CBX4 and NOX4 Expression Are Associated With Chemoradiotherapy Response and Overall Survival in Locally Advanced Cervical Squamous Cell Carcinoma

ABSTRACT

Background

Tissue biomarkers may improve risk stratification in locally advanced cervical cancer before definitive concurrent chemoradiotherapy (CCRT). We evaluated pretreatment expression of chromobox 4 (CBX4), NADPH oxidase 4 (NOX4), specificity protein 1 (SP1), and copper transporter 1 (CTR1) in relation to CCRT response and overall survival.

Methods

In this retrospective 1:1 matched case–control study, pretreatment biopsies from 40 patients with locally advanced cervical squamous cell carcinoma (20 CCRT‐resistant, 20 CCRT‐sensitive) were quantified by AI‐assisted digital pathology using H‐scores. A parsimonious CBX4 + NOX4 logistic model was evaluated by apparent ROC, leave‐one‐out cross‐validation (LOOCV), repeated stratified 5‐fold cross‐validation, and bootstrap optimism correction. Overall survival analyses were exploratory.

Results

CBX4 and NOX4 expression was lower in CCRT‐resistant tumors (median H‐score: CBX4, 40.47 vs. 79.62; NOX4, 7.00 vs. 53.93; both p < 0.001), whereas SP1 and CTR1 did not differ. Apparent AUCs were 0.840 for CBX4 and 0.843 for NOX4. The two‐marker model had an apparent AUC of 0.903, LOOCV AUC 0.845, repeated 5‐fold CV AUC 0.854 ± 0.022, and bootstrap‐corrected AUC 0.883. High CBX4 and NOX4 expression was associated with better 5‐year overall survival; in joint Cox models, hazard ratios per 1‐SD increase were 0.31 (95% CI 0.13–0.74) for CBX4 and 0.48 (0.26–0.90) for NOX4. No histochemically detectable copper deposits were observed.

Conclusions

Lower pretreatment CBX4 and NOX4 expression was associated with a CCRT‐resistant phenotype and poorer overall survival in this small, response‐enriched cohort. Internal resampling suggested discrimination, but same‐dataset marker selection may introduce residual optimism and does not establish generalizability. Independent validation in a consecutively recruited cohort is required before clinical application. No mechanisms involving DNA repair, ROS, or chemoradiotherapy sensitivity were established.

Abbreviations

AUC

area under the receiver operating characteristic curve

CBX4

chromobox 4

CCRT

concurrent chemoradiotherapy

CCRT‐R

CCRT‐resistant

CCRT‐S

CCRT‐sensitive

CI

confidence interval

CTR1

copper transporter 1

FIGO

International Federation of Gynecology and Obstetrics

H₂O₂

hydrogen peroxide

HR

hazard ratio

IHC

immunohistochemistry

IQR

interquartile range

LACC

locally advanced cervical cancer

LOOCV

leave‐one‐out cross‐validation

NOX4

NADPH oxidase 4

OS

overall survival

ROC

receiver operating characteristic

ROS

reactive oxygen species

SP1

specificity protein 1

SUMO

small ubiquitin‐like modifier

1. Introduction

Cervical cancer remains a major global health problem, with the greatest burden in low‐ and middle‐income settings [1]. For locally advanced cervical cancer (LACC), definitive concurrent chemoradiotherapy (CCRT) is a standard curative‐intent treatment [2, 3, 4, 5]. Nevertheless, treatment failure and subsequent recurrence remain clinically important, and conventional staging alone does not fully capture biological heterogeneity in treatment response. Tissue biomarkers measurable before treatment may therefore complement clinicopathological risk assessment.

Chromobox 4 (CBX4) is a polycomb‐group protein with SUMO E3 ligase activity that participates in chromatin regulation and DNA‐damage responses [6, 7, 8, 9, 10]. Its relationship with cancer treatment response appears context dependent. In esophageal squamous cell carcinoma, experimental CBX4 depletion increased radiosensitivity and altered autophagy‐related responses [7], whereas studies of the broader CBX family in pancreatic adenocarcinoma and glioblastoma have described heterogeneous expression patterns, genomic alterations, immune‐infiltration associations, and prognostic or recurrence‐related findings across different CBX members [11, 12]. These family‐level observations do not establish a corresponding prognostic or treatment‐response role for CBX4 specifically. These observations support evaluating CBX4 directly within cervical cancer rather than extrapolating a fixed functional direction across tumor types.

NOX4 is a constitutively active NADPH oxidase and an important source of intracellular reactive oxygen species (ROS). More broadly, NOX‐derived ROS can participate in redox signaling or contribute to cytotoxic oxidative stress, depending on biological context and ROS burden [13, 14, 15]. Its biological effects are therefore highly context dependent. Cervical cancer models have linked AQP3/NOX4‐derived hydrogen peroxide to invasion‐related signaling [16], whereas NOX4 has been reported to support tumorigenesis in renal cancer [17] but to constrain progression in hepatocellular and selected breast cancer models [18, 19]. Conversely, clinical and pan‐cancer analyses in pancreatic cancer and glioblastoma have associated NOX4‐related signatures or high NOX4 expression with adverse biological features in some settings [20, 21]. Thus, an observed clinical association between NOX4 expression and CCRT response should not be assumed to establish a specific ROS‐mediated mechanism.

SP1 regulates transcriptional programs relevant to proliferation and treatment resistance [22, 23], whereas CTR1 (SLC31A1) is a major copper transporter that also participates in cisplatin uptake [24, 25, 26]. Interest in copper‐dependent cell death has further raised the possibility that copper homeostasis may influence therapeutic response [27, 28]. Separately, a review has discussed autophagy as a potential therapeutic research area in cervical cancer [29], and experiments with ONC201 have demonstrated an integrated stress response and cell death in cervical cancer cell lines [30]. Neither publication directly tests CBX4 or NOX4 as determinants of chemoradiotherapy response; these reports are included only as broader cervical cancer research context.

In this tissue‐based exploratory study, we quantified CBX4, NOX4, SP1, and CTR1 in pretreatment cervical biopsies using AI‐assisted digital pathology. We evaluated their association with a predefined CCRT‐resistant versus CCRT‐sensitive phenotype, assessed a parsimonious CBX4 + NOX4 discrimination model with internal validation, examined exploratory overall survival associations, and documented histochemically detectable copper deposition. We specifically distinguish clinical association from causal mechanism throughout the revised analysis.

2. Materials and Methods

2.1. Study Design and Patient Cohort

This retrospective matched case–control study was approved by the Institutional Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (Approval No. 2024‐K0522; approval date January 3, 2025) and was conducted in accordance with the Declaration of Helsinki. Forty patients treated between January 2015 and October 2023 were included. Eligibility criteria were: (1) histologically confirmed primary cervical squamous cell carcinoma; (2) FIGO 2018 stage IB3‐IVA; (3) no prior antitumor treatment; (4) available pretreatment cervical biopsy tissue; (5) completion of definitive CCRT; and (6) adequate clinical and follow‐up information. Patients with another malignant tumor, previous antitumor therapy, severe comorbidity compromising treatment compliance, or loss to follow‐up were excluded.

Patients with local recurrence or distant metastasis within 3 years after CCRT were classified as CCRT‐resistant (CCRT‐R; n = 20). Patients without recurrence or metastasis within 3 years and with at least 3 years of disease‐free follow‐up were classified as CCRT‐sensitive (CCRT‐S; n = 20). A 1:1 matched sampling strategy targeted similarity in age (±3 years), FIGO stage group, and tumor size. Because the final sample was small, residual imbalance in individual baseline variables remained possible and is reported descriptively rather than interpreted as proof of complete balance.

2.2. Immunohistochemistry and AI‐Assisted Digital Pathology

Formalin‐fixed, paraffin‐embedded cervical biopsy specimens were cut into 4‐μm sections. After deparaffinization and rehydration, heat‐induced antigen retrieval was performed in Tris‐EDTA buffer (pH 8.0) for 30 min. Endogenous peroxidase activity was blocked with 3% hydrogen peroxide, followed by 3% bovine serum albumin for 30 min. Sections were incubated overnight at 4°C with primary antibodies against CBX4 (1:400; Proteintech, 18,544–1‐AP), NOX4 (1:1000; Servicebio, GB11347), SP1 (1:1000; Servicebio, GB114530), or CTR1 (1:1000; Servicebio, GB11587). After washing, slides were incubated for 50 min with HRP‐labeled goat anti‐rabbit IgG (1:200; Servicebio, GB23303), developed using 3,3′‐diaminobenzidine, and counterstained with hematoxylin.

Stained slides were digitized and analyzed using the Aipathwell AI‐assisted digital pathology platform (Servicebio, Wuhan, China). Tumor regions were identified on whole‐slide images and staining intensity was categorized as weak, moderate, or strong. The H‐score was calculated as: (percentage weak × 1) + (percentage moderate × 2) + (percentage strong × 3), yielding a range from 0 to 300 [31]. The platform was used for quantitative image analysis rather than as an autonomous clinical diagnostic system.

2.3. Copper Staining

Copper deposition was assessed using erythrine copper staining. Deparaffinized sections were incubated in copper erythramine staining solution at 37°C for approximately 3 days, washed in 70% ethanol, dehydrated, counterstained with nuclear red, cleared, and mounted. Dark green‐black deposits were regarded as histochemically detectable copper salt. All stained specimens were independently reviewed by two pathologists blinded to clinical outcomes. The assay identifies visible histochemical deposits; a negative result does not exclude altered copper flux, intracellular copper below the detection threshold, or other perturbations of copper metabolism.

2.4. Clinical Data and Follow‐Up

Clinical variables included age, FIGO 2018 stage, tumor diameter, histologic grade, and lymph node status. Follow‐up was scheduled every 3 months during the first 2 years, every 6 months during years 3–5, and annually thereafter. Recurrence or metastasis was documented by imaging (CT, MRI, or PET‐CT) and/or pathological confirmation when available. Overall survival (OS) was defined as the interval from initiation of definitive CCRT to death from any cause or last follow‐up. Because recurrence time was not consistently available as an analyzable time‐to‐event variable in the final dataset, recurrence‐free survival was not modeled as a separate survival endpoint in this revision.

2.5. Statistical Analysis

The revision‐specific statistical reanalysis was performed in Python 3.13.5 using SciPy 1.17.0, scikit‐learn 1.8.0, NumPy 2.3.5, and statsmodels. Continuous variables are reported as median (interquartile range [IQR]) and were compared with the Mann–Whitney U test. Categorical variables were compared with Fisher exact tests because of small cell counts. Pearson correlation coefficients were used for descriptive biomarker correlations and for correlation with binary CCRT sensitivity status (CCRT‐S = 1, CCRT‐R = 0).

Discrimination for the CCRT‐resistant phenotype was summarized by ROC AUC. To reduce model complexity, the original four‐marker combination was replaced by a parsimonious two‐marker logistic model containing CBX4 and NOX4, the two markers that showed clear individual discrimination. Internal validation included LOOCV, repeated stratified 5‐fold cross‐validation (100 repeats), and 2000 bootstrap resamples for optimism correction. The fitted study‐cohort linear predictor was LP = 4.975–0.05095 × CBX4–0.04641 × NOX4, with higher values indicating a greater probability of belonging to the CCRT‐R group within this response‐enriched sample. Because CCRT‐R and CCRT‐S patients were deliberately sampled 1:1, the intercept and predicted probabilities do not estimate population‐level absolute risk; therefore, we emphasize discrimination and do not interpret population calibration or clinical net benefit from this case–control sample. The marker selection was data‐informed within the same 40‐patient cohort; the reported resampling procedures did not fully repeat the initial candidate‐marker selection and therefore may not capture its entire associated optimism.

The OS analyses were prespecified here as exploratory because patient sampling was based on CCRT response and the cohort was small. Kaplan–Meier curves were dichotomized at the cohort median H‐score for visualization and compared by the log‐rank test. To limit overfitting, the joint Cox proportional hazards model included only standardized continuous CBX4 and NOX4 (20 deaths and 2 predictors); hazard ratios (HRs) therefore represent a 1‐standard‐deviation increase in marker expression. Two‐sided p < 0.05 was considered statistically significant. No multiplicity‐adjusted confirmatory inference was undertaken.

3. Results

3.1. Baseline Characteristics

The cohort comprised 40 patients (20 CCRT‐S and 20 CCRT‐R), with a median age of 53.0 years (IQR 46.0–59.0). None of the tested baseline comparisons reached statistical significance; however, the CCRT‐R group contained a larger proportion of Stage III–IV disease (70% vs. 40%), illustrating residual imbalance that can occur in a small matched sample. The median measured tumor diameter was similar between groups (4.2 vs. 4.0 cm; p = 0.747). Lymph node metastasis was present in four patients overall (two in each group). Baseline characteristics are summarized in Table 1.

TABLE 1.

Baseline clinicopathological characteristics of the study cohort.

Characteristic	Total (N = 40)	CCRT‐S (n = 20)	CCRT‐R (n = 20)	p
Age, years, median (IQR)	53.0 (46.0–59.0)	56.0 (48.3–59.3)	50.0 (45.5–55.0)	0.203
FIGO stage, n (%)
II	18 (45.0)	12 (60.0)	6 (30.0)	0.111
III–IV	22 (55.0)	8 (40.0)	14 (70.0)
Tumor diameter, cm, median (IQR)	4.0 (3.0–5.0)	4.0 (3.8–5.0)	4.2 (3.8–5.0)	0.747
Histologic grade, n (%)
G2 (moderately differentiated)	18 (45.0)	8 (40.0)	10 (50.0)	0.751
G3 (poorly differentiated)	22 (55.0)	12 (60.0)	10 (50.0)
Lymph node metastasis, n (%)
Negative	36 (90.0)	18 (90.0)	18 (90.0)	> 0.999
Positive	4 (10.0)	2 (10.0)	2 (10.0)

Note: p values: Mann–Whitney U test for continuous variables and Fisher exact test for categorical variables. CCRT‐S, CCRT‐sensitive; CCRT‐R, CCRT‐resistant; IQR, interquartile range. Stage was grouped for descriptive analysis because of the small numbers within individual substages.

3.2. Biomarker Expression and CCRT Response

CBX4 and NOX4 H‐scores were lower in the CCRT‐R group than in the CCRT‐S group. Median CBX4 H‐score was 40.47 (IQR 26.61–54.45) in CCRT‐R versus 79.62 (IQR 67.21–108.95) in CCRT‐S (p < 0.001). Median NOX4 H‐score was 7.00 (IQR 3.43–22.56) versus 53.93 (IQR 35.44–75.44), respectively (p < 0.001). SP1 did not differ between groups (p = 0.695), and CTR1 showed a nonsignificant trend toward lower expression in CCRT‐R (p = 0.068) (Figure 1).

FIGURE 1.

Pretreatment expression of NOX4, CBX4, SP1, and CTR1 in cervical cancer biopsies. (A–D) Representative immunohistochemical images of NOX4, CBX4, SP1, and CTR1. (E–H) H‐score distributions by CCRT response group. CBX4 and NOX4 were lower in CCRT‐R tumors (both p < 0.001), whereas SP1 (p = 0.695) and CTR1 (p = 0.068) were not significantly different. CCRT‐R, CCRT‐resistant; CCRT‐S, CCRT‐sensitive.

With sensitivity status coded as CCRT‐S = 1, CBX4 and NOX4 each showed a moderate positive correlation with the CCRT‐S phenotype (CBX4 r = 0.564, p < 0.001; NOX4 r = 0.565, p < 0.001). CTR1 (r = 0.291, p = 0.069) and SP1 (r = 0.036, p = 0.824) were not significantly correlated with sensitivity status. CBX4 and NOX4 were moderately correlated with each other (r = 0.423, p = 0.0066), and SP1 and CTR1 were also correlated (r = 0.403, p = 0.0099) (Figure 2).

FIGURE 2.

Correlation of biomarker expression with CCRT response. (A) Correlation matrix. (B and C) CCRT sensitivity versus CBX4 and NOX4. (D) NOX4 versus CBX4. (E and F) CCRT sensitivity versus CTR1 and SP1. (G) SP1 versus CTR1. Pearson correlation coefficients and two‐sided p values are shown.

3.3. Discrimination and Internal Validation

Individual apparent AUCs for identifying the CCRT‐R phenotype were 0.840 for CBX4, 0.843 for NOX4, 0.670 for CTR1, and 0.538 for SP1. The parsimonious CBX4 + NOX4 model yielded an apparent AUC of 0.903. Internal validation showed lower but still substantial discrimination: LOOCV AUC 0.845, mean AUC 0.854 ± 0.022 across 100 repetitions of stratified 5‐fold cross‐validation, and bootstrap optimism‐corrected AUC 0.883 (Table 2; Figure 3). These internally validated values, rather than the apparent AUC alone, are used to characterize model performance. Because the markers were selected in this dataset, these estimates remain susceptible to residual feature‐selection optimism and are not measures of performance in an independent cohort.

TABLE 2.

Discrimination and internal validation of biomarker models for the CCRT‐resistant phenotype.

Model/marker	Apparent AUC	Internal validation
CBX4	0.840	—
NOX4	0.843	—
CTR1	0.670	—
SP1	0.538	—
CBX4 + NOX4	0.903	LOOCV AUC, 0.845; repeated stratified 5‐fold CV (100 repeats), mean AUC 0.854 ± 0.022; bootstrap optimism‐corrected AUC, 0.883

Note: Internal validation was applied to the parsimonious two‐marker model. Because the cohort was sampled 1:1 by CCRT response status, discrimination rather than absolute‐risk calibration is emphasized.

Abbreviations: AUC, area under the receiver operating characteristic curve; CV, cross‐validation; LOOCV, leave‐one‐out cross‐validation.

FIGURE 3.

Discrimination and internal validation of tissue biomarkers for the CCRT‐resistant phenotype. (A) Apparent ROC curves for individual markers and the parsimonious CBX4 + NOX4 model. (B) LOOCV ROC curve for the two‐marker model; repeated stratified 5‐fold cross‐validation and bootstrap optimism‐corrected AUC estimates are also shown. Apparent performance should not be interpreted as external validation.

For comparison, the previous four‐marker model had an apparent AUC of approximately 0.900 but did not improve apparent discrimination over CBX4 + NOX4 and had weaker repeated cross‐validated performance in reanalysis. SP1 and CTR1 were therefore retained as evaluated candidate biomarkers but were not included in the revised parsimonious response model.

3.4. Exploratory Overall Survival Analysis

Twenty deaths occurred during follow‐up, all among patients sampled into the CCRT‐R group; the reverse Kaplan–Meier estimate of median follow‐up was 2689 days (7.36 years). Because response status defined cohort sampling, all survival analyses should be interpreted as exploratory associations within this enriched sample rather than population survival estimates.

Using median H‐score cutoffs for visualization, 5‐year OS was higher in the CBX4‐high group than the CBX4‐low group (74.7% vs. 28.0%; log‐rank p = 0.0038) and in the NOX4‐high group than the NOX4‐low group (76.8% vs. 25.0%; p < 0.001). SP1 and CTR1 were not significantly associated with OS. In a joint two‐biomarker Cox model using standardized continuous expression, higher CBX4 remained associated with lower mortality (HR 0.31, 95% CI 0.13–0.74; p = 0.008), as did higher NOX4 (HR 0.48, 95% CI 0.26–0.90; p = 0.023) (Table 3; Figure 4). We avoid the term independent prognostic factor because clinical covariate adjustment would be unstable in a cohort with only 20 events.

TABLE 3.

Exploratory overall survival analysis by biomarker expression.

Marker	Median cutoff	5‐year OS, high vs. low	Log‐rank p	Joint Cox HR per 1‐SD increase (95% CI)	p
CBX4	54.67	74.7% vs. 28.0%	0.0038	0.31 (0.13–0.74)	0.008
NOX4	31.06	76.8% vs. 25.0%	< 0.001	0.48 (0.26–0.90)	0.023
SP1	47.02	53.0% vs. 48.1%	0.725	—	—
CTR1	26.21	58.2% vs. 42.9%	0.161	—	—

Note: The Cox model included CBX4 and NOX4 only, entered simultaneously after standardization, to limit the number of model parameters given 20 deaths. These analyses are exploratory because the cohort was selected according to CCRT response status.

Abbreviations: CI, confidence interval; HR, hazard ratio; OS, overall survival; SD, standard deviation.

FIGURE 4.

Exploratory overall survival analysis. (A, B) Kaplan–Meier curves stratified at the median H‐score for CBX4 and NOX4 with numbers at risk. (C) Joint two‐marker Cox proportional hazards model using standardized continuous CBX4 and NOX4; hazard ratios represent a 1‐SD increase in expression. Survival analyses are exploratory because the cohort was sampled according to CCRT response status.

3.5. Copper Staining

No visible dark green‐black copper deposits were identified in the 40 tumor specimens using the erythrine histochemical method. This finding is limited to histochemically detectable tissue deposits and does not demonstrate that copper transport, intracellular copper concentration, or cuproptosis‐related pathways are biologically inactive in LACC. CTR1 expression was not significantly associated with CCRT response in this cohort.

4. Discussion

This retrospective analysis identified associations between pretreatment tissue CBX4/NOX4 expression and the response‐defined CCRT‐resistant phenotype in locally advanced cervical squamous cell carcinoma. The apparent AUC of 0.903 for the two‐marker model decreased to 0.845 with leave‐one‐out cross‐validation, while repeated 5‐fold cross‐validation yielded a mean AUC of 0.854 and bootstrap optimism correction yielded 0.883. These internally validated estimates describe discrimination within the study sample only. Because the two markers were selected after examination of the same cohort, internal resampling may not fully account for model‐selection bias; moreover, it cannot establish transportability to different case mixes, institutions, or treatment settings. The model should therefore be regarded as exploratory and hypothesis‐generating, rather than as a clinically validated predictor.

4.1. CBX4: Association Without Mechanistic Inference

CBX4 participates in SUMO‐dependent chromatin regulation and DNA‐damage processes in experimental systems [8, 9, 10]. In esophageal squamous cell carcinoma, experimental CBX4 depletion enhanced radiosensitivity and altered autophagy‐related responses [7]. In contrast, our observational cervical cancer cohort showed an association of lower pretreatment CBX4 expression with the response‐defined CCRT‐R phenotype and poorer OS. This between‐tumor difference does not imply an opposite biological function in cervical cancer; variation in tumor context, sampling, confounding, and treatment exposure may also contribute. Neither the direction nor the mechanism of a potential effect of CBX4 on chemoradiotherapy sensitivity can be determined from tissue H‐scores alone.

We did not experimentally perturb CBX4 or measure DNA repair kinetics, γ‐H2AX, autophagic flux, or SUMOylation targets. Thus, the present clinical association cannot demonstrate that CBX4 alters DNA repair, autophagy, or treatment sensitivity in cervical cancer. The reported molecular functions of CBX4 provide candidate hypotheses for future tests, including CBX4 knockdown and overexpression, clonogenic survival after radiation and cisplatin, and DNA‐damage/repair assays [8, 9, 10], together with exploratory LC3B/p62‐based autophagy measurements. These proposed mechanisms must be distinguished from the observations reported here.

4.2. NOX4 and the Context Dependence of ROS Biology

In this cohort, lower NOX4 immunohistochemical expression was associated with CCRT resistance as defined by subsequent disease failure and with poorer OS. This is a clinical association, not evidence that NOX4‐generated ROS caused radiosensitization. Previous experimental studies have reported tumor‐promoting and tumor‐restraining NOX4 functions in different cellular contexts [16, 17, 18, 19, 20, 21]. Such context dependence precludes extrapolation of a single NOX4 mechanism to our patients and does not resolve whether NOX4 expression reflects tumor‐cell signaling, other tissue components, or correlated disease characteristics.

The study did not measure ROS production, oxidative DNA damage, antioxidant responses, NOX4 enzyme activity, DNA repair, or downstream redox signaling. Consequently, a mechanism in which increased NOX4‐derived ROS enhances radiation‐ or cisplatin‐induced cytotoxicity is one possible hypothesis, not a finding of this study; alternative explanations remain plausible. Mechanistic testing would require NOX4 loss‐ and gain‐of‐function experiments with irradiation and cisplatin, direct ROS and oxidative‐damage measurements, rescue interventions, and treatment‐response assays. These experiments are needed before assigning a causal role to NOX4 in chemoradiotherapy sensitivity.

4.3. Model Performance and Clinical Interpretation

The two‐marker model achieved an apparent AUC of 0.903 in the 40‐patient development sample; LOOCV yielded an AUC of 0.845, repeated stratified 5‐fold cross‐validation a mean AUC of 0.854, and bootstrap optimism correction an AUC of 0.883. These figures should not be viewed as independently validated accuracy estimates. CBX4 and NOX4 were retained following examination of their associations and performance in this dataset, and the model‐development and assessment steps were not fully separated at the feature‐selection stage. Consequently, cross‐validation and bootstrap resampling may still underestimate the total optimism introduced by selecting markers and the model on the same observations. No estimate establishes generalizability beyond the current sample.

The deliberate 1:1 sampling of CCRT‐R and CCRT‐S patients also means that the model intercept and predicted probabilities cannot be interpreted as population‐level absolute risks or assumed to be calibrated in unselected clinical populations. We therefore do not advocate a classification cutoff, treatment escalation, or routine clinical implementation on the basis of these data. The necessary next step is independent validation of the prespecified CBX4 + NOX4 model, including its coefficients, in a sufficiently sized, consecutively recruited cohort, preferably across multiple centers. Only then should discrimination, calibration, clinical utility, and potential treatment‐stratification thresholds be investigated.

4.4. SP1, CTR1, and Copper Staining

SP1 and CTR1 were included because of their known roles in transcriptional regulation, copper homeostasis, and cisplatin transport [22, 23, 24, 25, 26]. Neither marker showed a statistically significant expression difference between response groups in this cohort, and adding them to the response model did not improve cross‐validated discrimination. These negative findings are informative but do not exclude smaller effects that a 40‐patient study is underpowered to detect.

Likewise, the absence of visible copper deposits should be interpreted narrowly. Experimental cervical cancer studies indicate that copper homeostasis and cuproptosis can be therapeutically relevant under specific perturbations [28], and radiotherapy‐cuproptosis interactions have been demonstrated in other experimental contexts [27]. Our histochemical result only indicates that conspicuous copper salt deposits were not detectable by the staining method used; it cannot rule out altered copper metabolism or the potential value of copper‐targeted strategies.

4.5. Limitations and Future Directions

This study has important limitations. First, the sample comprised only 40 patients at one center, selected in equal numbers by subsequent response status; selection and spectrum effects may make discrimination and survival associations unrepresentative of a consecutive cervical cancer cohort. No independent validation cohort was available. Second, the two‐marker model was developed after comparing candidate biomarkers in the same dataset. Although LOOCV, repeated cross‐validation, and bootstrap optimism correction were performed, internal validation cannot completely eliminate marker‐selection bias or establish external generalizability. Third, residual stage imbalance persisted despite attempted matching, and unmeasured clinical and treatment‐related confounding remains possible. Fourth, all 20 observed deaths were in the selected CCRT‐R group, so OS analyses remain descriptive and hypothesis‐generating rather than population prognostic estimates. Recurrence times were not complete enough for a separate recurrence‐free survival analysis.

Fifth, the cohort was restricted to cervical squamous cell carcinoma, and the study did not measure CBX4/NOX4 activity, ROS, oxidative DNA damage, repair kinetics, or autophagy, nor did it include functional interventions. The mechanisms discussed are informed by previous experimental literature but were not tested in this study. Future studies should consecutively enroll independent patients, ideally across centers; prespecify the two‐marker signature and analytic pipeline; collect complete stage, treatment‐delivery, recurrence, and survival information; and assess calibration and decision utility. Separate functional studies are required to determine whether CBX4 or NOX4 affects DNA repair, ROS signaling, or treatment sensitivity.

5. Conclusion

Lower pretreatment CBX4 and NOX4 expression was associated with the response‐defined CCRT‐resistant phenotype and poorer overall survival in this small, retrospectively selected cervical squamous cell carcinoma cohort. Although the exploratory two‐marker model retained discriminatory ability in internal resampling analyses, these findings may be affected by marker‐selection bias and cannot establish performance in new patients. Independent validation in a consecutively recruited cohort is required before clinical use can be considered. No causal role for CBX4, NOX4, DNA repair, or ROS signaling in chemoradiotherapy response was demonstrated.

Author Contributions

Xianfang Pang: methodology, software, writing – original draft, writing – review and editing, investigation. Xiyun Qin: methodology, investigation, writing – original draft, writing – review and editing. Lilin Que: formal analysis, visualization. Weiye Li: methodology, investigation. Zixuan Yang: formal analysis. Weili Yi: formal analysis. Jiao Tang: investigation. Yong Zhang: investigation. Shanshan Ma: supervision, investigation. Linjiang Pan: conceptualization, supervision, project administration, writing – review and editing, resources. Fang Wu: project administration, supervision, resources, conceptualization, writing – review and editing.

Funding

The authors have nothing to report.

Ethics Statement

This retrospective study was approved by the Institutional Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (Approval No. 2024‐K0522; approval date January 3, 2025) and was conducted in accordance with the Declaration of Helsinki. The requirement for informed consent was waived by the ethics committee because of the retrospective study design.

Consent

All reported data are anonymized and contain no direct individual identifiers.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The datasets used and/or analyzed during the current study are not publicly available because of patient privacy and confidentiality requirements but are available from the corresponding author on reasonable request, subject to institutional and ethical approvals.