Mediaboss Marketing

Identification of TRGs and their prognostic value

From the data obtained from the single-cell RNA-seq analyses of OC tissue (GSE184880 dataset), we identified six major types of cells, including T/NK cells, myeloid cells, Epithelial cells, Fibroblasts, B cells and endothelial cells (Fig.2A). Figure2B showed the expression of cell markers. We then extracted T/NK cells for further analysis. As result, T/NK cells could be re-clustered into CD8+ cytotoxic T, CD8+ exhausted T, NK, CD4+ exhausted T and CD4+ nave T based on expression pattern of cell markers (Fig.2C,D). Development trajectory analyses of T/NK cells unveiled that CD4+ nave T, CD8+ cytotoxic T, and NK were enriched in initial differentiation phase while CD4+ exhausted T and CD8+ exhausted T were enriched in terminal differentiation phase (Fig.2E). Based on the FindAllMarkers function of the Seurat package, we identified 384 TRGs. Compared with normal tissues, we obtained 9638 DEGs in OC tissues (Fig.2F), including 248 TRGs (Fig.2G) in TCGA dataset. Among these differentially expressed TRGs, a total of 41 genes were significantly associated with the prognosis of OC patients in TCGA dataset (Fig.2H, P<0.05).

Identification of TRGs and their prognostic value. (A) t-SNE plot showing the identified cell types of from 7 ovarian cancer sample. (B) Dotplot showing average expression levels of cell marker. (C,D) SNE plot of sub-cell types of T cells and dotplot of expression pattern of cell markers. (E) Developmental trajectory of T cells inferred by monocle, colored by pseudotime and cell subtype. (F) Volcano plot showing DEGs in ovarian cancer. (G) Overlap between DEGs and TRGs. (H) Potential biomarkers identified by univariate cox analysis.

These 41 potential prognostic biomarkers were submitted to an integrative machine learning procedure including 10 methods, with which we developed a stable TRPS. As a result, we obtained a total of 101 kinds of prognostic models and their C-index in training and testing cohorts were shown in Fig.3A. The data suggested that the prognostic signature constructed by Enet (alpha=0.3) method was considered as the optimal TRPS with a highest average C-index of 0.58 (Fig.3A). The optimal TRPS was developed by 18 TRGs. The formula of the risk score was shown in Supplementary methods and results. Using the best cut-off value, we then divided into ovarian cancer cases into high and low TRPS score. As expected, OC patients with high risk score had a poor OS rate in TCGA cohort (P<0.001), GSE14764 cohort (P=0.0146), GSE26193 cohort (P=0.0039), GSE26712 cohort (P=0.0013), GSE63885 cohort (P<0.001) and GSE140082 (P=0.0032) cohort (Fig.3BG), with the AUCs of 2-, 3-, and 4-year being 0.728, 0.783, and 0.773 in TCGA cohort; 0.629, 0.642, and 0.739 in GSE14764 cohort; 0.617, 0.644, and 0.616 in GSE26193 cohort; 0.607, 0.587, and 0.591 in GSE26712 cohort, 0.672, 0.646 and 0.721 in GSE63885 cohort, 0.608 and 0.617 in GSE140082 cohort, respectively (Fig.3BG).

Identification of TRPS by machine learning. (A) The C-index of 101 kinds prognostic models constructed by 10 machine learning algorithms in training and testing cohort. (BG) The survival curve of ovarian cancer patients with different TRPS score and their corresponding ROC curve in TCGA, GSE14764, GSE26193, GSE26172, GSE63885 and GSE140082 cohort.

To compare the performance of TRPS with other prognostic signatures in predicting the OS rate of OC cases, we randomly collected 45 OC-related prognostic signatures (Supplementary Table 1) and calculated their C-index. As a result, the C-index of TRPS was higher than most of these prognostic signatures in TCGA dataset (Fig.4A). Moreover, the C-index of TRPS was higher than that of tumor grade and clinical stage in training and testing cohorts (Fig.4BF). These evidences suggested that the predictive value of TRPS in predicting the clinical outcome of OC patients was higher than most of signatures and clinical characters. However, we could not evaluate the predictive value of TRPS in predicting the OS rate of OC patients in GSE26712 cohort due to the missing data of tumor grade and clinical stage. Based on the result of univariate and multivariate cox regression analysis, TRPS served as an independent risk factor for the clinical outcome of OC patients in TCGA, GSE14764, GSE26193, GSE63885 and GSE140082 cohort (Fig.4G,H, all P<0.05). To predict the 1-year, 3-year and 5-year OS rate of OC patients, we then constructed a nomogram based on TRPS, clinical stage and tumor grade using TCGA dataset (Fig.4I). The comparison between the predicted curve and the ideal curve showed a high coincidence in TCGA dataset (Fig.4J). Compared with TPRS, clinical stage and tumor grade, the AUC of nomogram were higher in TCGA dataset (Fig.4K).

Evaluation the performance of TRPS in predicting prognosis of OC patients. (A) C-index of TRPS and other 45 established signatures in predicting the prognosis of OC patients. (BF) The C-index of TRPS, tumor grade and clinical stage in predicting prognosis of OC patients in TCGA, GSE14764, GSE26193, GSE63885 and GSE140082 cohort. (G,H) Univariate and multivariate cox regression analysis considering grade, stage and TRPS in training and testing cohort. (I,J) Predictive nomogram and calibration evaluating the 1-y, 3-y and 5-y overall survival rate of OC patients. (K) ROC curve evaluated the performance of nomogram in predicting prognosis of OC patients.

As shown in Fig.5A, TRPS showed significant correlation with the abundance of immune cells in TCGA dataset (all P<0.05). More specifically, TRPS showed a negative correlation with immuno-activated cell infiltration, such as CD8+ T cells, plasma cells, macrophage M1 and NK cells in TCGA dataset (Fig.5BE, all P<0.05). Interestingly, higher risk score indicated a higher level of cancer-related fibroblasts in TCGA dataset (Fig.5F). Similar results were obtained in ssGSEA analysis, suggesting a higher abundance of immuno-activated cells in low risk score group, including aDCs, B cells, CD8+ T cells, Neutrophils, NK cells, Tfh and TIL in TCGA dataset (Fig.5G, all P<0.05). Previous studies showed that macrophage M2/M1 polarization played a vital role in the progression of cancer9,10. Our study showed that OC patients with high risk score had a higher macrophage M2/M1 polarization in TCGA, GSE26712, and GSE140082 cohort (Fig.5H, all P<0.05). Further analysis suggested a higher stromal score, immune score and ESTIMAE score in low risk score group in TCGA dataset (Fig.5I, all P<0.001). Moreover, higher risk score indicated a higher APC co-stimulation score, CCR score, cytolytic activity score, para-inflammation promoting score, parainflammation and T cell co-stimulation score in TCGA dataset (Fig.5J).

Correlation between immune microenvironment and TRPS in OC. (A) Seven state-of-the-art algorithms evaluating the correlation between TRPS and immune cell infiltration in OC. (BF) The correlation between TRPS and the abundance of CD8+ T cells, plasma cells, macrophage M1 and CAFs. (G) The level of immune cells in different TRPS score group based on ssGSEA analysis. (H) The macrophage M2/M1 ratio in different TRPS score group in TCGA, GSE26712 and GSE140082 dataset. (I,J) The stromal score, immune score, ESTIMAE score and immune-related functions score in different TRPS score group. *P<0.05, **P<0.01, ***P<0.001.

High HLA-related gene expression indicated wider range of antigen presentation, increasing the likelihood of presenting more immunogenic antigens, and the likelihood of benefiting from immunotherapy11. We found that OC patients with low risk score had a higher HLA-related genes in TCGA dataset (Fig.6A, all P<0.05). Immune checkpoints played a vital role in immune escape of cancer. Based on our results, the expression of most of immune checkpoints was higher in high risk score groups in OC in TCGA dataset (Fig.6B, all P<0.05). Previous study showed that high TMB score was correlated with a better response to immunotherapy12. IPS was a superior predictor of response to anti-CTLA-4 and anti-PD-1 antibody and high IPS indicated a better response to immunotherapy13. High TIDE score indicated a greater likelihood of immune escape and less effectiveness of ICI treatment14. As showed in Fig.6CF, OC patients with low risk score had a higher TMB score, higher PD1 immunophenoscore, CTLA4 immunophenoscore, and PD1&CTLA4 immunophenoscore, lower immune escape score, lower TIDE score, lower T cell exclusion and dysfunction score in TCGA dataset. Thus, OC patients with low risk score may have a better immunotherapy benefit. To further verify the predictive value of TRPS in immunotherapy benefits, we then applied two immunotherapy cohorts to further verify our results. As shown in Fig.6G, the risk score in non-responders was significantly higher than that in responders in IMvigor210 cohort (P<0.01). Moreover, high risk score indicated a poor clinical outcome and lower response rate in IMvigor210 cohort (Fig.6G). Similar results were obtained in GSE91061 cohort (Fig.6H). As the vital role of chemotherapy, targeted therapy and endocrinotherapy for the treatment of OC, we also detected the IC50 value of common drugs in OC patients. We found that the IC50 value of 5-Fluorouracil, Camptothecin, Cisplatin, Gemcitabine, Foretunib, KRAS inhibitor, Erlotinib, and Tamoxifen were higher in in OC patients with high risk score in TCGA dataset (Fig.7A, all P<0.05). Moreover, positive correlation was obtained between risk score and these drugs in TCGA dataset (Fig.7B). Thus, OC patients with low risk score may be better sensitivity to chemotherapy and targeted therapy.

TRPS as an indicator for immunotherapy response in OC. (A,B) The level of HLA-related genes and immune checkpoints in different TRPS score group. (BF) The TMB score, immunophenoscore, immune escape score and TIDE, T cell dysfunction and exclusion score in different TRPS score group. (G,H) The overall rate and immunotherapy response rate in patients with high and low risk score in GSE91061 and IMvigor210 cohort. *P<0.05, **P<0.01, ***P<0.001.

The IC50 value of common drugs in different TRPS score group. (A) Low risk score indicated a lower IC50 value of common drugs. (B) The correlation between IC50 value of common drugs and TRPS score.

We finally performed gene set enrichment analysis to explore the potential mechanism mediating the difference of OC patients in clinical outcome, immune infiltration, and therapy response. High risk score indicated a higher sore of angiogenesis, DNA repair, EMT, G2M checkpoint, glycolysis, hypoxia, IL2-STAT5 signaling, IL6-JAK-STAT3 signaling, MTORC1 signaling, NOTCH signaling, P53 pathway, and P13K-AKT-mTOR signaling in OC in TCGA dataset (Fig.8AL, all P<0.05).

Gene set enrichment analysis in different TRPS score group. High risk score indicated a higher score of angiogenesis (A), DNA repair (B), EMT (C), G2M checkpoint (D), glycolysis (E), hypoxia (F), IL2-STAT5 signaling (G), IL6-JAK-STAT3 signaling (H), MTORC1 signaling (I), NOTCH signaling (J), P53 pathway (K), and P13K-AKT-mTOR signaling (L).

To further verify the performance of TRPS, we selected ARL6IP5 that contributed the most to the TRPS for further analysis. We first examined the expression of ARL6IP5 in OC cell lines, which showed that the expression of ARL6IP5 was lower in OC cell lines (Fig.9A). Typical immunohistochemical of ARL6IP5 in OC and normal tissues were showed in Fig.9B. In the follow-up study, the results of the CCK-8 assay proved that overexpression of ARL6IP5 obviously inhibited the proliferation of SKOV3 and TOV21G (Fig.9C,D).

Validation of the potential function of ARL6IP5 in OC by in vitro assays. (A) Comparison of ARL6IP5 expressions in normal and OC cell lines. (B) Typical immunohistochemical of ARL6IP5 in OC and normal tissues. (C,D) CCK-8 assay showed that overexpression of ARL6IP5 obviously inhibited the proliferation of SKOV3 and TOV21G cells. *P<0.05, **P<0.01.

See the original post:
Machine learning developed a CD8+ exhausted T cells signature for predicting prognosis, immune infiltration and drug ... - Nature.com