2.1 Selection of prognosis-related cells
The contents of 64 immune and stromal cells in the TME of TNBC were calculated through xCell algorithm. An overview of the scores in the training cohort was shown in Fig. 2a. To further determine the prognostic value of these cells, a univariate Cox regression model was applied. CD4+ memory T cells (HR [hazard ratio] 0.324, 95%CI [confidence interval] 0.138–0.765, P < 0.05), M2 macrophages (HR 2.918, 95%CI 1.217–6.996, P < 0.05), mv endothelial cells (HR 2.780, 95%CI 1.152–6.707, P < 0.05), CD8+T cells (HR 0.373, 95%CI 0.158–0.878, P < 0.05), endothelial cells (HR 2.622, 95%CI 1.080–6.376, P < 0.03), and Th2 cells (HR 0.409, 95%CI 0.175–0.957, P < 0.05) were identified (Fig. 2b). Kaplan-Meier analysis of these 6 kinds of cells were shown in Fig. 2c. The log-rank test of OS showed consistent results that higher scores of CD4+ memory T cells, CD8+ T cells, and Th2 cells were associated with better prognosis, while M2 macrophages, endothelial cells, and mv endothelial cells did the opposite.
2.2 Model Construction And Patients Classification
In order to comprehensively quantify the TME of TNBC and better predict patients’ prognosis, the random survival forest (RSF) model applied to construct a novel scoring system. Above-mentioned 6 prognosis-related cells were further selected by variable importance (VIMP) and minimal depth analysis for the model construction (Fig. 3a-b). During the VIMP analysis, M2 macrophages showed the highest variable importance, followed by CD4+ memory T cells. However, Th2 cells should be discarded from the model for their negative VIMP. Through the minimum depth analysis, we found that the minimum depth of CD4+ memory T cells is the smallest, indicating that they are the most powerful in prognostic prediction. Th2 cells have the largest minimum depth and exceed the selection threshold, so they are excluded. As is well known, the endothelial cell subtypes are hard to distinguish by IHC assay. To facilitate clinical application, only CD4+ memory T cells, CD8+ T cells, and M2 macrophages were selected to build up the model. The decision tree was applied for phenotypes classification. Accordingly, patients were devided into 4 phenotypes basing on the scores of these 3 kinds of cells. Patients with a M2low feature were classified as type 1 (N = 79, 50%), M2highCD8+ThighCD4+Thigh as type 2 (N = 30, 19%), M2highCD8+ThighCD4+Tlow as type 3 (N = 8, 5%), and M2highCD8+Tlow as type 4 (N = 41, 26%) (Fig. 3c). In order to further quantify, the score of each patient was evaluated by the multivariate Cox regression model. The scores of type 3 and 4 were significantly higher than those of type 1 and 2, which was consistence with the results obtained by the RSF model, indicating that types 3 and 4 patients are in a relatively high-risk state (Fig. 3d). Hence, type 1 and 2 patients were merged into the low-risk group, while types 3 and 4 patients were in the high-risk group.
2.3 Prognostic And Diagnostic Value Of The Classifier
The analysis above was used to construct a comprehensive indicator and identify 4 phenotypes and 2 subgroups of TNBC patients with distinctive features. To verify the prognostic value of our scoring system, Kaplan-Meier analysis was applied in the training cohort. The 5-year survival curves of type 1 and type 2 showed better survival than those of type 3 and Type 4 (Fig. 4a). To validate the prediction accuracy of our scoring system, the time-dependent AUC and ROC curves were employed for further evaluation (Fig. 4b-c). The time-dependent AUC was higher than 0.6, suggesting that the prediction model may be an independent factor in predicting the prognosis of TNBC. We found that the AUC was 0.706 in the 1st year, 0.607 in the 3rd year, and 0.678 in the 5th year, indicating the moderate ability of our system in prognosis prediction. In the training cohort, Kaplan-Meier analysis of OS suggested that the low-risk group had a better prognosis than the high-risk group (P < 0.05) (Fig. 4d). To confirm that the scoring system shared similar prognostic value in different populations, we applied it to the validation cohorts. The result of the METABRIC cohort was consistent with that of the training cohort, disclosing that our scoring system was able to discriminate patients with better or worse OS (Fig. 4e). Though the difference did not reach statistical significance, the low-risk group showed a better OS in the GSE58812 cohort.
To further investigate the rationality and universality of our scoring system, IHC assay was performed in the SYSUCC cohort (N = 51) to facilitate its clinical application. Based on the evaluation of CD4+ T cells, CD8+ T cells, and M2 macrophages, patients were classified into abovementioned 4 phenotypes and consistent result could be achieved in the Kaplan-Meier analysis of OS (Figure S2a-b).
2.4 Classification Associated With Immune Pathways And Checkpoint Molecules
Further results showed the score of each patient and its correlation with the expressions of immune-related genes and clinical prognosis. A total of 404 differentially expressed genes were included in this analysis. Patients with lower scores expressed higher levels of immune-related genes, which indicated that these patients have a more active immune microenvironment. For overall survival, patients with higher scores seemed to have a higher rate of death, which indicates that the newly defined scoring system may be an independent and promising prognostic predictor for TNBC (Fig. 5a). Subsequent analyses proved that several key immune checkpoint molecules, PD-L1, PD-1, and CTLA4, were significantly upregulated in TNBC patients with lower scores, demonstrating that the low-risk group might be more sensitive to immunotherapy. The expression of some immune-related molecules, such as LAG3 and TIGIT, were significantly higher in the low-risk group, indicating their roles as potential therapeutic targets in TNBC treatment (Fig. 5b). We then applied GSEA to examine the relevant signaling pathways involved in patients with low scores. Our results revealed that several key immune-related pathways were significantly enriched in the low-risk group, including immune response regulating cell surface receptor signaling, T cell and B cell activation, lymphocyte and B cell mediated immunity, immunoglobulin production, cytokine receptor binding, T cell receptor complex, and antigen receptor mediated signaling (Fig. 5c and Figure S1). Moreover, the ESTIMATE method was employed to explore the overall TME status. A high immune score, but not the stromal score, were observed in the low-risk group (Fig. 5d). To further investigate the cooperative contribution of infiltrating immune cells between the two groups, CIBERSORT algorithm was used to estimate the proportion of 28 immune cells in the TCGA cohort (Fig. 5e). Patients in the low-risk group presented a higher percentage of antitumoral immune cells, including activated B cells (P < 0.01), activated CD4+ T cells (P < 0.01), activated CD8+ T cells (P < 0.01), NK cells (P < 0.01), and activated dendritic cells (P < 0.01). Taken together, these results showed that there was a more active immune microenvironment in the low-risk group compared with the high-risk one.