RNA sequencing data from 80 UVM patients, along with corresponding clinical information, were collected as the training cohort from The Cancer Genome Atlas (TCGA) database (https://www.cancer.gov/tcga). Dysregulated gene expression was analyzed using datasets from the Gene Expression Omnibus (GEO), including GSE22138 with 63 UVM samples and GSE29801 with 175 normal retinal pigmented epithelium-choroid complex samples. A validation cohort consisting of 370 UVM patients was obtained from the International Cancer Genome Consortium (ICGC) database (https://www.icgc-argo.org). Another validation GSE84976 database was further used to evaluate the prognostic analysis of individual genes, including 28 UVM patients.

Additionally, a total of three UVM tumor samples and matched adjacent normal tissues were collected as clinical validation samples. RNA sequencing (RNA-seq) of tissue samples was performed on an Illumina Nova X Plus (Novogene, Beijing, China) using a paired-end approach with 150-bp reads. The fragments per kilobase of transcript per million mapped reads (FPKM) of genes in clinical samples, including three UVM tumor samples and matched adjacent normal tissues, were provided in Supplementary Table 1 . All human samples used in this study were obtained with written informed consent from participating patients, in accordance with the ethical guidelines of The First Affiliated Hospital of Anhui Medical University. Specimens were anonymized and handled in compliance with the Declaration of Helsinki.

All data analyses were performed using R software (version 4.1.0) and associated Bioconductor packages. Additionally, 1,065 oxidative stress-related genes (OSGs) were retrieved from previous literature for further investigation (see Supplementary Table 2 ).

A total of 63 UVM samples from the GSE22138 dataset and 175 normal retinal pigmented epithelium-choroid complex samples from GSE29801 (21) were included in the analysis. To minimize batch effects in the combined RNA sequencing data from GSE22138 and GSE29801, the “normalizeBetweenArrays” function from the limma R package and the “ComBat” function from the sva R package were used. Differentially expressed genes (DEGs) between UVM and normal samples were identified using limma, with a significance threshold of P <0.01 and |log₂(fold change)| >1.5. Dysregulated OSGs were visualized in a Venn plot using the VennDiagram R package. Furthermore, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis was performed using the KEGG Orthology Based Annotation System (KOBAS) database (http://bioinfo.org/kobas/).

A PPI network was constructed using the STRING database (http://string-db.org/) (22). Subclusters within the PPI network of OSGs were created to identify candidate hub genes for further analysis, using a median degree threshold (degree cutoff > 61). The subclusters were visualized using the Cytoscape software (version 3.8.2).

Multivariate Cox regression analysis was conducted using the survminer R package to refine the set of OSGs with the best predictive performance. The formula for the risk signature model was defined as follows: risk score = Exp^DNM2 × 0.133679 + Exp^POMC × 0.591935 + Exp^HSP90B1× + Exp^POMC × 0.591935 + Exp^CALM1 × 0.221127 + Exp^ACACA × 0.889051. All UVM patients were classified into high-risk and low-risk groups based on the median risk score (0.960617). Kaplan–Meier survival curves were generated using the survminer package. Additionally, a receiver operating characteristic (ROC) curve was plotted using the survivalROC package to assess the predictive accuracy of the prognostic model.

Next, we investigated the relationship between OSG-related risk signature and clinicopathological characteristics, including age, gender, race, and stage. The “rms” R package was used to develop the nomogram to illustrate each patient’s 1-, 3-, and 5-year overall survival probability, integrating OSG-related risk signature and clinicopathological features. Calibration curves were used to confirm the consistency between the predicted and the actual overall survival.

To explore the mechanisms underlying the prognostic impact of OSGs in UVM, immune cell infiltration analysis was conducted using the Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORT) algorithm. Furthermore, the Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data (ESTIMATE) algorithm was employed to compute immune and stromal scores based on gene expression data, using the estimate R package. Correlations between these scores and either individual risk scores or specific hub genes were then assessed.

Six machine learning algorithms were applied to identify key genes with significant predictive power. The least absolute shrinkage and selection operator (LASSO) method, implemented via the glmnet R package, performed sparse regularization to select crucial genes. Random forest analysis was carried out using the randomForest package to determine gene importance via ensemble learning with decision trees. Support vector machine (SVM) regression, implemented with the e1071 package, contributed to gene selection. Gradient boosting machine (GBM), using the gbm package, applied a sequential ensemble method to screen for predictive genes. Extreme gradient boosting (XGBoost), accessed through the xgboost package, employed advanced tree penalization to refine gene selection. A neural network (NNET) model was constructed using the nnet package. All models were built using 10-fold cross-validation. Model performance on the validation set was evaluated using the mlr3 package, with predictive accuracy quantified by the area under the ROC curve (AUC), calculated using the pROC R package.

The UVM cell lines (MP65, MM28) were provided by the Stem Cell Bank, Chinese Academy of Sciences. All cells were cultured in Roswell Park Memorial Institute (RPMI)-1640 medium supplemented with 20% fetal bovine serum (FBS, #C0226, Beyotime Biotechnology, China) and 1% penicillin/streptomycin (#C0222, Beyotime Biotechnology, China) at 37°C in a humidified atmosphere containing 5% CO₂. UVM cells were treated with hydrogen peroxide (H₂O₂) at concentrations of 50, 100, and 200 µmol/L (Merck KGaA, Darmstadt, Germany) in serum-free RPMI-1640 medium for 24 h. The CALM1 overexpression vector (oeCALM1) and a negative control vector were synthesized by GenePharma (Shanghai, China). Transfection was performed using Lipofectamine™ 2000 (Invitrogen, Carlsbad, USA) according to the manufacturer’s instructions. The medium was replaced three times per week, and cells were passaged upon reaching confluence.

Cells were harvested and washed twice with ice-cold phosphate-buffered saline (PBS) to remove residual medium. They were then lysed in radioimmunoprecipitation assay (RIPA) buffer containing 1% protease inhibitor cocktail. The supernatant containing total protein was collected, and protein concentration was determined using a bicinchoninic acid (BCA) protein assay kit (#P0012, Beyotime Biotechnology, China). Equal amounts of protein (20 µg) were mixed with 4× Laemmli buffer and loaded onto a 10% SDS-polyacrylamide gel (SDS-PAGE). After electrophoresis, proteins were transferred to a polyvinylidene difluoride (PVDF) membrane using a wet transfer system. The membrane was blocked in 5% non-fat dry milk diluted in Tris-buffered saline with Tween-20 (TBST; 20 mM of Tris, 150 mM of NaCl, 0.1% Tween-20, pH 7.6) for 1 h at room temperature. The membrane was then incubated overnight at 4°C with primary antibodies diluted in blocking buffer at the following concentrations: 1:1,000 for anti-GAPDH, 1:2,000 for anti-SOD2, 1:2,000 for anti-CAT, 1:3,000 for anti-CASP3, 1:3,000 for anti-BAX, and 1:2,000 for anti-CALM1. Afterward, the membrane was washed three times for 10 min each with TBST to remove unbound primary antibodies. It was then incubated with horseradish peroxidase (HRP)-conjugated secondary antibodies (anti-rabbit IgG, #A0208, Beyotime Biotechnology, China) diluted 1:200 in blocking buffer for 1 h at room temperature. The membrane was washed three times again with TBST, and protein bands were visualized using an enhanced chemiluminescence (ECL) substrate (#P0018S, Beyotime Biotechnology, China) and detected with a GeneGnome XRQ chemiluminescence imaging system.

Cell viability was assessed using the CCK-8 assay (#C0038, Beyotime Biotechnology, China). All UVM cells (MP65 or MM28) were seeded at a density of 5 × 10⁴ cells per well in 96-well plates and transfected with oeCALM1 or the control vector for 24 h. After transfection, 10 µL of CCK-8 reagent was added to each well, and the plates were incubated for 2 h at 37°C in 5% CO₂. The optical density (OD) at 450 nm was measured using a microplate reader (Thermo Fisher Scientific, USA).

Additionally, the concentrations of superoxide dismutase (SOD, #S0101S), malondialdehyde (MDA, #S0131S), and lactate dehydrogenase (LDH, #P0393S) were measured using ELISA kits (all from Beyotime Biotechnology, China), following the manufacturer’s protocols.

For migration, we obtained harvest cells (e.g., MP65 or MM28) and resuspended them in serum-free medium and seeded 2 × 10⁵ cells in the upper chamber. The complete medium (with 10% FBS) was added to the lower chamber as a chemoattractant. For oxidative stress conditions, cells were treated with 200 μmol/L of H₂O₂ in the upper chamber. For CALM1 overexpression groups, the overexpressing vector and the control negative vector were transfected into cells prior to seeding. Then, all cells were incubated for 48 h at 37°C with 5% CO₂. Next, we removed non-migrated cells from the upper chamber with a cotton swab. The migrated cells were fixed with 4% paraformaldehyde and stained with 0.1% crystal violet. The images were captured under a microscope, and cell counts were determined by analyzing three random fields per insert. For the cell invasion assay, the same steps as above were followed, but the transwell membrane was precoated with Matrigel (50 μg/mL, (BD Biosciences, San Jose, USA)) to simulate extracellular matrix barriers. All the cells were solidified by Matrigel for 4 h at 37°C before cell seeding. After incubation for 48 h, a microscope was used to capture the image.

Apoptosis was assessed using Annexin V-fluorescein isothiocyanate (FITC)/propidium iodide (PI) flow cytometry, as described in previous studies (23). Cells were stained with an Annexin V-FITC/PI Apoptosis Detection Kit (MedChemExpress, USA, #HY-K1073), and apoptosis was analyzed using a flow cytometer.

All results were expressed as mean ± standard deviation (SD). Statistical analyses were conducted using R software (version 4.1.0). Continuous variables were compared using the Wilcoxon rank-sum test. One-way analysis of variance (ANOVA) was applied to determine statistical significance among experimental groups. A P-value <0.05 was considered statistically significant unless otherwise indicated. A flowchart of the study design is presented in Figure 1 .

To analyze the differential expression of OSGs in UVM, 175 normal retinal pigmented epithelium-choroid complex samples from the GSE29801 dataset and 63 UVM samples from the GSE22138 dataset were included in the present study after removing batch effects. Principal component analysis (PCA) was conducted to assess batch effect correction ( Figures 2A, B ). A total of 2,576 DEGs were identified, consisting of 1,164 upregulated and 1,412 downregulated genes, using a threshold of P <0.01 and |log₂(fold change)| >1.5 ( Supplementary Table 3 ). A Venn diagram showed that 185 genes overlapped between the DEGs and OSGs ( Figure 2C ). These intersection genes were subsequently visualized in a hierarchical clustering heatmap ( Figure 2D ). Enrichment analysis revealed that these dysregulated OSGs were significantly enriched in pathways related to melanoma, cancer, HIF-1 signaling, FoxO signaling, PI3K-Akt signaling, T-cell receptor signaling, PD-L1 expression and the PD-1 checkpoint pathway, and apoptosis ( Figure 2E ).

To further investigate the interactions among OSGs, the dysregulated OSGs identified above were analyzed using the STRING database to construct a PPI network. The resulting network revealed extensive interactions among dysregulated OSGs ( Figure 3A ). To identify hub genes, subclusters of the PPI network were generated using a median degree cutoff (>61), resulting in 42 hub OSGs for subsequent analysis ( Figure 3B ). In addition, 54 prognostic genes were identified based on Kaplan–Meier survival curves. The intersection of hub genes from the PPI network and prognostic genes yielded 15 common genes ( Figure 3C ). Kaplan–Meier survival analysis demonstrated significant associations between gene expression levels and patient prognosis ( Figures 3D–R ). High expression was strongly correlated with poor survival, including ACACA ( Figure 3D ), CALM1 ( Figure 3F ), CALR ( Figure 3G ), CXCR4 ( Figure 3I ), DNM2 ( Figure 3J ), EDN1 ( Figure 3K ), HMOX1 ( Figure 3L ), HSP90B1 ( Figure 3M ), IL6 ( Figure 3N ), POMC ( Figure 3O ), TEK ( Figure 3Q ), and TNFSF10 ( Figure 3R ). Conversely, low expression of AKT2 ( Figure 3E ), CDK2 ( Figure 3H ), and SPP1 ( Figure 3P ) was associated with poor prognosis.

To identify novel genes with diagnostic potential, six machine learning algorithms were applied: LASSO regression, random forest, SVM, NNET, GBM, and XGBoost. ROC curve analysis showed that all models achieved area under the curve (AUC) values exceeding 0.85 ( Figures 4A–F ). Compared with random forest ( Figure 4B ), higher diagnostic potential was observed in the models including LASSO ( Figure 4A ), SVM ( Figure 4C ), NNET ( Figure 4D ), GBM ( Figure 4E ), and XGBoost ( Figure 4F ). Based on feature importance rankings, the top gene identified by each algorithm was selected for further analysis: ACACA (LASSO), CALM1 (random forest), HSP90B1 (SVM), DNM2 (NNET), ACACA (GBM), and POMC (XGBoost) ( Figure 4G ).

Using multivariate Cox regression analysis, a risk signature model was constructed based on the coefficients and expression levels of key genes identified through machine learning ( Figure 5A ). Patients were stratified into high- and low-risk groups according to the median risk score (0.960617). The high-risk group exhibited significantly poorer overall survival compared to the low-risk group ( Figure 5B ), suggesting that the risk score may serve as an independent prognostic factor in UVM. Risk score distribution ( Figure 5C ), survival status ( Figure 5D ), and gene expression profiles ( Figure 5E ) for each patient in the TCGA cohort were also visualized. ROC analysis demonstrated that the model had the highest sensitivity in predicting 5-year overall survival (AUC = 0.844), followed by 3-year (AUC = 0.759) and 1-year (AUC = 0.691) predictions ( Figure 5F ). To validate the model, the ICGC dataset was used in this study. Similar to the training set, the high-risk group in the validation cohort also exhibited significantly worse prognosis (P = 6.67e−03; Figure 6A ). Distributions of risk scores, survival outcomes, and gene expression profiles are shown in Figures 6B–D . ROC curves indicated moderate predictive sensitivity for 1-, 3-, and 5-year overall survival ( Figures 6E–G ).

To investigate the association between risk score and immune infiltration, the CIBERSORT algorithm was applied. The analysis of immune-related functions revealed that the OSG high-risk group exhibited elevated infiltration scores for CD4⁺ memory-activated T cells, activated NK cells, and resting mast cells ( Figure 7A ). In contrast, the OSG low-risk group showed higher infiltration scores for CD4⁺ memory resting T cells, resting NK cells, and monocytes ( Figure 7A ). Additionally, the correlations between novel OSGs in the risk signature model and immune cell infiltration scores were assessed. The results showed that these OSGs were positively correlated with CD4⁺ memory-activated T cells and resting mast cells ( Figure 7B ). Further analysis revealed that risk scores were positively correlated with CD4⁺ memory-activated T cells (cor = 0.388, P = 0.001078) and mast cells resting (cor = 0.386, P = 0.00108) while negatively correlated with resting NK cells (cor = −0.317, P = 0.007704) and CD4⁺ memory resting T cells (cor = −0.235, P = 0.04937) ( Figure 7C ). Moreover, ESTIMATE scores were compared between the OSG high- and OSG low-risk groups. The OSG high-risk group showed significantly lower immune scores, while stromal scores and total ESTIMATE scores did not differ significantly between the groups ( Figure 7D ).

To further accurately assess the prognostic value of OSGs in clinical information, the tumor stage, patient age, gender, and race were used to analyze the correlation with overall survival in this study. Our results demonstrate that the high-risk group consistently exhibited a significantly higher risk score ( Supplementary Figure 1A ), regardless of whether patients were younger or older than 60 years. Although the overall survival between patients aged >60 and ≤60 did not show a statistically significant difference ( Supplementary Figure 1B ), further stratified analysis revealed that the high-risk group showed significantly poorer survival outcomes in patients ≤60 years ( Supplementary Figure 1C ). Similarly, in patients >60 years, the high-risk group also displayed a comparable and statistically significant association with adverse prognosis ( Supplementary Figure 1D ).

Similarly, in our sex-based subgroup analysis, we observed comparable findings. Although no statistically significant difference in prognosis was detected between male and female patients overall, further stratification revealed that high-risk status remained strongly associated with adverse outcomes in both groups ( Supplementary Figures 1E, F ). Among female patients, high-risk individuals similarly demonstrated higher risk scores and worse prognosis ( Supplementary Figure 1G ). Likewise, among male patients, the high-risk group exhibited significantly elevated risk scores and poorer survival outcomes ( Supplementary Figure 1H ).

Due to limitations in the clinical annotation of TCGA datasets, our race-stratified analysis was restricted to comparing “White” versus “Other” populations. Despite this broad categorization, consistent patterns emerged regarding high-risk patients who demonstrated significantly elevated risk scores and worse overall survival in the White subgroup (P = 0.000989, Supplementary Figures 1I–K ).

Similarly, high-risk individuals of other racial subgroups showed poorer survival outcomes (P = 0.03089, Supplementary Figure 1L ). Notably, our risk stratification model revealed consistent prognostic value in both subgroups. In early-stage (I–II) patients, high-risk individuals exhibited significantly elevated risk scores (mean score: 2.326 vs. 0.453; P = 1.1317E−06, Supplementary Figure 1M ). When categorizing patients into early-stage (I–II) and late-stage (III–IV) disease, we observed that while the overall survival difference between these groups did not reach statistical significance, early-stage cases demonstrated a trend toward better prognosis ( Supplementary Figure 1N , P = 0.2286). Additionally, early-stage patients within high-risk groups showed worse clinical outcomes (log-rank P = 0.000774, Supplementary Figure 1O ). In advanced-stage (III–IV) patients, similar results showed comparable risk score differentiation (mean score: 3.123 vs. 0.543; P = 0.000105, Supplementary Figure 1M ), with poorer survival in the high-risk group (log-rank P = 0.01491, Supplementary Figure 1P ).

Additionally, we could calculate each patient’s total points and the corresponding overall survival probability using the constructed nomogram ( Supplementary Figure 1Q ). Calibration curves demonstrated high consistencies between the predicted overall survival and the actual overall survival rates at 1 year ( Supplementary Figure 1R ), 3 years ( Supplementary Figure 1S ), and 5 years ( Supplementary Figure 1T ). These results reinforced that our risk stratification model maintained robust predictive value across age, sex, race subgroups, and disease stages, potentially identifying high-risk patients who might benefit from more aggressive therapeutic interventions regardless of initial staging.

A total of five key genes were screened using machine learning and prognosis analysis in the TCGA training dataset. Further prognostic analysis of individual genes in the GSE84976 ( Supplementary Figures 2A–E ) database demonstrated that there were four key genes with consistent correlation of overall survival, with genes having high expression showing poor overall survival, including ACACA ( Supplementary Figure 2A , log-rank P = 0.000538), CALM1 ( Supplementary Figure 2B , log-rank P = 0.019632), DNM2 ( Supplementary Figure 2C , log-rank P = 0.03332), and POMC ( Supplementary Figure 2E , log-rank P = 0.002665). HSP90B showed no significant correlation with either expression or overall survival in the GSE84976 validation data ( Supplementary Figure 2D , log-rank P = 0.407925).

Moreover, the ICGC validation set also revealed consistent correlations between overall survival and three key genes: ACACA, CALM1, and DNM2, all showing high expression associated with poor prognosis ( Supplementary Figures 3A–C ), aligning with our initial results. HSP90B1 exhibited a discordant prognostic trend compared to our training data and GSE84976 validation cohort, with higher expression correlating with better survival in the ICGC data (P = 0.044791, Supplementary Figure 3D ). POMC showed no significant correlation with either expression or overall survival in the ICGC validation cohort (P = 0.762312, Supplementary Figure 3E ). Based on their reproducible prognostic associations, ACACA, CALM1, and DNM2 were prioritized for subsequent mechanistic studies.

The expression levels of key genes were determined in the TCGA training and ICGC testing data ( Supplementary Figures 4A–J ). The results revealed consistent expression between the high- and low-risk groups, including ACACA ( Supplementary Figures 4A,F ), CALM1 ( Supplementary Figures 4B, G ), DNM2 ( Supplementary Figures 4C, H ), and POMC ( Supplementary Figures 4E, J ). In our ICGC validation analyses, we observed a discordant expression pattern of HSP90B1 between the TCGA training set and the ICGC validation set. Specifically, HSP90B1 was significantly upregulated in the high-risk group within the TCGA cohort (P = 0.000159, Supplementary Figure 4D ), whereas it showed higher expression in the low-risk group in the ICGC dataset (P = 3.611E−31, Supplementary Figure 4I ). We opted not to include HSP90B1 in further analyses to ensure the robustness of our model.

In addition, ROC analysis was performed to evaluate diagnostic efficacy. In the TCGA training cohort, CALM1 demonstrated the highest AUC (0.9225), followed by POMC (AUC = 0.8556) and ACACA (AUC = 0.8356) ( Supplementary Figure 5A ). Similarly, in the ICGC validation data, CALM1 maintained high diagnostic performance (AUC = 0.8378) ( Supplementary Figure 5B ), exceeding the predefined threshold of AUC >0.8. We further evaluated the diagnosis of these genes using clinical samples, including three UVM tumor samples and matched adjacent normal tissues. The results demonstrated that CALM1 and HSP90B1 exhibited superior discriminatory power, with significantly higher AUC values compared to other candidate genes ( Supplementary Figure 5C ). Based on these findings, CALM1 was selected for further functional analysis to elucidate its biological role.

To evaluate the impact of oxidative stress on cell viability, UVM cells (MP65, MM28) were treated with various concentrations of H₂O₂ (0, 100, 200, 300, and 400 µmol/L) for 24 h. Cell viability was assessed using the CCK-8 assay. The results indicated that cell viability decreased significantly in a dose-dependent manner in MP65 ( Figure 8A ) and MM28 ( Supplementary Figure 6A ). A concentration of 200 µmol/L (cell viability: 57.2% in MP65 and 47.3% in MM28) was selected for subsequent experiments.

To further investigate the role of CALM1 in H₂O₂-induced oxidative stress, UVM cells were transfected with the CALM1 overexpression vector (oeCALM1 group) or a control vector and then treated with H₂O₂. Western blot analysis confirmed low expression of CALM1 in H₂O₂-treated cells compared to the control (CN) group in MP65 cells ( Figure 8B ). In the H₂O₂-challenged environment (200 µmol/L, 24 h duration), the CALM1 overexpression system achieved a robust 1.44 ± 0.012-fold increase in CALM1 expression in MM28 (Western blot quantification, normalized to GAPDH, P = 5.36E−05, Supplementary Figure 6B ), meeting the predetermined threshold for successful genetic manipulation. This overexpression efficiency remained stable throughout the oxidative stress exposure period (24 h post-treatment), establishing an appropriate experimental platform for probing CALM1’s functional role in redox regulation.

To evaluate the effect of CALM1 on oxidative stress, the levels of SOD, MDA, and LDH were measured using ELISA assays. H₂O₂ treatment significantly decreased SOD levels (P = 0.03698; Figure 8C ) and increased MDA and LDH levels (P = 0.008023 and P = 0.007522, respectively; Figures 8D, E ), compared to untreated controls (CN). Notably, oeCALM1 effectively reversed the reduction in SOD and suppressed the accumulation of MDA and LDH ( Figures 8C–E ) in H₂O₂-treated MP65 cells. Compared with the H₂O₂ negative control group (vector + H₂O₂), oeCALM1 also restored the expression of SOD2 (P = 0.02858) and CAT (P = 0.00918), both of which were suppressed by H₂O₂ treatment ( Figure 8F ). This protective pattern was consistently replicated in MM28 cells, where oeCALM1 similarly mitigated H₂O₂-mediated SOD suppression (50.39% recovery, P = 0.00046, Supplementary Figure 6C ) and reduced oxidative damage markers (MDA: 23.53% decrease; LDH: 29.81% decrease; both P < 0.05 vs. H₂O₂-treated vector controls, Supplementary Figures 6D, E ). Western blot results also revealed that oeCALM1 restored the expression of SOD2 (P = 0.02858) and CAT (P = 0.00918) in H₂O₂ treatment ( Supplementary Figure 6F ). The concordant results across both cell lines demonstrated the robust capacity of CALM1 to counteract oxidative stress regardless of cellular context.

Next, we examined the effect of CALM1 on apoptosis induced by oxidative stress. Flow cytometry analysis showed that H₂O₂ significantly increased the rate of apoptosis in MP65 cells, compared to untreated controls ( Figure 9A ). Importantly, this effect was attenuated by oeCALM1 transfection ( Figures 9A, B ), while necrosis remained unaffected. Compared with the CN group, the expression levels of CASP3 (caspase-3; P = 0.000797) and BAX (P = 0.00021) were significantly elevated following H₂O₂ treatment ( Figure 9C ). Overexpression of CALM1 mitigated the H₂O₂-induced increases in CASP3 (P = 0.00355) and BAX (P = 0.01071) expression levels ( Figure 9C ). These protective effects were consistently observed in MM28 cells, where oeCALM1 similarly attenuated both the percentage of apoptotic cells (49.14% reduction, P = 0.000464, Supplementary Figures 7A, B ) and the expression levels of CASP3 (50.60% decrease) and BAX (26.62% decrease) following H₂O₂ exposure ( Supplementary Figure 7C ).

Our investigation extended to examine the impact of CALM1 overexpression on proliferation, migration, and invasion under H₂O₂-induced oxidative stress (200 µmol/L, 24 h). CCK-8 assays demonstrated that oeCALM1 significantly enhanced cell proliferation by 37.58% (P = 0.00383, Supplementary Figure 8A ) in MP65 cells and 43.35% (P = 0.00734, Supplementary Figure 8B ) in MM28 cells at 72 h, effectively reversing the H₂O₂-mediated growth inhibition. This pro-proliferative effect was consistently observed across both cell lines. Furthermore, transwell assays revealed that oeCALM1 restored migratory capacity by 1.51-fold (P = 0.00288) in MP65 cells and 1.62-fold (P = 0.00436) in MM28 cells, compared to H₂O₂-treated controls ( Supplementary Figures 8C, D ). The oeCALM1 increased invasive potential by 1.84-fold (P = 0.00069) in MP65 cells and 2.22-fold (P = 6.013E−06) in MM28 cells, relative to oxidative stress conditions ( Supplementary Figures 8E, F ). The concordant results in both MP65 and MM28 cell lines established CALM1 as a multifunctional regulator capable of counteracting oxidative stress-induced impairment of critical oncogenic processes. These findings suggest CALM1 may serve as a key mediator in maintaining cellular functionality during redox imbalance.