Establishing Molecular Subgroups of CD8+ T Cell-Associated Genes in the Ovarian Cancer Tumour Microenvironment and Predicting the Immunotherapy Response

Background: The mechanism by which infiltrating CD8+ T lymphocytes in the tumour microenvironment influence the survival of patients with ovarian cancer (OC) remains unclear. Methods: To identify biomarkers to optimise OC treatment, 13 immune-cell-line-associated datasets, RNA sequencing data, and clinical data from the GEO, TCGA, and the ICGC were collected. Gene expression in OC was assessed using quantitative reverse transcription polymerase chain reaction (qRT-PCR) and immunohistochemistry (IHC) staining. Results: We identified 520 genes and three immunological clusters (IC1, IC2, and IC3) associated with CD8+ T cells. Higher IFN scores, immune T cell lytic activity, and immune cell infiltration and upregulated expression of immune-checkpoint-related genes indicated that IC3 is more responsive to immunotherapy, whereas IC1 and IC2 have a poorer prognosis. A 10-gene signature, including SEMA4F, CX3CR1, STX7, PASK, AKIRIN2, HEMGN, GBP5, NSG1, and CXorf65, was constructed, and a multivariate Cox regression analysis revealed a significant association between the 10-gene signature-based risk model and overall survival (p < 0.001). A nomogram was constructed with age and the 10-gene signature. Consistent with the bioinformatics analysis, IHC and qRT-PCR confirmed the accuracy of the signatures in OC tissue samples. The predictive ability of the risk model was demonstrated using the Imvigor210 immunotherapy dataset. Conclusions: The development of a novel gene signature associated with CD8+ T cells could facilitate more accurate prognostics and prediction of the immunotherapeutic response of patients with OC.


Introduction
Worldwide, approximately 230,000 new cases of ovarian cancer (OC) are diagnosed and 150,000 patients die every year [1]. Among gynaecological cancers, OC has the greatest mortality rate and only a 46% 5-year survival rate [2], with most patients eventually relapsing and developing resistance to chemotherapy. This loss in efficacy of conventional therapy necessitates the development of novel treatment plans for patients with OC.
Although a high proportion of homologous repair deficiency tumours in OC exhibit a high tumour mutational burden (TMB), increased CD8+ lymphocyte infiltration, and high tumour antigen expression that can independently trigger an antitumour response [3][4][5], clinical studies of OC immunotherapy have not yielded satisfactory results. Additional and more effective biomarkers are therefore required to identify patients that will be sensitive to OC checkpoint inhibitors.
This study aimed to develop a polygenic signature and predictive model for OC by identifying genes associated with CD8+ T cells using data from immune cell lines. This will allow identification of high-risk subpopulations most likely to benefit from immunotherapy, which will improve prognostics and prediction of immunotherapy efficacy in patients with OC.

Data Pre-Processing
The National Cancer Institute Genomic Data Commons API was used to download TCGA-OV RNA sequencing (RNA-seq) and clinical data. The International Cancer Genome Consortium provided the OV-AU dataset, and the Gene Expression Omnibus (GEO) provided the GSE26193, GSE30161, GSE63885, and GSE9891 datasets for GSE-OV cohort microarray and survival data.
The TMB was estimated from the TCGA mutect2-processed mutation dataset, and the single-sample GSEA approach was used to calculate IFN scores from Th1/IFN gene signatures as previously described [13].
Affy's RMA was used to process each immune-cell-line-associated dataset, and the limma "RemoveBatchEffect" R-package was used to eliminate batch effects.
TCGA-OV and OV-AU RNA-seq data samples lacking follow-up, survival time, or living status were eliminated. Ensemble gene IDs were transformed into gene symbols, and median values were determined when numerous gene symbols were expressed. Batch effect adjustments were made.
The GSE-OV dataset excluded normal tissue samples, samples without follow-up, OS, and living status. An annotation file transformed probes into gene symbols. The Affy RMA algorithm processed each immune-cell-line-associated dataset. Batch effects were corrected.
A single dataset was created from 13 immune-cell-line-associated datasets by merging the RNA-seq data of TCGA-OV and OV-AU dataset samples into the RNA-seq dataset, and merging the GSE26193, GSE30161, GSE63885, and GSE9891 microarray data into a single GSE-OV dataset. Principal component analysis verified the batch effect correction of the three merged datasets. The datasets were dispersed (Supplementary Figure S1A,C,E) but harmonised after batch effect correction (Supplementary Figure S1B,D,F).

Weighted Gene Correlation Network Analysis (WGCNA)
The WGCNA R-package was used to construct a weighted co-expression network based on the expression patterns of co-expressed coding genes and modules. Co-expression module screening used an eight-threshold soft threshold. The co-expression network was scale-free, with the log(k) of the node, where K represents the connection, negatively associated with log (P(k)), where P represents probability of occurrence, with a correlation coefficient of >0.8. The expression matrix was converted into an adjacency matrix and subsequently a topology matrix. Using average linkage hierarchical clustering, we constructed a topological overlap matrix of clustered genes. Hybrid dynamic tree cut-off criteria included 200 genes per gene network module. The dynamicTreecut R-package was used to identify gene modules and calculate the eigenvector values of each module. Modules were clustered and merged by applying height = 0.25, deepSplit = 2, and minModuleSize = 200.

Immunotyping
We performed a univariate Cox analysis of CD8+ T-cell genes and intersecting genes from the RNA-seq and GSE-OV datasets. The ConsensusClusterPlus R-package grouped the 466 RNA-seq samples according to the predicted CD8+ T-cell gene expression. The cumulative distribution function (CDF) was used to determine the optimal cluster number and the CDF delta area curve was used to determine the most stable number of groups. Subsequently, we evaluated the immune subtype characteristics of each group. The same analyses were performed for the GSE-OV group. The pooled Akaike information criterion (AIC) was used in stepwise regression using the MASS package. Briefly, the stepAIC technique involves starting with the model with the most variables and sequentially deleting variables to decrease the AIC, as a smaller AIC is indicative of a better fitted model. The workflow is shown in Supplementary Figure S2.

Specimen Collection
Surgical OC resection specimens and adjacent normal tissues were snap-frozen in liquid nitrogen and kept at −80 • C until RNA extraction. This study was approved by the Ethics Committee of the Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College (17-099/1355).

Statistical Analysis
Statistical analyses were performed with R software 3.5.3 and GraphPad Prism v. 8.01 (GraphPad Software, La Jolla, CA, USA). A Student's t test was used to compare values between test and control groups. p values of <0.05 indicated statistical significance.

CD8+ T Cell-Associated Genes
Hierarchical clustering generated 179 immune-cell-line-associated expression profiles (Supplementary Figure S3A). To obtain a scale-free network, we set β = 8 (Supplementary Figure S3B); 12 modules were obtained (Supplementary Figure S3C). Grey module genes could not be aggregated into other modules. The pink module contained 520 genes, most of which were positively linked to CD8+ T cells (Supplementary Figure S3D).
Gene ontology (GO) functional annotation determined the top ten biological processes (BPs, Supplementary Figure S4A Figure S4D) identified 10 significantly enriched immunological pathways involving primary immunodeficiency and T cell receptor signalling.

Immunophenotyping
Univariate analyses of the RNA-seq and GSE-OV datasets revealed 71 and 84 prognostic genes, respectively. Nine genes were included in both datasets (Supplementary Figure S5A), indicating that CD8+ T cell-associated gene expression may vary across sequencing platforms. We also analysed 146 prognostically important genes from both datasets (p < 0.05). When k = 3, the RNA-seq dataset consensus clustering analysis yielded three stable immunological clusters (ICs; Supplementary Figure S5B,C). The three ICs exhibit prognostic ability (Supplementary Figure S5D), with IC3 exhibiting a better prognosis than IC1 and IC2. These findings were successfully replicated in an independent GSE-OV cohort using the same methodology (Supplementary Figure S5E).

Relationship between Immunophenotyping and TMB and Gene Mutation
The TMB differed significantly between IC1 and IC3 ( Figure 1A). Comparison of the mutated gene counts indicated that IC1 had significantly more mutated genes than IC3 ( Figure 1B). Chi-square tests of 2431 genes with mutation frequencies of >3 and significantly high-frequency mutations in each IC yielded 202 genes (p < 0.05). The mutation signatures of the 15 genes in each IC group are shown in Figure 1C.

Immunophenotyping
Univariate analyses of the RNA-seq and GSE-OV datasets revealed 71 and 84 prognostic genes, respectively. Nine genes were included in both datasets ( Supplementary Figure S5A), indicating that CD8+ T cell-associated gene expression may vary across sequencing platforms. We also analysed 146 prognostically important genes from both datasets (p < 0.05).
When k = 3, the RNA-seq dataset consensus clustering analysis yielded three stable immunological clusters (ICs; Supplementary Figure S5B,C). The three ICs exhibit prognostic ability (Supplementary Figure S5D), with IC3 exhibiting a better prognosis than IC1 and IC2. These findings were successfully replicated in an independent GSE-OV cohort using the same methodology (Supplementary Figure S5E).

Relationship between Immunophenotyping and TMB and Gene Mutation
The TMB differed significantly between IC1 and IC3 ( Figure 1A). Comparison of the mutated gene counts indicated that IC1 had significantly more mutated genes than IC3 ( Figure 1B). Chi-square tests of 2431 genes with mutation frequencies of >3 and significantly high-frequency mutations in each IC yielded 202 genes (p < 0.05). The mutation signatures of the 15 genes in each IC group are shown in Figure 1C.

Immunophenotyping Chemokines and Immunological Checkpoint Genes
Examination of the RNA-seq cohort revealed differential chemokine expression for 28 of 33 chemokines (84.85%; Figure 2A), suggesting differences in immune cell infiltration

Immunophenotyping Chemokines and Immunological Checkpoint Genes
Examination of the RNA-seq cohort revealed differential chemokine expression for 28 of 33 chemokines (84.85%; Figure 2A), suggesting differences in immune cell infiltration and immunotherapeutic responses in the three ICs. Immunophenotyping also identified differential expression for 14 of 18 chemokine receptor genes ( Figure 2B). the angiogenesis scores of IC2 and IC3 subgroups were lower than those of the IC1 subgroup ( Figure 2E).
Evaluation of mean GZMA and PRF1 expression levels, used to assess intratumoural immune T-cell lysis activity [18], revealed substantial variation among the three ICs ( Figure 2D), with the IC3 group exhibiting greater immunological T-cell lysis activity than the IC1 and IC2 groups.
Based on the angiogenesis-related gene set obtained from a previous study [19], the angiogenesis score of each patient was calculated by the ssGSEA method in the R-package GSVA, and significant differences were found between the subgroups. It was found that the angiogenesis scores of IC2 and IC3 subgroups were lower than those of the IC1 subgroup ( Figure 2E).

Immune Signatures and Pathways in Different ICs
A CIBERSORT analysis of 22 immune cell scores in each sample of the RNA-seq dataset revealed substantial immunological signature differences between the subgroups ( Figure 3A). CD8+ T cells, resting memory CD4+ T cells, and M0, M1, and M2 macrophages were all highly expressed ( Figure 3B).

Immune Signatures and Pathways in Different ICs
A CIBERSORT analysis of 22 immune cell scores in each sample of the RNA-se taset revealed substantial immunological signature differences between the subgr ( Figure 3A). CD8+ T cells, resting memory CD4+ T cells, and M0, M1, and M2 m phages were all highly expressed ( Figure 3B).
Analysis of 10 carcinogenic pathways reported in a previous study [20] showed 9 pathways were significantly different between the three subtypes ( Figure 3C). IC hibited the highest microenvironment infiltration ( Figure 3D).
A comparison of six previously reported immunophenotypes of pan-carcinoma and the ICs in the current study indicated that the ICs intersected with four of th published immune subtypes ( Figure 3E, F). This suggests that the three ICs in this s are complimentary to the four previously published immunophenotypes.

Tumour Immune Dysfunction and Exclusion (TIDE) among ICs
The clinical effects of immunotherapy on the three ICs were assessed using (http://tide.dfci.harvard.edu/ (accessed on 20 August 2021)), with a higher TIDE predi score indicating a greater likelihood of immune escape and lower immunotherapy cacy. Analysis of the RNA-seq dataset indicated that IC2 and IC3 could benefit more immunotherapy than IC1 ( Figure 4A). Comparison of projected T cell dysfunction rejection ratings revealed that IC2 exhibits reduced T cell dysfunction, and the proj T cell rejection scores were greater for IC1 and IC2 than for IC3 ( Figure 4B,C). Consi trends were observed for the GSE dataset ( Figure 4D-F). Analysis of 10 carcinogenic pathways reported in a previous study [20] showed that 9 pathways were significantly different between the three subtypes ( Figure 3C). IC3 exhibited the highest microenvironment infiltration ( Figure 3D).
A comparison of six previously reported immunophenotypes of pan-carcinoma [21] and the ICs in the current study indicated that the ICs intersected with four of the six published immune subtypes ( Figure 3E,F). This suggests that the three ICs in this study are complimentary to the four previously published immunophenotypes.

Tumour Immune Dysfunction and Exclusion (TIDE) among ICs
The clinical effects of immunotherapy on the three ICs were assessed using TIDE (http://tide.dfci.harvard.edu/ (accessed on 20 August 2021)), with a higher TIDE prediction score indicating a greater likelihood of immune escape and lower immunotherapy efficacy. Analysis of the RNA-seq dataset indicated that IC2 and IC3 could benefit more from immunotherapy than IC1 ( Figure 4A). Comparison of projected T cell dysfunction and rejection ratings revealed that IC2 exhibits reduced T cell dysfunction, and the projected T cell rejection scores were greater for IC1 and IC2 than for IC3 ( Figure 4B,C). Consistent trends were observed for the GSE dataset ( Figure 4D-F).

Differential Analysis of ICs in Immunotherapy/Chemotherapy
We analysed variations in immunotherapy and chemotherapy efficacy. The GSE78220 dataset was used to examine immunological cluster similarity and immunotherapy efficacy via subclass mapping, with increasing similarity reflected by decreasing p-values. IC1 was similar to anti-PD-1_no response (anti-PD-1_NR) in both the RNA-seq ( Figure 4G) and GSE ( Figure 4I) datasets, suggesting it may be insensitive to PD-1 inhibitors. Evaluation of cisplatin, sunitinib, paclitaxel, crizotinib, and bleomycin indicated that IC2 was less responsive to these five medications than IC1 and IC3 ( Figure 4H,J).

Prognostic Risk Model
Training and validation sets were constructed from the 466 samples in the RNA-seq dataset. All samples were randomly clustered 100 times with replacement to prevent random allocation bias from affecting modelling stability. The training-to-validation ratio was 4:1 for group sampling.
Using the coxph function of the survival R-package and a filtering threshold of p < 0.01 on the training set, a univariate Cox proportional hazards regression model was constructed for each CD8+ T cell-related gene and survival data. In total, 11 prognostic genes were identified in the RNA-seq and GSE datasets. A multivariate Cox analysis (Supplementary Figure S6) and the stepwise regression algorithm reduced the 11 genes to 10 genes, including TXK, STX7, PASK, AKIRIN2, SEMA4F, HEMGN, GBP5, CX3CR1, NSG1, and CXorf65.
The 10-gene signature formula was as follows:

Differential Analysis of ICs in Immunotherapy/Chemotherapy
We analysed variations in immunotherapy and chemotherapy efficacy. The GSE78220 dataset was used to examine immunological cluster similarity and immunotherapy efficacy via subclass mapping, with increasing similarity reflected by decreasing p-values. IC1 was similar to anti-PD-1_no response (anti-PD-1_NR) in both the RNA-seq ( Figure 4G) and GSE ( Figure 4I) datasets, suggesting it may be insensitive to PD-1 inhibitors. Evaluation of cisplatin, sunitinib, paclitaxel, crizotinib, and bleomycin indicated that IC2 was less responsive to these five medications than IC1 and IC3 ( Figure 4H,J).

Prognostic Risk Model
Training and validation sets were constructed from the 466 samples in the RNA-seq dataset. All samples were randomly clustered 100 times with replacement to prevent random allocation bias from affecting modelling stability. The training-to-validation ratio was 4:1 for group sampling.
Using the coxph function of the survival R-package and a filtering threshold of p < 0.01 on the training set, a univariate Cox proportional hazards regression model was constructed for each CD8+ T cell-related gene and survival data. In total, 11 prognostic genes were identified in the RNA-seq and GSE datasets. A multivariate Cox analysis (Supplementary Figure S6) and the stepwise regression algorithm reduced the 11 genes to 10 genes, including TXK, STX7, PASK, AKIRIN2, SEMA4F, HEMGN, GBP5, CX3CR1, NSG1, and CXorf65. The risk score distribution (Supplementary Figure S7A) was generated for each sample based on its expression level in the RNA-seq training dataset. The timeROC R-package was used to classify RiskScore prognosis, and the prognostic prediction of the risk model at 1, 3, and 5 years assessed using a receiver operating characteristic (ROC) curve analysis (Supplementary Figure S7B) revealed a high area under the curve (AUC) for the model. RiskScore z-scores classified samples as high risk (>0) or low risk (<0), and subsequently constructed Kaplan-Meier (KM) curves indicated that the high-risk group had a significantly reduced survival rate (p < 0.0001) (Supplementary Figure S7C).

Prognostic Risk Model Validation
The RiskScore distribution of samples in the RNA-seq validation set is shown in Supplementary Figure S8A. An ROC analysis showed high AUC values for the risk model at 1, 3, and 5 years (Supplementary Figure S8B). KM curves showed lower survival in the high-risk group (p < 0.01; Supplementary Figure S8C).
Supplementary Figure S9A shows the RiskScore distribution of all the RNA-seq datasets. An ROC analysis revealed a high AUC for the risk model at 1, 3, and 5 years (Supplementary Figure S9B), and the KM curves indicated significantly reduced survival rates for the high-risk group (p < 0.0001; Supplementary Figure S9C).
Supplementary Figure S10A shows the RiskScore distribution of samples in the independent GSE validation dataset. The RiskScore prognostic classification was evaluated using an ROC analysis (Supplementary Figure S10B), which indicated a high AUC value. The KM analysis results are consistent with the results of the RNA-seq datasets (Supplementary Figure S10C, p < 0.001).
In addition, KM and ROC curves were constructed for RiskType using the survival information of TCGA cohort (Supplementary Figure S11A Figure S12I,J). High-risk patients aged 60 years or older, stage III-IV, and G1/2 and G3/4 subgroups exhibited a shorter OS than low-risk patients. Our risk model exhibited good predictive ability in diverse clinical clusters.

Nomograms and Forest Plots
Using the full TCGA dataset, we constructed an age-RiskScore nomogram ( Figure 5A). The RiskScore function of the 10-gene-based risk model had the greatest survival prediction ability. Calibration curves and DCA graphs are shown in Figure 5B,C. The nomogram performed well.

Predicting Immunotherapy Efficacy with the Constructed Risk Model
An immunotherapy dataset (Imvigor210) with transcriptome data was obtained to test the 10-gene signature risk model. KM curves showed that immunotherapy patients with higher RiskScores had poorer survival rates (Supplementary Figure S14A). RiskScore ROC curves showed higher AUC values (Supplementary Figure S14B). Immunotherapy responders and non-responders differed significantly in both high-risk and low-risk groups (Supplementary Figure S14C). The MCPcounter R-package calculated immune cell scores for Imvigor210 samples. TMB, NEO, and immune cell scores were inversely related to RiskScore (Supplementary Figure S14D).
Furthermore, among groups, immunotherapy efficacy, immune cell grouping, tumour cell grouping, and immunophenotypic grouping differed significantly according to the RiskScore (Supplementary Figure S15A-D).

Gene Expression in OC Tissues
To validate the 10-gene signature, we analysed the expression of four genes in paired tumour and adjacent non-tumour tissues from patients with OC. The clinical information of the patients with OC is shown in Supplementary Table S3. IHC and qRT-PCR revealed downregulated expression of HEMGN and TXK in OC tissues compared to normal tissues ( Figure 6A,B), whereas SEMA4F and STX7 were highly expressed in OC tissues. These results validate the bioinformatics analyses.

Predicting Immunotherapy Efficacy with the Constructed Risk Model
An immunotherapy dataset (Imvigor210) with transcriptome data was obtained to test the 10-gene signature risk model. KM curves showed that immunotherapy patients with higher RiskScores had poorer survival rates (Supplementary Figure S14A). RiskScore ROC curves showed higher AUC values (Supplementary Figure S14B). Immunotherapy responders and non-responders differed significantly in both high-risk and low-risk groups (Supplementary Figure S14C). The MCPcounter R-package calculated immune cell scores for Imvigor210 samples. TMB, NEO, and immune cell scores were inversely related to RiskScore (Supplementary Figure S14D).
Furthermore, among groups, immunotherapy efficacy, immune cell grouping, tumour cell grouping, and immunophenotypic grouping differed significantly according to the RiskScore (Supplementary Figure S15A-D).

Gene Expression in OC Tissues
To validate the 10-gene signature, we analysed the expression of four genes in paired tumour and adjacent non-tumour tissues from patients with OC. The clinical information of the patients with OC is shown in Supplementary Table S3. IHC and qRT-PCR revealed downregulated expression of HEMGN and TXK in OC tissues compared to normal tissues ( Figure 6A,B), whereas SEMA4F and STX7 were highly expressed in OC tissues. These results validate the bioinformatics analyses. Biomedicines 2023, 11, x FOR PEER REVIEW

Discussion
Bevacizumab has been one of the most studied targeted agents in OC in rec ades; however, despite its widespread use and efficacy, patient selection and tim treatment remain controversial and the survival benefit in patients with advanced mains limited [22]. Currently, immune checkpoint blockade (ICB) therapy is effe cancers such as melanoma and non-small cell lung cancer, but the therapeutic v immunotherapy in OC is still in the research stage. A phase II clinical study eva the role of combined anti-PD-1 antibodies and anti-VEGF antibodies in recurrent lial OC showed an overall response rate of 28.9%, with a 40% overall response platinum-sensitive OC [23]. The important reason limiting the response to immu apy and disease progression in OC is that the tumour microenvironment of OC state of immunosuppression [24]. Immunologically activated tumour-infiltrating l cytes (TILs) in ovarian tumour tissue demonstrate that the immune system is the for this tumour [25]. Therefore, the present study was conducted to construct a gen associated with CD8+ T lymphocytes to identify patients with OC who would bene immunotherapy and predict prognosis.
We identified three OC ICs and a risk prediction model based on genes ass with CD8+ T cells. The prognosis was worse for IC1 and IC2 than for IC3. TMB number of gene mutations were considerably higher in IC1 than in IC3. Chemokin receptor genes, and angiogenesis-related genes varied substantially between clust had a better prognosis than IC1 and IC2 owing to greater IFN scores, immune T c activity, immune cell infiltration, and immune checkpoint gene expression. IC higher TIDE score than IC2 and IC3, indicating a larger risk of immune escape a therapeutic benefits. IC1 and IC3 were more sensitive to cisplatin, sunitinib, pa crizotinib, and bleomycin than IC2, suggesting that the three ICs can identify h patients with OC and help clinicians choose treatment drugs.
In total, 10 CD8+ T cell-related genes (TXK, STX7, PASK, AKIRIN2, SE HEMGN, GBP5, CX3CR1, NSG1, and CXorf65) were identified in patients with O expression of STX7, SEMA4F, and CX3CR1 was associated with a high risk and poo nosis, whereas high expressions of TXK, PASK, AKIRIN2, HEMGN, GBP5, NSG CXorf65 were associated with a low risk and better prognosis. qRT-PCR and IH firmed the accuracy of the signatures in OC tissue samples and the bioinformatic ation. Human TXK primarily regulates IFNγ gene transcription in Th1/Th0 cells [2 taxin 7, encoded by STX7, affects lysosome trafficking and phagosome-lysosome

Discussion
Bevacizumab has been one of the most studied targeted agents in OC in recent decades; however, despite its widespread use and efficacy, patient selection and timing of treatment remain controversial and the survival benefit in patients with advanced OC remains limited [22]. Currently, immune checkpoint blockade (ICB) therapy is effective in cancers such as melanoma and non-small cell lung cancer, but the therapeutic value of immunotherapy in OC is still in the research stage. A phase II clinical study evaluating the role of combined anti-PD-1 antibodies and anti-VEGF antibodies in recurrent epithelial OC showed an overall response rate of 28.9%, with a 40% overall response rate in platinum-sensitive OC [23]. The important reason limiting the response to immunotherapy and disease progression in OC is that the tumour microenvironment of OC is in a state of immunosuppression [24]. Immunologically activated tumour-infiltrating lymphocytes (TILs) in ovarian tumour tissue demonstrate that the immune system is the trigger for this tumour [25]. Therefore, the present study was conducted to construct a gene model associated with CD8+ T lymphocytes to identify patients with OC who would benefit from immunotherapy and predict prognosis.
We identified three OC ICs and a risk prediction model based on genes associated with CD8+ T cells. The prognosis was worse for IC1 and IC2 than for IC3. TMB and the number of gene mutations were considerably higher in IC1 than in IC3. Chemokines, their receptor genes, and angiogenesis-related genes varied substantially between clusters. IC3 had a better prognosis than IC1 and IC2 owing to greater IFN scores, immune T cell lytic activity, immune cell infiltration, and immune checkpoint gene expression. IC1 had a higher TIDE score than IC2 and IC3, indicating a larger risk of immune escape and less therapeutic benefits. IC1 and IC3 were more sensitive to cisplatin, sunitinib, paclitaxel, crizotinib, and bleomycin than IC2, suggesting that the three ICs can identify high-risk patients with OC and help clinicians choose treatment drugs.
In total, 10 CD8+ T cell-related genes (TXK, STX7, PASK, AKIRIN2, SEMA4F, HEMGN, GBP5, CX3CR1, NSG1, and CXorf65) were identified in patients with OC. High expression of STX7, SEMA4F, and CX3CR1 was associated with a high risk and poor prognosis, whereas high expressions of TXK, PASK, AKIRIN2, HEMGN, GBP5, NSG1, and CXorf65 were associated with a low risk and better prognosis. qRT-PCR and IHC confirmed the accuracy of the signatures in OC tissue samples and the bioinformatics evaluation. Human TXK primarily regulates IFNγ gene transcription in Th1/Th0 cells [26]. Syntaxin 7, encoded by STX7, affects lysosome trafficking and phagosome-lysosome fusion [27]. PASK affects lipid and glucose metabolism, mitochondrial respiration, phosphorylation, and gene expression [28]. AKIRIN2 controls embryonic development and innate immunity, but its role in carcino-genesis is unclear [29]. AKIRIN2 expression is correlated with chemotherapy resistance in patients with OC [30]. SEMA4F regulates embryologic axon guidance and cancer-induced neurogenesis [31]. HEMGN is upregulated in thyroid carcinoma tissues and cells [32] and regulates cellular proliferation and apoptosis via the PI3K/Akt signalling pathway [33]. GBP5 has prognostic power for OS in OC [34]. The chemokine receptor CX3CR1 can identify distinct populations within monocyte, macrophage, and dendritic cell lineages [35]. NSG1 is a neuronal cell-expressed endosomal protein and a direct transcriptional target gene of the tumour suppressor p53 [36].
Although our study was a retrospective investigation, we mitigated potential biases by validating the model with data from multiple cohorts.

Conclusions
We identified three immunophenotypes of OC and constructed a 10-gene signaturebased risk model that accurately predicts the prognosis and response to immunotherapy of patients with OC.