SCLC samples of 102 subjects were collected from January 2014 to December 2018 in Shanghai Pulmonary Hospital, China. We used the eighth edition of tumor, node, metastasis classification for lung cancer for stage identification and prognosis evaluation.20

For all formalin-fixed, and paraffin-embedded slides, xylene and alcohol were used for dewaxing, followed by aqua destillata for rinsing. Next, we used the target retrieval solution (Dako, DM828/DM829) and recovered antigens under high pressure and heat. To decrease background staining, 0.3% hydrogen peroxide was used. We applied purified anti-human FOXP3 (1:100, BioLegend 320101) as primary antibody and goat-anti-rabbit/anti-mouse IgG that was conjugated with horseradish peroxidase as secondary antibody for testing.

Two pathologists (CW and LZ) reviewed and determined clinically pathological samples independently. Disagreements were resolved by consensus in these two reviewers. The cut-off value of FOXP3 level on TILs was defined as ≥1% staining. We performed survival analysis for cut-off determination.21 22

We adapted machine learning algorithm XGBoost for importance evaluation and immune risk model construction.23 In addition to discovering non-linear relationships by working with data of the first and second derivatives, this machine learning method also controls the overfitting as well as overcomplexity of one predictive model by employing regularization item. XGBoost algorithm can also score the importance of each attribute as value measure.

In this study, we repeated 1000 times for cohort division and model construction to make full use of the sample information. The whole 102 samples were divided into training set and calibration set (7:3) randomly. By incorporating all immune biomarkers, the XGBoost-based model was constructed and the top three features were selected for further survival prediction. Then, we used the calibration set to validate the model performance and performed receiver operating characteristic (ROC) curve analysis for visualizing its predictive value. To construct immune risk assessment system for patients with SCLC, we combined the importance outcomes that ranked the top three of XGBoost model and the coefficients of multivariate Cox regressions that included these immune biomarkers. The final equation was calculated: immune risk score=(−0.067×FOXP3)−(0.023×PD-L1 on TILs)+(0.002×CD8). Further, we confirmed the prognostic role of immune risk score in SCLC and compared its predictive performance with single-indicator-based and two-indicator-based immune risk score systems.

The entire set was randomly divided into the training and validation set (n=71 and n=31, respectively) by r. Based on the independently predictive factors of multivariate Cox regression analyses, we adopted nomogram algorithm for relapse prediction in SCLC. After establishing the nomogram, the calibration plots for relapse-free survival (RFS) of 1, 3, and 5 years graphically demonstrated the correlation between predicted and observed risks of each outcome for appraising the predictive ability of this model. To fully test its performance, the nomogram was subjected to 100 bootstrap resamples in the training cohort for internal verification. We also used C-index and ROC curve analyses of 1, 3, and 5 years to assess its discriminating ability.

To reveal the clinical value of this prognostic nomogram, we further regrouped all subjects and made low-risk, moderate-risk, and high-risk stratifications. Survival analyses were conducted for each group and were compared by log-rank test.

For further assessing the clinical value of FOXP3 in patients with SCLC, the clinical data from cBioportal Database were used (https://www.cbioportal.org). We screened samples based on the inclusion criteria24: (1) mRNA sequencing (mRNA-seq) of SCLC tissues, (2) complete data for mRNA expression, and (3) complete prognostic information from patients with SCLC.

To figure out the expression of FOXP3, CD274 that encodes PD-L1 protein and CD8A that encodes CD8 protein in SCLC tissues, we used the Gene Expression Omnibus (GEO) Database for verification (https://www.ncbi.nlm.nih.gov/geo/). mRNA expression data enrolled must follow the inclusion criteria: (1) complete data for mRNA expression, (2) mRNA-seq of SCLC tissues, and (3) mRNA-seq of normal tissues. The exclusion criteria were as follows: (1) insufficient data for comparing gene expression, (2) sample from animals, and (3) mRNA-seq for SCLC cell lines.

To investigate relevant biological pathways between distinct FOXP3 expression status in SCLC, GSEA software V.4.1.0 was applied.25 Based on FOXP3 level, mRNA expression dataset in the GEO was divided into high and low expression groups. The number of gene set permutations was 1000 times, and phenotype label was set as ‘high expression versus low expression’. GSEA analysis according to correlation coefficient (CC) was also conducted. In our study, the absolute values of a normalized enrichment score (NES) of >1 and a false discovery rate (FDR) q value of <0.25 were considered as meaningful GSEA sets.

We used the method of CIBERSORT as a tool to analyze the immune landscape of patients with SCLC in the tumor microenvironment (TME). On the bases of linear support vector regression and deconvolution, CIBERSORT is an online database to analyze immune infiltration of tumor tissues through 22 immune-cell phenotypes.26 Leukocyte signature matrix (LM22), which contains source data and gene expression matrix, includes 547 genes distinguishing 22 hematopoietic cell types of human: naïve and memory B cells, plasma cells, seven T-cell phenotypes, myeloid subsets, and NK cells. We divided 23 SCLC samples into low-immune risk and high-immune risk groups, which was on the basis of the expression levels of FOXP3, PD-L1, and CD8. After that, we investigated the correlations among distinct immune cells and revealed the heterogeneity of tumor-infiltrating immune cells in theses SCLC samples between two immune risk groups.

The correlation between different clinical factors and FOXP3 expression level was appraised by Chi-square tests. By Spearman rank correlation test, we also investigated the correlations of FOXP3 levels with other immune biomarkers. Then, considering distinct clinical characteristics, we conducted univariate as well as multivariate logistic regression analyses to predict FOXP3 expression in patients with SCLC . In order to evaluate prognosis condition, we also used Kaplan-Meier and Cox regression method for survival analysis. Pearson correlation test was used for continuous variables. p<0.05 was defined as statistical significance. Data analysis and visualization were conducted based on statistical tool SPSS V.26.0, GraphPad Prism V.7, and R Programming Language (R V.3.6.1).

One hundred and two subjects with SCLC were included (online supplemental table 1). The baseline information of clinical characteristics among these patients with SCLC were summarized (online supplemental table 2). The median age was 63.5 years (lower limit:38; upper limit: 81). Most subjects were male (84, 82.4%) and 18 (17.6%) were female. There are slightly more non-smokers (58, 56.9%) than smokers (44, 43.1%). All patients enrolled were in clinical staging I–III, and 98 patients (96.1%) did not develop tumor metastasis.

We performed IHC and pathological test and revealed the expression of biomarkers on both TILs and SCLC cells (figure 1). In this study, FOXP3 expression was tested positive in 88 samples on TILs (figure 1) (86.3%, ≥1% staining), while 14 (13.7%) did not. The positive expression of PD-1 was 41 (40.2%) on TILs and 0 (0.0%) on tumor cells. Two patients (2.0%) tested positive for the expression of PD-L1 on SCLC cells and 38 (37.3%) tested positive for the expression of PD-L1 on TILs. Expression of FOXP3 on TILs was significantly associated with clinical staging (p=0.013). We observed no correlation between FOXP3 expression on TILs and clinical data including age (p=0.320), gender (p=0.689), smoking status (p=0.077), metastasis (p=0.504) and chemotherapy (p=0.360, online supplemental table 3).

The connection of FOXP3 levels and immune checkpoints or other immune markers was analyzed (online supplemental table 4). Interestingly, we found that FOXP3 expression on TILs was closely associated with checkpoints including PD-1 on TILs (CC=0.327, p=0.001), as well as PD-L1 on TILs (CC=0.307, p=0.002). Furthermore, FOXP3 level was also positively correlated with CD3 (CC=0.366, p<0.001), CD4 (CC=0.307, p=0.002), and CD8 levels (CC=0.215, p=0.030). Regretfully, FOXP3 levels showed no correlation with PD-L1 on SCLC cells (p=0.573). We failed to investigate the correlation between FOXP3 expression and PD-1 levels on tumor cells for the limited samples.

We performed both univariate and multivariate logistic analysis for FOXP3 expression. After adjusting relevant parameters, the analysis results were summarized in online supplemental table 5. The ORs for FOXP3 expression on TILs were 4.375 (95% CIs 1.268 to 15.091, p=0.019) for patients at stages I and II compared with those at stage III, and 0.051 (95% CI 0.006 to 0.406, p=0.005) when comparing specimens with positive CD3 on TILs with those with negative CD3.

In all subjects involved, the median RFS of 102 patients was 32.000 months (95% CI 12.106 to 51.894), and 56/102 (54.9%) patients had relapses. For 60 patients (60/102, 58.8%) with stage I–II SCLC, 25 (25/102, 41.7%) relapsed (median RFS: 63.000 months, 95% CI 26.271 to 99.729). For 42 patients (42/102, 41.2%) with stage III SCLC, 31 (73.8%) had the end event (median RFS: 14.7000 months, 95% CI 10.360 to 19.040), which showed a significantly lower RFS compared with patients with stage I–II SCLC (p=0.004, online supplemental table 6). We then investigated the correlation between FOXP3 levels and RFS status in SCLC. Interestingly, FOXP3 level on TILs was significantly correlated with RFS status in SCLC, and the positive group showed longer RFS compared with the negative group (41.200 months, 95% CI 26.937 to 55.463, vs 14.000 months, 95% CI 8.133 to 19.867; p=0.008; figure 2A).

We then combined diverse expression patterns based on FOXP3 levels and performed subgroup analysis (figure 2B–D and online supplemental figure 1). Importantly, for immune checkpoints expressed on TILs, FOXP3 and PD-L1 double positive (vs either FOXP3 or PD-L1 positive, or FOXP3 and PD-L1 double negative; 41.300 months, 95% CI 41.076 to 41.524, vs 16.200 months, 95% CI 0.000 to 38.336, vs 14.000 months, 95% CI 8.133 to 19.867; p=0.003), FOXP3 and PD-1 double positive (vs either FOXP3 or PD-1 positive, or FOXP3 and PD-1 double negative; 45.000 months, 95% CI undefined, vs 32.000 months, 95% CI 0.926 to 63.074, vs 14.000 months, 95% CI 8.133 to 19.867; p=0.017), and FOXP3 and LAG-3 double positive (vs either FOXP3 or LAG-3 positive, or FOXP3 and LAG-3 double negative; 45.000 months, 95% CI 27.333 to 62.667, vs 18.000 months, 95% CI 0.000 to 44.807, vs 11.300 months, 95% CI 6.208 to 16.392; p=0.021) showed a statistical correlation with longer RFS. Besides, for other immune biomarkers, FOXP3 and CD3 double positive (vs either FOXP3 or CD3 positive, or FOXP3 and CD3 double negative), FOXP3 and CD8 double positive (vs either FOXP3 or CD8 positive, or FOXP3 and CD8 double negative), and FOXP3 and CD4 double positive (vs either FOXP3 or CD4 positive, or FOXP3 and CD4 double negative) also showed higher survival with statistical significance (p=0.005, p=0.002, and p=0.001). For checkpoints expressed on tumor cells, the subgroup analysis of combing FOXP3 on TILs as well as PD-L1 level on SCLC cells revealed a significant prognostic value as well (p=0.017). However, the sample size of checkpoint expression in tumor cells was too limited. Further investigations are required.

The univariate analysis suggested that clinical stage (p=0.006; HR=0.474, 95% CI 0.279 to 0.804) and FOXP3 levels on TILs (p=0.011; HR=2.352, 95% CI 1.221 to 4.531) were predictive factors for RFS in SCLC. By multivariate COX regression, we found that clinical stage (I–II vs III–IV; p=0.016, HR=0.516, 95% CI 0.301 to 0.886) and FOXP3 on TILs (negative vs positive; p=0.041, HR=2.008, 95% CI 1.027 to 3.927) were considered as independently predictive factors (table 1).

The subgroup analysis revealed great significance of FOXP3 levels on TILs in combination with other immune markers to prognosis evaluation in SCLC. Thus, we hypothesized that FOXP3 levels on TILs connected closely with other immune molecules. We used XGBoost algorithm for importance assessment and feature selection to validate this assumption. By repeating 1000 times considering all immune parameters, the importance feature map demonstrated the first rank of FOXP3 on TILs, followed by PD-L1 on TILs and CD8 (figure 3A). To validate the model performance, we performed ROC curve analysis in the calibration set. The machine learning model by XGBoost matched the actual one well, with an area under the curve (AUC) of 0.866 (figure 3B). Further, integrating the first three biomarkers also showed well-predictive ability (AUC=0.655, figure 3B).

Given the importance results of the XGBoost model, FOXP3, PD-L1 on TILs, and CD8 were selected as variables for construction of immune risk model in SCLC. As shown in the survival diagram by Kaplan-Meier analysis, patients with SCLC with high-risk scores (52/102, 51.0%) had poorer prognosis (vs patients with SCLC with low-risk scores; 14.700 months, 95% CI 10.107 to 19.293, vs 41.300 months, 95% CI 35.773 to 46.827; p=0.012; figure 3C). We further conducted univariate Cox regression, and the regression result revealed a statistically predictive role of immune risk score in RFS of SCLC (HR=0.505, 95% CI 0.293 to 0.872; p=0.014). Given that FOXP3 on TILs was incorporated in the construction of immune risk score, a multivariate Cox regression model of RFS was built based on SCLC staging and immune risk score. Both SCLC staging (HR=0.478, 95% CI 0.281 to 0.814; p=0.007) and immune risk score (HR=0.512, 95% CI 0.296 to 0.884; p=0.016) were independently and significantly correlated with RFS in SCLC (table 1).

To further confirm the prognostic role of immune risk score in SCLC, we compared the predictive performance between immune risk score and single indicators at first (figure 3D). The AUC values of time-dependent ROC analyses showed this immune risk score system performed better than single biomarkers including FOXP3, PD-L1 on TILs, and CD8 (figure 3D). Also, considering the significant survival differences of subgroup analysis based on FOXP3 and TIL PD-L1, TIL PD-1, TIL LAG-3, CD3, CD4, CD8, or PD-L1 on tumor cells, we established seven unique immune risk scores using multivariate Cox analysis and compared their performance with XGBoost-based risk score. All combined immune systems showed good performance (figure 3E). Among distinct combined biomarkers, the XGBoost-based immune risk score obtained the best AUC value of 0.715, which performed better than either biomarkers analyzed previously in predicting relapse of SCLC at stages I–III (figure 3E).

By integrating immune risk score and SCLC staging, we used nomogram algorithm to predict the probability of RFS of 1, 3, and 5 years in SCLC using training dataset. The nomogram indicated the immune risk score contributed slightly less to patients’ prognosis compared with clinical stage (figure 4A). Calibration plots and C-index were conducted for the probability of RFS of 1, 3, and 5 years, and the C-index was 0.639 (figure 4B–D). Furthermore, we assessed and validated its effectiveness by time-dependent ROC analysis. The AUC values for RFS of 1, 3, and 5 years were 0.656, 0.737, and 0.698, respectively, in the training cohort, and 0.608, 0.608, 0.714, respectively, in the validation cohort (figure 4E,F), which highlighted that the predictive model performed well in predicting relapse of patients with SCLC in stages I–III.

Based on predictive scores, the cut-off values were determined by regrouping all subjects in the entire SCLC cohort, as well as the training and validation cohorts into three subgroups with distinct prognosis. Survival analyses showed that in the entire SCLC, training, and validation cohorts, the low-risk group had the highest 1-year RFS at 81.1%, 86.4%, and 67.6%, respectively, followed by the moderate-risk group at 67.0%, 75.5%, and 48.6%, respectively. The high-risk group suggested the lowest 1-year RFS for the entire, training, and validation cohorts: 44.4%, 52.4%, and 27.2%, respectively (figure 4G–I). Further, the low-risk groups in all cohorts were correlated with higher RFS compared with moderate-risk and high-risk groups (p<0.001 in the entire SCLC cohort, p=0.002 in the training cohort, and p=0.126 in the validation cohort, respectively).

We used cBioportal Database27 to retrieve SCLC clinical datasets and downloaded the suitable one with 81 samples. After screening gene panels, RNA-seq data of patients with SCLC were available for nine samples (online supplemental table 7). Then, we investigated the coexpression correlation between protein FOXP3 and other immune biomarkers (online supplemental figure 2A). The results showed some similar correlations. Significant significance was analyzed in the correlation of FOXP3 expression with the expression of PD-1, LAG-3, CD3, and CD4 (p=0.001, p=0.014, p=0.002, and p=0.015, respectively) (online supplemental table 8). Among them, PD-1 expression showed the highest relevance (CC=0.894). In addition, the results demonstrated no significant correlation of FOXP3 expression with the expression of PD-L1 (p=0.584) and CD8 (p=0.977). For the relationship between FOXP3 and clinical data, the mRNA expression of FOXP3 had no statistical association with diagnosis age (p=0.206) as well as gender (p=0.379). Further, we identified the relationship between FOXP3 expression and patients’ prognosis in SCLC (online supplemental figure 2B). The survival analysis showed that the group with high FOXP3 expression had longer survival compared with the group with low FOXP3 expression (17.000 months, 95% CI 2.300 to 31.700, vs 10.000 months, 95% CI 1.180 to 18.820; p=0.076). We also extended the risk model to the prognosis evaluation of these patients with SCLC (online supplemental figure 2C). The survival results indicated that patients with SCLC with high risk showed shorter OS compared with the low-risk and moderate-risk groups (10.000 months vs 17.000 months, p=0.603).

For better understanding of the relative expression of mRNA FOXP3, CD274 (PD-L1), and CD8A (CD8) in SCLC, we verified their levels, respectively, in an SCLC cohort. After retrieving gene expression of SCLC, the GSE43346 dataset that included 23 SCLC samples and 42 normal tissues was chosen available for analysis. For FOXP3 expression, significant difference between SCLC tissues and normal lung controls (p=0.0061 in GSE43346 dataset) was calculated (online supplemental figure 3A). For PD-L1 (CD274), the GEO cohort also suggested a significant difference of expression between SCLC tissues and control subjects (p=0.034 in GSE43346 dataset) (online supplemental figure 3B). For CD8, there was no significant difference shown in GSE43346 dataset (p=0.8; online supplemental figure 3C). In particular, the expression of FOXP3 in SCLC tissues was lower than that of the normal group; and PD-L1 expression in SCLC tissues was higher than that of the controls.

To investigate relevant biological pathways and potential regulatory genes related to FOXP3 expression in SCLC, we further performed GSEA in the GSE43346 dataset to excavate putative targets related to FOXP3. The whole cohort was divided into high and low FOXP3 expression groups according to the median value of FOXP3 expression. In 174 gene sets that were upregulated or downregulated between the FOXP3 high and low expression groups, 108 gene sets upregulated in the SCLC group with high FOXP3 expression (108/174, 62.1%) were analyzed, while 66 gene sets upregulated in the group with low FOXP3 expression (66/174, 37.9%) were analyzed. High FOXP3 expression-related genes were enriched in several immune-related pathways such as intestinal immune network for IgA production (|NES|=2.300, FDR q value<0.001), primary immunodeficiency pathway (|NES|=2.267, FDR q value<0.001), and cytokine–cytokine receptor interaction (|NES|=2.158, FDR q value<0.001). The top three FOXP3 high expression-related pathways were depicted, with the absolute value of NES >1 and FDR q value<0.25 (figure 5A). Figure 5B demonstrated a high overlapping rate and close correlations among the top three high FOXP3 expression correlated pathways. Specifically, through leading edge analysis, a total of 17 genes were found important in at least two pathways (17/53, 32.1%), and CD40LG overlapped in all of these pathways. After verifying expression levels of the aforementioned overlapped genes, four showed significant differences between SCLC tissues and normal specimens (p<0.05) (figure 5C). Further, by Benjamini and Hochberg multiple testing to correct p value, inducible co-stimulator (ICOS), with the absolute value of log2-fold change of >1, was found differentially expressed between SCLC and normal tissues. By t-test, overexpression of CD28, CD40LG, HLA-DMA, and interleukin (IL)-2 was observed in high FOXP3 expression group (all p<0.05, figure 5D). In addition, we conducted a correlation analysis and found that FOXP3 was positively associated with IL-2 (p=0.011) and CD28 (p=0.026) (figure 5E). Two significantly enriched pathways were observed in the low expression cohort. Detailed enrichment profiles and related genes were also depicted (online supplemental figure 4).

Besides, based on CC for GSEA analysis, genes negatively correlated with FOXP3 were also significantly enriched in another three pathways ‘KEGG_PARKINSONS_DISEASE’, ‘KEGG_HUNTINGTONS_DISEASE’, and ‘KEGG_OXIDATIVE_PHOSPHORYLATION’ (online supplemental figure 5). Among all overlapped genes that presented in more than two pathways, expression of cytochrome c oxidase subunit 8A (COX8A) suggested the highest negative correlation with FOXP3 level, indicating it might play an important role in the regulation of FOXP3 (CC=−0.648, p=0.001; online supplemental figure 5). For genes positively correlated with FOXP3, overlapped genes among the top four pathways only presented low to moderate correlations with FOXP3 (online supplemental figure 6).

To figure out tumor-infiltrating condition of patients with SCLC in two immune risk groups, we further performed CIBERSORT using LM22.26 In the SCLC GEO dataset, two heatmaps were first performed with immune features between these two immune risk groups of high or low level, which showed the difference in the proportion of 22 diverse immune cells in different samples (figure 6A,B). After analysis, we summarized the correlations among these immune cells in TME, respectively, in both groups. The interactions vary between different cells. For the high-immune risk group, Treg cells indicated a strong correlation with monocyte (CC=0.820, p=0.002) and activated mast cells (CC=0.980, p<0.001), while a relatively weak correlation was shown between naïve B cells and neutrophils (CC=−0.326, p=0.328; figure 6C). For the low-immune risk group, naïve B cells showed an extremely high correlation with monocytes (CC=0.942, p<0.001), while its relationship with other cell types was quite modest (figure 6D). Further, immune-cell proportions were explored between high-immune and low-immune risk groups (figure 6E–H). By incorporating the expression levels of the three predictive biomarkers, significant differences were found in activated CD4⁺ memory T cells (p=0.014) and plasma cells (p=0.049), and marginally marked differences in gamma delta T cells (γδ T cells) (p=0.054) and resting dendritic cells (p=0.053). Specifically, three immune cells including activated CD4⁺ memory T cells, γδ T cells, and plasma cells were higher in the low-immune risk group compared with those in the high-score group. Conversely, in comparison with low-risk group, resting dendritic cell level was suggested higher in the high-immune risk group. Therefore, the heterogeneity of tumor-infiltrating immune cells might represent a promising feature for risk prediction in SCLC. In return, these meaning findings also validated the crucial role of immune risk scores in TME.