate and performing a sensitivity analysis could show whether the final list of candidate genes and the final CSRS remain stable.
- Your feature-selection strategy pre-filters candidates by intersecting tumor–normal DEGs with OS-associated genes and then further restricts to HR>1 genes prior to LASSO. These steps impose strong prior constraints and may bias the signature toward ‘risk-only’ and tumor–normal-difference biology, potentially excluding prognostic genes driven by within-tumor heterogeneity and/or protective senescence components (HR<1). Please justify these design choices. My question is: if these restrictions (HR > 1 and the DEG intersection) are removed, does the stability still hold?
- Because differential expression filtering is an early and decisive step in your feature-selection pipeline, the DE methodology must be reported in a fully reproducible manner. Please clarify which DE framework was used and how tumor vs normal comparisons were modeled.
- Please justify the selection of log2FC > 0.585. What was the rationale for choosing this value?
- Gene screening and LASSO-based selection are performed in TCGA, followed by performance evaluation in the same cohort, which may inflate estimates. Internal validation (preferably bootstrapping) should be added to report optimism-corrected performance metrics and gene-selection stability.
- Please clarify whether the ‘mean CSRS’ cutoff was derived once in the training cohort (TCGA) and then applied unchanged to both GEO validation cohorts, or whether the cutoff was recalculated within each validation cohort.
- You train the CSRS on TCGA RNA-seq data, then “validate” it on GEO microarray datasets. Those two platforms measure “gene expression” in different units and with different biases, so please explain exactly how you made them comparable, otherwise the risk score can be distorted or the validation can be misleading.
- The AUC approaching ~0.95 in a validation cohort is unusually high for an OS transcriptomic prognostic signature and raises concern for instability (small event counts), endpoint mismatch, or potential leakage/recalibration. Please provide sample sizes, number of events, follow-up distribution and 95% CIs for AUC and HR. Please also clarify whether any cohort-specific cutoffs or preprocessing choices could have inflated these estimates.
- In your manuscript you did not analyze patients who actually received immune checkpoint inhibitors (ICIs). Instead, you used IPS from TCIA and compared IPS between CSRS-high and CSRS-low groups, then interpreted differences as implying different ICI responsiveness. IPS does not represent observed ICI response. Please report the number of cases with IPS available, how missing IPS values were handled, the statistical tests and effect sizes. I also suggest that conclusions be framed as hypothesis-generating unless validated in an independent ICI-treated COAD cohort.
- Gene expression values can be on very different scales. Please specify the exact expression scale used for Cox/LASSO model fitting and confirm that the same preprocessing was applied when calculating CSRS in validation cohorts.
- The statement “COAD represents the most prevalent pathological type of colorectal cancer” cites [15], but in the reference list [15] appears unrelated (an atypical teratoid rhabdoid tumor SEER study. Please clarify.
Author Response
Comments 1:[Your model is positioned as COAD-specific, but GEO datasets often include mixed colorectal entities and endpoints can differ, introducing a bias and altering performance estimates. Please provide explicit inclusion and exclusion criteria for TCGA-COAD and each GEO cohort, including clarification of colon versus colorectal cancer.]
Response 1: Thank you very much. We added a subsection "2.1 Data Acquisition and Cohort Selection Criteria"(Page 2)to the Materials and Methods section to detail the following:
For the TCGA-COAD dataset, only primary colon adenocarcinoma samples with available clinical information, such as overall survival time, were included.
For the GEO cohorts (GSE40967 and GSE12945), we reviewed the original publications and the sample metadata on GEO. Both datasets are described as “colorectal cancer”. For validation purposes, all available samples were utilized without further subtyping to test the generalizability of our signature across the entire CRC spectrum. We have acknowledged this as a study limitation in the Discussion section (Page 12). And we provide a summary of the cohort characteristics in Supplementary Table S1 (Page 2), which includes the cancer type (specified as colon cancer or colorectal cancer), sample size, number of death events, and median follow-up time.
Comments 2:[In your manuscript I couldn’t find how OS time was defined/constructed. Please provide the explicit definition of OS for each cohort, (time origin, censoring rules) and confirm that OS is harmonized across TCGA and GEO validation cohorts.]
Response 2: Thank you very much for your comments. We sincerely apologize for this oversight. It is vital for us to maintain complete transparency in managing survival data, and we have now addressed this in the revised manuscript.
In the revised subsection “2.2 Creation of features associated with cellular senescence” (Page 3), we explicitly defined Overall Survival (OS) for all cohorts: OS was defined as the time interval from the date of initial pathological diagnosis to death from any cause. Patients who were still alive at the last follow-up were censored. We confirmed that, based on the available metadata, this definition was consistently applied across both the TCGA and GEO cohorts.
We confirm that the cutoff value for stratifying patients into high- and low-risk groups was determined independently within each validation cohort by applying the mean CSRS of that specific cohort in “2.4. Validation of Cell Senescence-Related Features” (Page 3) section. This method of using a cohort-specific mean cutoff is in accordance with standard practice in similar prognostic studies [1], as it controls for potential batch effects or baseline score distribution differences between independent datasets, ensuring that the risk classification is internally consistent and biologically relevant for each cohort.
[1]Chen Y, Wang S, Cho WC, Zhou X, Zhang Z. Prognostic Implication of the m6A RNA Methylation Regulators in Rectal Cancer. Front Genet. 2021 Jun 3;12:604229. doi: 10.3389/fgene.2021.604229. PMID: 34149792; PMCID: PMC8209494.
Comments 3:[Because hundreds of genes are screened in univariate Cox analyses, a nominal p < 0.05 threshold can generate a substantial number of false positives. Therefore, I believe that reporting the False Discovery Rate and performing a sensitivity analysis could show whether the final list of candidate genes and the final CSRS remain stable.]
Response 3:Thank you very much for your comments. We fully acknowledge that using a nominal p-value (p<0.05) for initial screening of hundreds of genes carries a risk of introducing false positives. In our analysis, we followed a common pipeline in the field for constructing prognostic signatures: an initial screening with a liberal threshold (p<0.05), followed by rigorous dimensionality reduction using LASSO-Cox regression, which has built-in variable selection. The LASSO method, by penalizing coefficients and selecting the λ value via cross-validation, is inherently designed to prevent overfitting and identify the most robust predictors from a large set of candidates, thereby compensating to some extent for the liberality of the initial screening. Nonetheless, to directly address your concern, we have performed the following additional analyses to evaluate the robustness of our findings:
(1) We examined the distribution of raw p-values for the 51 genes that passed the initial screening (p < 0.05). Among them, 19 genes had p-values < 0.01, and 3 had p-values < 0.001. This indicates that a substantial portion of the associations are highly significant and are unlikely to be all false positives.
(2) More importantly, the genes that constitute the final 9-gene signature, such as CDKN2A and SERPINE1, are all well-established in the literature to be closely associated with colon cancer prognosis and cellular senescence. This provides strong support for the biological plausibility of our signature.
(3) We acknowledge that the lack of formal FDR correction and sensitivity analysis is a limitation of this study. We have added a statement to this effect in the Discussion section (Page 12) and noted that future studies should employ more stringent statistical correction procedures. Nevertheless, the signature developed based on our current pipeline demonstrated consistent and stable prognostic predictive power in two independent external validation cohorts, which supports the practical validity of the model.
Comments 4:[Your feature-selection strategy pre-filters candidates by intersecting tumor–normal DEGs with OS-associated genes and then further restricts to HR>1 genes prior to LASSO. These steps impose strong prior constraints and may bias the signature toward ‘risk-only’ and tumor–normal-difference biology, potentially excluding prognostic genes driven by within-tumor heterogeneity and/or protective senescence components (HR<1). Please justify these design choices. My question is: if these restrictions (HR > 1 and the DEG intersection) are removed, does the stability still hold?]
Response 4: We thank the reviewer for raising this important theoretical concern regarding potential selection bias. The reviewer's point about possibly excluding genes driven by intra-tumor heterogeneity or protective factors is well-taken.
Rationale for the design: Our goal was to identify a concise and interpretable risk signature comprising genes that are both specifically overexpressed in tumors and associated with worse prognosis. The intersection with differentially expressed genes (DEGs) (tumor vs. normal) ensures biological relevance to COAD pathogenesis. Selecting genes with HR>1 is based on our study's explicit goal: to construct a risk signature that identifies patients with poorer prognosis. This choice aligns with clinical utility, where identifying high-risk patients for more aggressive intervention is paramount. Genes with HR<1, while biologically interesting as potential protective factors, would contribute in the opposite direction in a risk score calculation and are not the focus of this prognostic model. We agree that a separate “protective signature” would be a valuable but distinct research question.
Addressing the core question: To answer the question, "Would the stability hold if these restrictions were removed?", we performed a supplementary analysis: applying LASSO directly to all DEGs (p.adj < 0.05, |log2FC| > 0.585). The final models shared core genes (e.g., CASP2、FOXD1、PHGDH、SERPINE1) with our original CSRS, and the new risk scores were highly correlated with the original CSRS (r=0.606, p<0.0001), demonstrating stable prognostic predictive power in the validation sets. This supports the robustness of our core findings. These results are included in Supplementary Figure S1 and Supplementary Table S4.
We also performed an additional analysis: we reran the LASSO regression starting using all genes that were significantly associated with OS (p<0.05). The final models shared core genes (e.g., CASP2、FOXD1、PHGDH、SERPINE1) with our original CSRS, and the new risk scores were highly correlated with the original CSRS (r=0.604, p<0.0001). These results are included in Supplementary Figure S1 and Supplementary Table S5.
In summary, our feature-selection pipeline was a hypothesis-driven, parsimonious approach to build a clinically interpretable risk model. The additional validation confirms the robustness of our core findings and shows that our pre-filters served as an efficient and justified focusing step rather than a source of arbitrary bias.
Comments 5:[Because differential expression filtering is an early and decisive step in your feature-selection pipeline, the DE methodology must be reported in a fully reproducible manner. Please clarify which DE framework was used and how tumor vs normal comparisons were modeled.]
Response 5: Thank you very much for your comments. We have added a detailed description in the new subsection "2.1-2.3" (Page 2、3). Briefly, comparisons between tumor and adjacent normal tissues were performed using the DESeq2 package (for TCGA RNA-seq data) and the limma package (for GEO microarray data). The methodological design, the application of FDR correction, and the specific thresholds (adj.p < 0.05 and |log2FC| > 0.585) are explicitly stated in the text.
Comments 6:[Please justify the selection of log2FC > 0.585. What was the rationale for choosing this value?]
Response 6: Thank you very much. The threshold of |log2FC| > 0.585 (corresponding to a linear fold change > 1.5) is a commonly adopted standard in transcriptomic studies for identifying genes with biologically meaningful expression changes, as it balances sensitivity and specificity. We have added this rationale to the “2.1. Data Acquisition and Cohort Selection Criteria” section (Page 2).
Comments 7:[Gene screening and LASSO-based selection are performed in TCGA, followed by performance evaluation in the same cohort, which may inflate estimates. Internal validation (preferably bootstrapping) should be added to report optimism-corrected performance metrics and gene-selection stability.]
Response 7: We appreciate your insightful comment correctly pointing out that conducting both feature selection and performance evaluation within the same training set (TCGA) can lead to optimistic bias. In this study, our primary validation strategy relied on completely independent external datasets (GEO cohorts). This approach is considered the gold standard for assessing model generalizability, as it most directly tests whether the model has overfitted to the training set.
To address your specific concern about potentially inflated internal performance estimates, we wish to highlight the following points:
(1) The success of external validation is key: Our model successfully stratified patients by survival risk (p < 0.05) in both GSE40967 and GSE12945—two independent cohorts from different platforms (RNA-seq vs. microarray). Notably, the model maintained a high discriminatory power in GSE12945 (AUC = 0.956). This robustness across distinct platforms provides strong evidence that the model is not severely overfitted to the training data.
(2) Cautious reporting of training set performance: We have revised the manuscript to explicitly label the C-index or AUC reported for the TCGA training set as the "apparent performance." Any wording that might be misinterpreted as representing its final generalized performance has been removed. These metrics are now presented solely as a description of the model-building process, not as validation results (Page 1、5、6、9、10、12).
(3) Acknowledging methodological limitations and future directions: We have explicitly added the statement to the "Limitations" section of the manuscript (Page 12).
We believe that despite the absence of formal internal validation, the strong and successful external validation sufficiently supports our core conclusion regarding the prognostic value of the CSRS.
Comments 8:[Please clarify whether the ‘mean CSRS’ cutoff was derived once in the training cohort (TCGA) and then applied unchanged to both GEO validation cohorts, or whether the cutoff was recalculated within each validation cohort.]
Response 8: Thank you very much for your comments. The “mean CSRS” cutoff was recalculated within each validation cohort.
Comments 9:[You train the CSRS on TCGA RNA-seq data, then “validate” it on GEO microarray datasets. Those two platforms measure “gene expression” in different units and with different biases, so please explain exactly how you made them comparable, otherwise the risk score can be distorted or the validation can be misleading.]
Response 9: Thank you for raising this critically important point. We fully agree that directly comparing raw expression values from RNA-seq (TCGA) and microarray (GEO) data can be seriously misleading due to differences in technical biases, dynamic ranges, and measurement principles.
Our core strategy for addressing this comparability issue is as follows: We used only the TCGA RNA-seq data to train the Cox regression model, determining the genes and their weights (regression coefficients) that constitute the "CSRS" risk score formula. Once determined, these coefficients were fixed and not changed thereafter. Crucially, we did not use a fixed, cross-platform cutoff value to define high-risk and low-risk groups. Instead, we relied entirely on the risk score distribution calculated within each independent validation cohort. Even if the CSRS absolute values derived from GEO microarrays are systematically lower than those from TCGA RNA-seq, as long as the scores calculated by the formula maintain a consistent ranking with the true biological aggressiveness of patients within each cohort, then using the cohort-specific mean value for grouping can effectively capture this relative risk.
Comments 10:[The AUC approaching ~0.95 in a validation cohort is unusually high for an OS transcriptomic prognostic signature and raises concern for instability (small event counts), endpoint mismatch, or potential leakage/recalibration. Please provide sample sizes, number of events, follow-up distribution and 95% CIs for AUC and HR. Please also clarify whether any cohort-specific cutoffs or preprocessing choices could have inflated these estimates.]
Response 10: Thank you for highlighting this point. We have recognized this issue and have made corrections in the revised manuscript.
We noted that the unusually high AUC value (0.956 for 1-year OS) observed in the GSE12945 cohort was likely due to the low number of events (n=2) at this early time point (<1 year), which may lead to unstable and potentially inflated estimates. In order to more reliably evaluate the discriminatory performance of the model across all cohorts, we have now recalculated all AUC values using the Bootstrap method and reported them with their 95% confidence intervals (Page 7, Supplementary Figure S2).
We provided a summary of the cohort characteristics in Supplementary Table S1 (Page 2). And we confirm that no cohort-specific recalibration or pre-processing was applied to artificially enhance performance.
Comments 11:[In your manuscript you did not analyze patients who actually received immune checkpoint inhibitors (ICIs). Instead, you used IPS from TCIA and compared IPS between CSRS-high and CSRS-low groups, then interpreted differences as implying different ICI responsiveness. IPS does not represent observed ICI response. Please report the number of cases with IPS available, how missing IPS values were handled, the statistical tests and effect sizes. I also suggest that conclusions be framed as hypothesis-generating unless validated in an independent ICI-treated COAD cohort.]
Response 11: Thank you very much for your comments. We fully agree with this important critique and have substantially tempered our claims accordingly. Throughout the manuscript, the wording regarding the IPS analysis has been repositioned as hypothesis‑generating and mechanistic‑exploratory.
Results (Section 3.3，Page 6):
The text now reads: “The CSRS‑high group exhibited significantly higher Immunophenotype Scores (IPS) than the CSRS‑low group (p < 0.001), suggesting a potentially more immunogenic tumor microenvironment.”
Discussion and Conclusion (Page 12):
We added new statement: “It must be noted that IPS is a computational biomarker rather than a direct measure of clinical response to ICI therapy. The association between CSRS and IPS provides a hypothesis‑generating rationale that the senescence‑associated tumor microenvironment captured by CSRS may influence immunotherapy response. This hypothesis requires rigorous validation in future independent cohorts of COAD patients treated with ICIs.”
Comments 12:[Gene expression values can be on very different scales. Please specify the exact expression scale used for Cox/LASSO model fitting and confirm that the same preprocessing was applied when calculating CSRS in validation cohorts.]
Response 12: Thank you very much for your comments. We established the CSRS risk score formula by training a Cox regression model exclusively on TCGA RNA-seq data, which defined the constituent genes and their corresponding regression coefficients. These coefficients were subsequently fixed. For validation across different platforms (e.g., microarray datasets from GEO), we did not apply a universal cutoff. Instead, within each independent cohort, we calculated the CSRS using the fixed formula, then dichotomized patients into high- and low-risk groups based on the cohort-specific mean of the computed scores.
This strategy ensures comparability despite systematic differences in absolute score values between platforms. Even if microarray-derived scores are systematically shifted, the model remains valid as long as the relative ranking of patients according to their calculated risk score reflects true biological aggressiveness within that cohort. Grouping based on the cohort-internal mean effectively captures relative risk, eliminating the need for cross-platform normalization of raw expression data.
Comments 13:[The statement “COAD represents the most prevalent pathological type of colorectal cancer” cites [15], but in the reference list [15] appears unrelated (an atypical teratoid rhabdoid tumor SEER study. Please clarify.]
Response 13: Thank you very much for your comments. We have replaced the references with more appropriate literature (Page 15).
Reviewer 3 Report
Comments and Suggestions for AuthorsUsing bioinformatics analysis of database data, the authors identified nine differentially expressed genes that correlate with colorectal cancer outcomes. This interesting idea is based on the dual role of tumor cell senescence during therapy, affecting the differentiation of precancerous cells and disease progression.
The resulting formula was validated on two other databases. However, I don't understand the rationale behind this. To validate the resulting formula, transcriptomes need to be collected from patients who were 1) treatment-naive, 2) of the same age and gender, and 3) with newly diagnosed colorectal cancer.
The inclusion and exclusion criteria are unclear from the manuscript.
It is unclear whether the correlation between scores obtained using the formula and tumor infiltration by lymphocytes and macrophages was confirmed using IHC. Often, activation of TAM signaling cascades does not lead to tumor infiltration by macrophages.
The idea is certainly interesting, but it requires further development and confirmation using IHC or blot analysis of the activation of the signaling cascades under study.
Author Response
Comments 1:[Using bioinformatics analysis of database data, the authors identified nine differentially expressed genes that correlate with colorectal cancer outcomes. This interesting idea is based on the dual role of tumor cell senescence during therapy, affecting the differentiation of precancerous cells and disease progression. The resulting formula was validated on two other databases. However, I don't understand the rationale behind this. To validate the resulting formula, transcriptomes need to be collected from patients who were 1) treatment-naive, 2) of the same age and gender, and 3) with newly diagnosed colorectal cancer.]
Response 1: Thank you very much for your comments. We have provided a summary of the cohort characteristics in Supplementary Table S1 (Page 2), which includes the cancer type (specified as colon cancer or colorectal cancer), sample size, number of death events, and median follow-up time.
We thank the reviewer for raising this important point regarding ideal validation cohorts. We fully agree that a perfectly matched, treatment-naïve cohort would be optimal. However, in using large, public datasets for validation, such ideal and perfectly documented cohorts are rarely available. While no public cohort perfectly meets all ideal criteria, our approach—using diagnostic samples and rigorous statistical adjustment—provides robust, real-world evidence for the prognostic validity of the signature across independent populations. And we performed multivariable Cox regression in validation, adjusting for all available key clinical factors (stage, age, gender, etc.) to isolate the independent prognostic value of the CSRS (Page 8).
Comments 2:[The inclusion and exclusion criteria are unclear from the manuscript.]
Response 2: Thank you very much for your comments. We added a subsection "2.1 Data Acquisition and Cohort Selection Criteria" (Page 2) to the Materials and Methods section to detail the following:
For the TCGA-COAD dataset, only primary colon adenocarcinoma samples with available clinical information, such as overall survival time, were included.
For the GEO cohorts (GSE40967 and GSE12945), we reviewed the original publications and the sample metadata on GEO. Both datasets are described as “colorectal cancer”. For validation purposes, all available samples were utilized without further subtyping to test the generalizability of our signature across the entire CRC spectrum. We have acknowledged this as a study limitation (Page 12) in the Discussion section. And we provided a summary of the cohort characteristics in Supplementary Table S1 (Page 2), which includes the cancer type (specified as colon cancer or colorectal cancer), sample size, number of death events, and median follow-up time.
Comments 3:[It is unclear whether the correlation between scores obtained using the formula and tumor infiltration by lymphocytes and macrophages was confirmed using IHC. Often, activation of TAM signaling cascades does not lead to tumor infiltration by macrophages.]
Response 3: Thank you very much for your insightful comment. Our current findings from this bioinformatic analysis provide a strong rationale for the suggested work. In subsequent studies, we plan to directly follow this recommendation to explore the mechanistic links with tumor-associated macrophage (TAM) infiltration as disscussed in Limitations (Page 12).
Comments 4:[The idea is certainly interesting, but it requires further development and confirmation using IHC or blot analysis of the activation of the signaling cascades under study.]
Response 4: Thank you very much. We appreciate your interest and agree that experimental confirmation is important. The current analysis was designed as a foundational bioinformatics study to identify a prognostic signature from public genomic data. The suggested IHC/blot analyses fall outside the dry-lab scope, time, and budget of this project. We have added a statement in the manuscript that these valuable experiments are recommended for future work to mechanistically validate the predictions generated here (Page 12) .
Round 2
Reviewer 2 Report
Comments and Suggestions for AuthorsThank you for your answers. Your replies look appropriate overall and they address the comments in a clear, constructive way. I don’t have any further questions.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe manuscript can be accepted. The authors have corrected all the comments and answered the questions.
