A CTL/M2 macrophage-related four-gene signature predicting metastasis-free survival in triple-negative breast cancer treated with adjuvant radiotherapy

This study aimed to develop and validate a prognostic model for metastasis-free survival (MFS) based on genes that may functionally interact with cytotoxic T lymphocytes (CTLs) and M2 macrophages in patients with triple-negative breast cancer (TNBC) who underwent adjuvant radiotherapy. The transcriptional and phenotypic profiles of TNBC and other breast cancer subtypes were downloaded from gene expression omnibus (GEO). The abundance of infiltrated immune cells was evaluated through CIBERSORTx or MCP-counter. A weighted linear model, the score for MFS (SMFS), was developed using the least absolute shrinkage and selection operator (LASSO) in GSE58812 and validated in GSE2034 and GSE12276. The biological implication of the SMFS was explored by evaluating its associations with TNBC molecular subtypes and other radiosensitivity- or immune-related signatures. A model consisting of the PCDH12/ELP3, PCDH12/MSRA, and FAM160B2/MSRA gene expression ratios with non-zero coefficients finally selected by LASSO was developed using GSE58812. In GSE2034 (treatment with adjuvant radiotherapy), the SMFS was significantly associated with MFS in TNBC patients (hazard ratio (HR) = 8.767, 95% confidence interval (CI) 1.856–41.408, P = 0.006) and, to a lesser extent, in non-TNBC patients (HR = 2.888, 95% CI 1.076–7.750, P = 0.035). However, the interaction of subtype (TNBC vs non-TNBC) and the SMFS tended to be significant (Pinteraction = 0.081). In contrast, the SMFS was not significantly associated with MFS in either TNBC patients (P = 0.499) or non-TNBC patients (P = 0.536) in GSE12276 (treatment without radiotherapy). Among the four TNBC molecular subtypes, the c1 and c4 subtypes exhibited higher CTL infiltration and lower SMFS values than the c2 and c3 subtypes. In addition, the SMFS was positively correlated with the abundance of endothelial cells (r = 0.413, P < 0.001). The proposed model has the potential to predict MFS in TNBC patients after adjuvant radiotherapy, and the SMFS may represent a measurement of tumor immune suppression.


Introduction
Breast cancer is a malignant tumor with a high incidence worldwide and in China [1,2]. Triple-negative breast cancer (TNBC) accounts for 15-20% of all breast cancers cases. TNBC has the worst prognosis of all breast cancer subtypes and the 5-year mortality rate can reach 40% after the initial diagnosis compared with luminal breast cancer [3]. TNBC is highly aggressive; nearly 46% of patients with TNBC will develop distant metastasis, with a median survival time of only 13.3 months. Most distant metastases of TNBC occur within 3 years after the initial diagnosis, with a high He Xiao and Dong Wang were equally charged for supervising this study.
probability of brain metastasis and visceral metastasis [4,5]. According to the National Comprehensive Cancer Network (NCCN) guidelines for breast cancer, due to the young age of onset and the lack of targets for endocrine therapy and anti-HER2 therapy, radiotherapy (RT) and chemotherapy are usually used as adjuvant treatments for TNBC [6]. Although several transcriptomic signatures have been proposed for predicting the efficacy of RT for breast cancer [7][8][9], currently, no such indicators specific to TNBC have been identified. Therefore, screening the population suitable for RT to treat TNBC is a problem worthy of study.
TNBC is defined as breast cancer with negative estrogen receptor, progesterone receptor, and HER2 receptor expression [10]. Compared with luminal breast cancer, TNBC has specific biological characteristics, such as high proliferation, a high overlap ratio with basal-like tumors, more mesenchymal stem cells, and homologous recombination defects and BRCA1/2 inactivation [3,11,12]. Moreover, recent studies have demonstrated that TNBC has a higher level of T-cell infiltration accompanied by a higher level of immunosuppression than luminal breast cancer [13,14]. In one study, from among antigen presentation process genes, genes that interact with RT and affect the disease-specific survival (DSS) of breast cancer patients were screened and an immune signature (IMS) was constructed [15]. Another study showed that high tumor-infiltrating lymphocytes (TILs) in the primary tumor may independently reduce the risk of an ipsilateral breast tumor recurrence (IBTR) in breast cancer, whereas patients with low TILs may attain a greater benefit from RT with regard to the risk of IBTR [16]. However, this way of developing indicators using candidate gene sets as discovery sets or using a rough estimate of TILs to reflect immune status does not take into account the subclasses of infiltrated immune cells and their functions in TNBC tumors. Therefore, the performance of these signatures derived from the perspective of immunity to predict the benefit from RT should be further tested in TNBC.
Many studies have proposed radio-resistant gene signatures or recurrence prediction models after mastectomy or breastconserving surgery [7][8][9]. These signatures or models were primarily developed based on luminal breast cancer. Moreover, Sjöström found that the post-RT recurrence score calculated by the same gene signature had inconsistent effects on the prognosis for IBTR between the ER-and ER + groups receiving adjuvant RT. This result suggests that the mechanisms of RT resistance may be different among various breast cancer subtypes with different immune microenvironments and tumor cell characteristics [8]. Considering the high degree of immune infiltration in TNBC and the important impact of TILs on the prognosis of breast cancer patients undergoing adjuvant RT, we speculate that the subclass and number of immune infiltrating cells inherent in tumor masses and their cellular functional status may affect the efficacy of RT, resulting in different patient prognoses. Therefore, inspired by the method proposed by Jiang [17], a more general and robust method was developed to screen genes that may affect the function of certain immune cells in tumors, thereby affecting the survival of TNBC patients after RT. Then, based on these genes, we developed a new model, the score for metastasis-free survival (SMFS), which contains four genes and can be used as a prognostic marker of TNBC distant metastasis after RT.

Dataset retrieval and preprocessing
Gene expression omnibus (GEO) was screened with the keyword "breast cancer." Only tumor gene expression profiles obtained using arrays containing TNBC primary tissue samples were selected. All datasets used in this study are shown in Table 1. The raw data (CEL files) of each dataset were downloaded from GEO. The rma function implemented in the limma R package was used to perform background correction, normalization, and probe summarization for each dataset [18]. A total of twelve datasets were merged, and the batch effect was corrected using the main function virtualAr-rayComBat in the virtualArray R package [19]. For genes represented by several probes, the probe with the maximum interquartile range (IQR) was finally selected to determine the expression of that gene. Ultimately, a total of 12,402 genes were included. To expand the number of TNBC samples included in the consensus clustering analysis, a classifier discriminating TNBC samples from non-TNBC samples was first developed in the GSE21653 dataset using the support vector machine (SVM) algorithm. In brief, the expression of 749 differentially expressed genes (DEGs) identified in 87 TNBC samples and 179 non-TNBC samples through limma was used as the training expression matrix to construct the SVM classifier using the penalizedSVM R package [20]. The classifier was then verified in the GSE12276 dataset. The sensitivity and specificity of the SVM classifier in the training and validation sets were 94.25% and 96.65% and 95.45% and 93.33%, respectively. The number of TNBC samples used by the SVM classifier for GSE20685, GSE7390, and GSE31448 are shown in Table 1. A total of 920 TNBC samples were finally included in the subsequent analysis. We followed Reporting Recommendations for Tumor Marker Prognostic Studies (REMARK) guidelines in reporting this study [21].

Evaluation of the abundance of tumor-infiltrating immune cells, consensus clustering, and annotation
The abundance of tumor-infiltrating immune cells was evaluated using both MCP-counter [22] and CIBERSORTx [23] in GSE58812 and whole TNBC samples. In the MCP-counter method, the abundance of ten types of immune cells was represented as the log2 geometric mean of transcriptomic markers of these immune cell, called the MCPcounter score. Although the MCP-counter score cannot be presented as the actual fraction of each immune cell subpopulation in bulk tumor tissues, it has a numerical advantage for downstream statistical analysis. CIBERSORTx was used with the LM22 signature matrix and B-mode batch correction. The absolute score was used for enumeration of the abundance of the 22 types of immune cells. Consensus clustering of 4 subtypes was achieved with the CancerSubtypes R package using genes with 5% maximum variation in whole TNBC samples [24]. To characterize the biological features of each subtype, the log fold change (FC) values of all 12,402 genes were calculated for each subtype versus the rest, and the genes were subjected to gene set enrichment analysis with the gsePathway function implemented in the clusterProfiler and ReactomePA R packages [25].

Development of the MFS transcriptional score (SMFS) and validation
Having been inspired by the method proposed by Jiang [17], a similar approach was used to identify genes that could influence the function of certain immune cells and affect metastasis-free survival (MFS). In the GSE58812 dataset, the association of MFS with the abundance of immune cells was determined via univariate Cox regression. The abundance of cytotoxic T lymphocytes (CTLs) inferred from MCP-counter and M2 macrophages inferred from CIBERSORTx were found to be significantly positively and negatively associated with MFS, respectively (Supplementary Table 1). Based on these two univariate Cox models, gene expression and the interaction of the abundance of immune cells (CTLs or M2 macrophages) and gene expression were added, and a likelihood ratio test was used to evaluate the significance of these two terms. This analysis was performed for CTLs and M2 macrophages separately. In the Cox model with three terms, only genes in which the significance of immune cells and gene expression was less than 0.05 and the significance of the interaction and likelihood ratio test was less than 0.01 were chosen. In this situation, a z value of the interaction term less than zero indicated that the expression of the gene could improve the antitumor immune effect of CTLs, whereas a z value of the interaction term greater than zero indicated that the expression of the gene could enhance the tumor-promoting effect of M2 macrophages in terms of MFS. The gene expression ratio was calculated by dividing the expression of genes with z > 0 in the interaction with M2 macrophages (M2) by the expression of genes with z < 0 in the interaction with CTLs. The least absolute shrinkage and selection operator (LASSO) was used to perform penalized Cox regression. Fivefold cross-validation was used to select the lambda value using the cv.glmnet function. The final values of the coefficients of the gene expression ratios that were not zeroed out were used as the weights for the linear weight model for MFS prediction, which was defined as the score for MFS (SMFS). The linear weight model for MFS was validated in GSE2034.

Computation of previously published immune signature scores and radiosensitivity indices
Because the immune status of the tumor microenvironment has a clear impact on the long-term clinical outcomes of breast cancer patients treated by breast-conserving surgery followed by adjuvant RT, two immune-related signature scores, the IFNγ signature [26] and tumor immune dysfunction and exclusion (TIDE) [17], were used to further characterize the SMFS. The IFNγ signature score was calculated by averaging 17 genes in the "Expanded immune gene signature" shown in Table 2 published in the original article. Because of the prevalence of high CTL infiltration in TNBC tumor tissues (> 1%), only the T-cell dysfunction score was used. Ninety-three genes were retained to calculate the TIDE score. The TIDE value itself is a measurement of the degree of irreversible exhaustion of CTLs. The higher the TIDE value is, the more irreversible the exhaustion of CTLs [17]. In addition to immune-related signatures, the 10-gene radiosensitivity index (RSI) [27] and 27-gene adjuvant radiotherapy intensification classifier (ARTIC) [7] were also included to illustrate differences in the TNBC specificity of these radiosensitivity indices as well as the SMFS. However, only 22 of the 27 originally published genes were retained to calculate the ARTIC score. The IMS was calculated according to the original formula [15].

Molecular subtype
Lehmann's TNBC subtyping [28] was performed using a web tool. The z-score-transformed expression matrix of a total of 603 cases was loaded into the web service after removing ER-positive samples from 920 TNBC samples based on the criteria proposed by the authors [28]. Because some signature genes representing four subtypes in Burstein's classification were missed in the expression matrix of the 920 TNBC samples, Burstein's molecular classification was achieved through the method of nearest template prediction [29] using GSE76124 as the training set [30]. Fifty genes were selected to represent each of four original subtypes and a total of 200 genes composed the classifier. Pearson correlation coefficients were used as metrics of similarity. All cases with P values less than 0.1 were retained for subsequent analysis.

Statistical analysis
The abundance of immune cells, signature scores, and other continuous variables are presented as the median value and IQR and visualized with scatter plots and box plots. The difference in the abundance of immune cells and signature scores among the four TNBC subtypes was evaluated using a Kruskal-Wallis test accompanied by pairwise comparisons. A chi-square test and Mann-Whitney U test were used to evaluate differences in category and continuous variates between TNBC and non-TNBC subsets, respectively. The Kaplan-Meier method and a log-rank test were used to compare the MFS between different subgroups. Univariate and multivariate Cox regression analyses were used to evaluate independent prognostic factors for MFS. All statistical analyses were performed using SPSS 17.0 software (IBM SPSS, Chicago, IL, USA). All tests were two-sided, and P < 0.05 was considered statistically significant.

Development and characteristics of the SMFS
Because almost all TNBC patients enrolled in GSE58812 (96.26%, 103/107) had received adjuvant RT without systemic therapy, GSE58812 was used as the discovery and training set. The z values of the interaction term of CTLs or M2 macrophages with all genes were separately subjected to Reactome analysis. The Reactome pathways that were enriched for improvement of CTL functions (enrichment in z < 0) were involved in "Antigen processing-Cross presentation," "Interferon gamma signaling," and "PD-1 signaling." The "Collagen biosynthesis and modifying enzymes," "Extracellular matrix organization," and "Integrin cell surface interactions" pathways possibly participated in the enhancement of M2 functions (Fig. 1, Supplementary Table 2). Overall, 58 and 115 genes that had z < 0 and z > 0 in the interaction term for CTLs and M2 macrophages, respectively, were identified (Supplementary Table 3 Table 2). All patients were divided into low-and high-risk groups according to the 75th percentile of the SMFS, and patients with low risk had significantly longer MFS than those with high risk (5-year MFS: 79.9% vs. 32.3%, log rank P < 0.001) (Fig. 2b).

Validation of SMFS in GSE2034 and GSE12276
Comparisons of baseline clinical characteristics between TNBC and non-TNBC samples in GSE2034 and GSE12276 are shown in Supplementary Table 4. RT was given to 87% of patients in GSE2034, whereas none of the patients in GSE12276 received adjuvant RT. There were 41 TNBC patients in the GSE2034 dataset. In the subset of TNBC patients in GSE2034, only the SMFS was significantly associated with MFS (HR = 8.767, 95% CI 1.856-41.408, P = 0.006) ( Table 3). The patients in the whole GSE2034 dataset were also categorized into low-and high-risk groups based on the 75th percentile of the SMFS. Patients with low risk exhibited profoundly superior MFS compared with patients with high risk in the TNBC subset (Fig. 2c).
Although the SMFS was also significantly related to prognosis in the non-TNBC subpopulation, the effect was less pronounced with the same cut-off value (Fig. 2d). Although the SMFS and proliferation were independent prognostic factors for MFS in the non-TNBC patient subset in GSE2034 (Table 4), the interaction of TNBC and the SMFS tended to be significant in the whole GSE2034 population   (Table 3). These results suggest that the SMFS had a more pronounced association with prognosis in TNBC patients who received adjuvant RT than the ARTIC value, which was developed mainly based on patients with luminal breast cancer. Notably, the SMFS did not exhibit any relationships with prognosis in either the TNBC or non-TNBC subset in the GSE12276 dataset, roughly illustrating the RT specificity of the SMFS (Table 3, Fig. 2e, f). In fact, age, T stage, and the ARTIC score remained significantly associated with MFS in multivariate Cox regression in the non-TNBC subset in GSE12276 (Table 4).

Subtype and SMFS
To gain insight into the biological characteristics of the SMFS, the associations of the SMFS with TNBC subtypes and other immune signatures were examined. Consensus clustering separated 920 TNBC samples into four subtypes, c1 (n = 208), c2 (n = 314), c3 (n = 206), and c4 (n = 192). The average silhouette width was 0.88 (Fig. 3a). Reactome analysis revealed that the c1 subtype had the properties of highly active mitosis (R-HSA-69620), the c2 subtype had an EMT phenotype (R-HSA-3000171), the nuclear receptor transcription pathway predominated in the c3 subtype (R-HSA-383280), and high immune cell signaling was found in the c4 subtype (R-HSA-909733) (Supplementary Table 5). The abundance of CTLs and M2 macrophages, IFNγ signature, RSI, ARTIC score, TIDE score, IMS, and SMFS were significantly different among the four subtypes (Fig. 3b, Supplementary Table 6). In particular, the abundance of CTLs and score of IFNγ in the c1 and c4 subtypes were significantly higher than those in the c2 and c3 subtypes, respectively, whereas the abundance of M2 macrophages, RSI, and SMFS in the c1 and c4 subtypes were significantly lower than those in the c2 and c3 subtypes. The SMFS was slightly but significantly correlated with the IFNγ signature, TIDE score, RSI, and ARTIC value. However, the RSI was strongly negatively correlated with IFNγ score and the abundance of CTLs (Fig. 3c). Nevertheless, the SMFS was moderately positively correlated with the abundance of endothelial cells inferred by MCP-counter (Fig. 4). Burstein and Lehmann subtypes were available for 510 samples among a total of 920 TNBC patients (Supplementary Table 7). The c1 subtype was observed to almost only be composed of both basal-like immune-activated (BLIA) and basal-like immunosuppressed (BLIS) subtypes of the Burstein classification, the c2 subtype consisted of mostly BLIS tumors, the c3 subtype almost only included the luminal androgen receptor (LAR) and mesenchymal (MES) subtypes of both the Burstein and Lehmann classifications, and the c4 subtype was similar to c1 and involved a large proportion of the BLIA subtype of the Burstein classification (Fig. 5).

Discussion
In this study, we developed an SMFS based on a gene expression ratio which was calculated by dividing the expression of genes enhancing the tumor-promoting effect of M2 macrophages by the expression of genes improving the antitumor immune effect of CTLs. This ratio takes into account the possible interaction between CTLs and M2 macrophages to some extent [31]. Therefore, we speculate that the higher the SMFS value is, the higher the degree of tumor immunosuppression and the lower the SMFS value is, the stronger the antitumor immune effect. The SMFS was validated after stratification of patients for TNBC status and RT, creating four groups (TNBC with RT, non-TNBC with RT, TNBC with no RT, and non-TNBC with no RT). In the RT group (GSE2034), the SMFS was an independent prognostic factor for MFS in both patients with TNBC and non-TNBC. However, the SMFS was not associated with prognosis in terms of MFS in the no RT group (GSE12276), regardless of TNBC or non-TNBC status. GSE2034 and GSE12276 were two independent datasets, and the administration of RT was not random, thus verifying the interaction between the SMFS and RT in one cohort was impossible to achieve in a rigorous manner. However, these results could indirectly indicate the RT specificity of the SMFS. In addition, the SMFS showed an interaction trend with the TNBC subtype in GSE2034 (P interaction = 0.081), demonstrating that the SMFS may be more strongly linked to prognosis in TNBC patients than in non-TNBC patients. Compared with non-TNBC patients, TNBC patients with rapid distant recurrence (MFS < 1 year) had a higher median SMFS (P = 0.001). These results suggest that the SMFS could be a prognostic indicator for MFS specific to TNBC patients treated with adjuvant RT.
We also compared the performance of the SMFS with previously published immune or RT signatures in breast cancer. First, we found that proliferation and the SMFS were independent prognostic factors for MFS in the non-TNBC  (Table 4). This finding indicates that proliferation is a very important factor affecting the impact of RT in patients with non-TNBC. Previous studies have also demonstrated that the radiosensitivity signature (RSS) and single-sample predictors (SSPs) can predict the prognosis for IBTR in patients with ER + tumors owing to their biological effect on proliferation [8,9]. Second, previous studies found that the RSI was associated with IBTR prognosis in the ER-RT + group [8,27]. However, our results showed that the RSI was associated with prognosis in terms of MFS in the training set (TNBC with RT), determined by univariate Cox regression, but the association was not significant in multivariate analysis (Table 2). Furthermore, the performance of the RSI was poor in all groups in the validation sets, likely because the RSI and ARTIC were developed with survival fraction at 2 Gy (SF2) [32] and locoregional recurrence (LRR) [7] as endpoints, which mainly reflect the local control effect of RT, that is, the direct effect of RT. However, RT might have an indirect impact on long-term survival through the immune system, which would not be reflected by the ARTIC or RSI. In fact, certain evidence has indicated that the RT sensitivity of solid tumors is associated with immune activation [33]. Third, in our analysis, the IMS showed an ability to predict MFS only in the non-TNBC with RT group (GSE2034) (univariate P = 0.003, multivariate P = 0.069). The IMS was developed in the E-TABM-158 dataset [34], which only contains 15% TNBC patients, which might explain why the IMS did not show prognostic efficacy in the TNBC with RT group. Finally, we found that T3/T4 patients had a worse prognosis than T1 patients in the no RT group (GSE12276), but no such phenomenon was observed in the RT group (GSE2034) (Tables 3 and 4). According to the clinical guidelines for breast cancer, T3 (tumor size > 5 cm) tumors should be treated with RT after surgery [6], which likely indicates that T3/T4 patients in GSE12276 might obtain a better prognosis by receiving RT.
Many previous studies have shown that the immune system plays an important role in mediating the antitumor effects of RT [35,36]. For instance, RT can activate an antitumor immune effect by inducing maturation of dendritic cells and enhancing activation of T cells [37,38]. In other words, RT can eliminate the immunosuppressive state of cancer and turn immunologically "cold" tumors "hot" [39]. In our study, patients in the low-risk group (a low SMFS) may derive a greater benefit from indirect RT effects through the immune system, leading to a better prognosis. The SMFS developed by our approach may allow selection of breast cancer patient populations, especially those with TNBC, suitable for RT from the perspective of immunity. RT is generally considered to be an important local control method in breast cancer treatment [40], but the endpoint we used in the training set and validation sets was MFS, rather than the usual endpoints of IBTR or LRR. The reasons are as follows. First, the clinical outcome indicators available in public datasets are limited, especially in satisfying the conditions of both RT and a sufficient number of patients with TNBC. Second, at present, the local control effect of RT for breast cancer has been very good, and there are many indicators to predict the local control efficacy of RT [7,8,27]. Our focus is on the interaction between immunity and RT in TNBC, and the antitumor immune effect is usually considered to be related to the long-term cancer outcomes [41,42]. At present, TNBC patients are prone to developing distant metastasis after first-line treatment, and the survival time after distant metastasis is very short [4,5]. Antitumor immune effects play an important role in the distant metastasis of tumors [43,44]. Therefore, investigating how to combine the antitumor effects of RT and the immune system and maximizing their role in the treatment of TNBC is the fundamental goal of this study. Recently, a study on the application of RT combined with immunotherapy in metastatic TNBC showed encouraging results [45]. However, currently, no predictive biomarkers exist for RT combined with immunotherapy. We developed and validated a CTL/ M2 macrophage-related four-gene signature (SMFS) that had prognostic value for MFS in TNBC patients undergoing RT, which could provide information to achieve this goal, but the SMFS needs to be further verified in large randomized clinical trials or even in trials of checkpoint blockade therapy plus RT.
Each of our four TNBC subtypes seemed to comprise two or several Burstein subtypes, illustrating moderate heterogeneity between these two classification systems. We found that the c1 and c4 subtypes had the highest proportion of BLIA tumors but had a lower SMFS value and longer MFS. This further confirms that the lower the SMFS is, the stronger the antitumor immune activation. We also found that the SMFS was significantly positively correlated with the abundance of endothelial cells inferred by MCP-counter. Single-cell sequencing research in lung cancer showed that endothelial cells in tumors can downregulate immune cell homing and genes correlated with T-cell activity [46]. This suggests that the angiogenesis of tumors may be different from that of normal tissues and may impair the antitumor immune effect. The SMFS included 4 genes, among which ELP3 (Gene ID: 55,140) and MSRA (4482) are genes that can improve the antitumor immune effect of CTLs, while FAM160B2 (64,760) and PCDH12 (51,294) are genes that can enhance the tumor-promoting effect of M2 macrophages. However, these genes did not overlap with the markers of endothelial cells or CTLs inferred by MCP-counter or the markers of M2 macrophages inferred by CIBERSORTx (data not shown). These four genes were incorporated into the SMFS model in the form of PCDH12/ELP3, PCDH12/ MSRA, and FAM160B2/MSRA gene pairs. Therefore, these genes may be expressed on any cells in the tumor bulk and reflect the immunosuppression status of the tumor and immune ecosystem in the form of the SMFS.
There are some limitations in our study. First, the number of TNBC cases in the validation sets was too small and the administration of RT was not randomized which means that it was impossible to verify the predictive effect of the SMFS and the interaction effect on RT. Second, the current clinical treatment for invasive breast cancer almost always includes chemotherapy and RT and even neoadjuvant therapy has been widely used in the treatment of TNBC. However, most patients in GSE58812 and GSE2034 only received adjuvant RT, and patients in GSE12276 received chemotherapy but not RT. Third, in terms of the long-term survival influenced by the antitumor immune effects of RT, overall survival may be more appropriate as the endpoint to further develop classifiers [42]. Fourth, the biological characteristics of the SMFS need to be further analyzed with single-cell sequencing data of TNBC to evaluate the precise intercellular communications by which the functions of TILs or macrophages are impacted.
In conclusion, we developed a CTL/M2 macrophagerelated four-gene signature (SMFS) that had prognostic value for MFS in TNBC patients undergoing RT and then validated our SMFS in two independent datasets of patients with or without RT. Our research provides an idea on how to use transcriptional data to screen genes interacting with tumor-infiltrating immune cells to develop prognostic or predictive indicators for RT. This study may provide new ideas for development of biomarkers to guide the combined use of RT with immunotherapy in the future.