Significantly different immunological score in lung adenocarcinoma and squamous cell carcinoma and a proposal for a new immune staging system

ABSTRACT

TNM stage is not enough to accurately predict the prognosis of patients with non-small cell lung cancer (NSCLC). This study aimed to establish the immunological score (IS) in lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), separately, and propose a new staging system in NSCLC. We used the multiplex fluorescent immunohistochemistry (mIHC) technology to detect 17 immune biomarkers of 304 patients with NSCLC. The LASSO-COX regression model was used to establish the IS_NSCLC in the training cohorts. The IS_NSCLC was then validated in the validation cohort. The constructed IS_LUAD contained three immune features: CD4⁺CD73⁺_{core of tumor (CT)}, PD-L1⁺_CT, and IDO⁺_{invasive margin (IM)}. IS_LUSC also contained two immune features: CD8⁺CD39⁻CD73⁻_CT, CD8⁺Tim-3⁺_IM. In the training cohort, significant prognostic differences were found upon comparing low-IS_NSCLC patients with high-IS_NSCLC patients. For LUAD, the 5-y disease-free survival (DFS) rates were 54.7% vs. 8.1% and the 5-y overall survival (OS) rates were 82.4% vs. 36% (all P< .0001). For LUSC, the 5-y DFS rates were 74.0% vs. 14.7% and the 5-y OS rates were 78.2% vs. 17.6% (all P< .0001). Multivariate analyses indicated that IS_NSCLC was an independent indicator for prognosis. Finally, we combined IS_NSCLC with clinicopathological factors to establish a TN-I staging system and two nomogram models for clinical use. The TN-I stage had better prediction accuracy than TNM stage. The newly established IS_LUAD and IS_LUSC were completely different, and both were excellent indicators for the prognostic prediction. The TN-I stage could effectively improve prognostic accuracy and facilitate clinical application.

Abbreviations

NSCLC, non-small cell lung cancer; IS, immunological score; mIHC, multiplex fluorescent immunohistochemistry; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; CT, core of tumor; IM, invasive margin; DFS, disease-free survival; OS, overall survival; SITC, the Society for Immunotherapy of Cancer; FFPE, formalin-fixed paraffin-embedded; MWT, microwave treatment; DCA, decision curve analysis; ROC, receiver operating characteristic; AUC, area under the curve; EGFR, epidermal growth factor receptor.

KEYWORDS:

• Non-small cell lung cancer
• lung adenocarcinoma
• lung squamous cell carcinoma
• immunological score
• prognosis
• staging

Introduction

Non-small cell lung cancer (NSCLC) is one of the deadliest human malignant tumors and the leading cause of global death from cancers.^1,2 The two predominant histological phenotypes of NSCLC are lung adenocarcinoma (LUAD; ~50%) and lung squamous cell carcinoma (LUSC; ~40%).^3,4 Typically, LUADs arise from distal airways, whereas LUSCs are associated with the more proximal airways. LUSCs usually have a stronger correlation with smoking and inflammatory diseases compared with LUADs, ^3,4 and should be regarded as distinctly different diseases from the aspect of biopathology, therapeutic stratification, and therapeutic effect.⁵

The current tumor-node-metastasis (TNM) staging system is the most acknowledged clinicopathological factor for prognostication and therapy determination of NSCLC. However, the survival outcomes of patients with an equal stage and similar treatment options vary widely.⁶ In some advanced patients, the cancer can remain stable for a long time, and a small number of patients even have spontaneous regression of metastatic tumors. Conversely, even after undergoing R0 resection and without evidence of tumor residual or metastasis, about 25% of patients with stage I/II tumors experienced rapid tumor progression or death.⁶ These phenomena suggest that the information provided by the NSCLC TNM staging system is insufficient for prognostication. Therefore, it is necessary to adopt new strategies for improved survival prediction and treatment stratification. A retrospective study established a three-class risk model based on clinicopathological factors (age, T stage, nodes, and grading) and its nomogram in resected LUSC, which had a good prognostic accuracy (0.67 for DFS), provided a practical tool for evaluating the prognosis of patients with LUSC.⁷ Next, the researchers conducted an external multicenter validation which confirmed the wide applicability and reliability of this prognostic model, and found that adjuvant/neoadjuvant treatment seems to be more beneficial for intermediate/high-risk patients.⁸ In addition, it has been recognized that tumor-infiltrating immune cells play a pivotal role in tumor progression depending on the density and location of various immune cell subpopulations.^9,10 The heterogeneity of immune features may be the reason for the difference in the prognosis of patients within the same TNM stage. Immunological score (IS) indicates the abundance of immune infiltration on the basis of the total number of lymphocytes in the core of tumors (CT) and invasive tumor margin (IM). In colorectal cancer, IS was combined with the standard pathological TNM stage and was named as TNM-Immune, which was a more specific prognostic indicator.^11–13 To further validate the performance of IS on patients with colorectal cancer and prompt its implementation into clinical practice, an international coalition has been established, proposed by the Society for Immunotherapy of Cancer (SITC).^13–15 Numerous studies of the NSCLC tumor microenvironment have proved the important role of immune factors (such as CD8, CD3, CD4, mDC, and B cells, etc.) in predicting the prognosis of lung cancer;^16–21 and the researchers did a lot of attempts to use those factors to establish the IS in NSCLC.^22–24 With the emergence and development of multiplex immunohistochemical staining techniques, more comprehensive information about immune cell subgroups can be obtained. Therefore, in this study, we used the multiplex immunohistochemical techniques to try to construct a more refined and personalized new IS system in NSCLC. The least absolute shrinkage and selection operator method (LASSO) has become even more popular in selecting significant variables from high-dimensional markers. LASSO-COX proportional hazards (LASSO-COX) model is also broadly used in survival analysis when the data are complex and multidimensional.^25–28

In this study, we used the LASSO-COX model to construct a novel IS system named IS_NSCLC on the basis of 112 immune features, covering most of the immune cell subsets, for the prediction of postoperative recurrences and clinical outcomes of LUAD (IS_LUAD) and LUSC (IS_LUSC). We further combined the IS_NSCLC with TNM to form a TN-immunological score (TN-I) staging system for better risk stratification of NSCLC. At last, we established two nomograms to increase the clinical utility of the IS_NSCLC.

Materials and methods

Study population and tumor specimens

Data are collected from consecutive patients from Tianjin Medical University Cancer Institute and Hospital, Tianjin, China, with pathologically confirmed primary NSCLC diagnosed from February 2013 to December 2014. The median follow-up time is 52.5 months. Inclusion criteria were presence of invasive tumor components in the hematoxylin and eosin slides, availability of clinicopathological data, complete follow-up data, and no history of cancer treatment before the surgery. Before the surgery, all the patients had undergone chest CT examination, and pathological diagnosis of lung cancer was confirmed by biting biopsy test. Some patients underwent PET-CT examination for auxiliary examination. Most patients received lobectomy, some received pneumonectomy, wedge, and segmentectomy were less, and the specific numbers are listed in Supplementary Tables S3 and S4. According to the 2013 and 2014 NCCN Clinical Practice Guidelines in NSCLC, all patients underwent radical surgery with/without postoperative adjuvant treatment. Adjuvant therapy is mainly platinum-based chemotherapy, combined with radiotherapy or targeted therapy when necessary. Formalin-fixed paraffin-embedded (FFPE) blocks of 304 patients with resected stage I, II, and III NSCLC were collected. Stage IV patients were excluded because of their heterogeneity and poor representation. Among them, 204 patients were diagnosed with LUAD, and 100 patients were diagnosed with LUSC. In LUAD and LUSC, we randomly divide all patients into two cohorts, one cohort as the training cohort (LUAD N = 102, LUSC N = 50) and the other cohort as the validation cohort (LUAD N = 102, LUSC N = 50) (Supplementary Fig. S1). This retrospective study carried out according to the Helsinki Declaration and was approved by the Ethics Committee of Tianjin Medical University Cancer Institute and Hospital. Appropriate informed consent was received from all the patients.

Multiplex immunohistochemistry and multispectral imaging

According to previous studies, ^{9–11,13,22,23,29–34} we selected 17 markers and divided them into three PANELs for seven-color multiplex immunohistochemistry (mIHC) staining: CD8, Foxp3, CD4, CD69, CD39, and CD73 in PANEL-1; programmed death ligand-1 (PD-L1), CD68, CD163, CD11c, CD20, and indoleamine 2,3-dioxygenase (IDO) in PANEL-2; and CD8, CD66b, programmed death-1 (PD-1), lymphocyte activation gene 3 (Lag-3), T cells immunoglobulin domain and mucin domain protein-3 (Tim-3), and T cell Ig and ITIM domain (TIGIT) in PANEL-3 (Supplementary Fig. S2A-D). Detailed information is shown in Supplementary Table S1. The representative graphs of the high/low-level distributed positive cells in the CT/IM areas for each marker are shown in Supplementary Fig. S2A-D. The single-stained cell and multiple-stained cell densities were calculated, respectively, in the CT and IM areas of the NSCLC tissue. Common labeling combinations form a total of 112 immune features (Supplementary Table S2). Slides were stained with multiplex fluorescence by using the PerkinElmer Opal 7-Color technology Kit (NEL81001KT) containing seven fluorophores. Xylene was used for dewaxing, and ethanol was used for rehydration. Microwave treatment (MWT) was performed for antigen retrieval with EDTA Buffer (pH 9.0). At room temperature, the tissue was blocked using an antibody-blocking buffer for up to 10 min. The slides were then incubated overnight with primary antibodies in a 4°C refrigerator. Next day, incubation with secondary antibody (poly-HRP-Ms/Rb) for 10 min was performed at room temperature. TSA Visualization and signal amplification were performed by using Opal TSA Plus (1:100). Thereafter, the Ab-TSA complex was removed by heating through MWT with EDTA 9.0 buffer. Each multiplex staining was performed by repeating these staining steps in series, with MWT step at the end of each cycle. Finally, all 17 targets of interest were detected with a corresponding Opal fluorophore (Supplementary Table S1). At last, multiplex staining was finished with the cell nuclei staining with DAPI and the slides were covered by anti-fluorescence-quenching sealing agent and coverslips.

Slides were scanned and visualized using the PerkinElmer Mantra Quantitative Pathology Imaging System at 200× magnification with the same exposure times. Multispectral images unmixing was performed using PerkinElmer inForm Image Analysis software (version 2.4.0). In CT and IM areas, five random areas of tumor tissue without hemorrhage, necrosis, or detachment were selected and scanned. The cell density was determined by the average count of positive cells per mm².

Cutoff score selection and IS_NSCLC construction

X-tile (Yale University, version 3.6.1) was used to determine the optimum density cutoff of each single-stained or multiple-stained feature based on the patients’ DFS. Next, we used the log-rank test P-value of Kaplan–Meier analysis for DFS to identify 20 immune features with top 20 smallest P-values to construct the prognostic model in LUAD and LUSC.

In the training cohort, the most potent prognostic markers were selected from the 20 chosen immune features using the LASSO-COX regression model. Finally, a scoring system based on multi-immune features was constructed for survival predicting.^25–28 LASSO-COX regression model analysis was performed using the package “glmnet” of R software (version 3.5.3).

Statistics

The COX regression model was utilized to perform the univariate and multivariate analyses. Chi-square tests were used to compare two groups of categorical variables. P< .05 indicated statistical significance. DFS and OS estimation in different groups were performed by the Kaplan–Meier method with the log-rank test. Nomograms were generated by the logistic regression coefficients. We put all the important clinicopathological factors into the multivariate logistic regression, which adopts a stepwise backward method with an inclusion criterion of P < .05 and an exclusion criterion of P > .1. The finally screened factors and TNM stages were used to construct the nomograms. To evaluate the performance characteristics and assess the clinical practicality of the nomograms, respective calibration plots, and decision curve analysis (DCA) were drawn. Time-dependent receiver operating characteristic (ROC) curves were drawn to show the accuracy of different prognostic markers. Nomograms, DCA, and calibration plots were generated using EmpowerStats (version 2.0, http://www.empowerstats.com). Heatmap was performed using the package “pheatmap” of R software. Other calculations were performed using SPSS Statistics (version 24.0, IBM) and R software (version 3.5.3).

Results

Patient characteristics and IS_NSCLC construction

Total 304 patients were enrolled in this study. The clinicopathological characteristics of the LUAD/LUSC patients in the training/validation cohorts are listed in Supplementary Tables S3 and S4. In this study, the volume of the tumor was calculated by the formula: (π × length × width × height)/6. For LUAD, according to postoperative pathology reports, micropapillary and solid patterns were marked as poorly differentiated; acinar or papillary patterns were marked as moderately differentiated; lepidic patterns were regarded as well differentiated. For LUSC, the postoperative pathology report directly reports the differentiation status (well, moderately, and poorly differentiated). Supplementary Fig. S3 shows the heatmaps of all markers for each patient from the LUAD and the LUSC group.

In the training cohort, X-tile was used to determine the optimum cutoffs (Supplementary Fig. S4). Supplementary Table S2 lists the cutoff values and the Kaplan–Meier analysis results for DFS of all the 112 immune features in the LUAD/LUSC cohorts. If the feature’s density was greater than its cutoff, the status of this feature was equivalent to 1; otherwise, the status was equivalent to 0. Based on the Kaplan–Meier analysis and its P-value, we chose 20 features with the top 20 smallest P-values in LUAD and LUSC for further model construction. Next, in the training groups, the LASSO-COX regression model was used to establish the prognostic scoring system named IS_NSCLC based on the status of the 20 chosen features. We first screened some closely related prognostic indicators through LASSO-COX regression analysis, and obtained the corresponding regression coefficients of these indicators. Finally, through the calculation and deformation of the COX regression model, the IS_NSCLC we defined is equal to the sum of the product of each marker value multiplied by their coefficients, which can effectively predict the prognostic risk. The IS_NSCLC was designated as IS_LUAD and IS_LUSC for LUAD and LUSC, respectively. IS_LUAD included three features: CD4⁺CD73⁺_CT, PD-L1⁺_CT, and IDO⁺_IM. IS_LUSC included another two features: CD8⁺CD39⁻CD73⁻_CT, CD8⁺TIM-3⁺_IM (Figure 1). Supplementary Fig. S5 shows the COX regression risk scores of the raw data for the indicators included in the IS_LUAD/IS_LUSC. It also shows the distribution of these indicators among patients and their relationship to prognosis. The risk score plots suggested that, in LUAD, CD4⁺CD73⁺_CT was a favorable prognostic factor, whereas PD-L1⁺_CT and IDO⁺_IM were negative factors. In LUSC, CD8⁺CD39⁻CD73⁻_CT was a protective factor, whereas CD8⁺TIM-3⁺_IM was a negative factor for the prognosis. Two formulas were also constructed to obtain the score for each patient on the basis of the expression levels of the above-selected features, where IS_LUAD = – (1.372 × CD4⁺CD73⁺_CT status) + (1.499 × PD-L1⁺_CT status) + (1.014 × IDO⁺_IM status) and IS_LUSC = – (1.641 × CD8⁺CD39⁻CD73⁻_CT status) + (1.319 × CD8⁺TIM-3⁺_IM status). Based on X-tile, we chose 0 as the IS_NSCLC cutoff in both the LUAD and LUSC cohorts to classify the patients into a high-IS_NSCLC or a low-IS_NSCLC group (Supplementary Fig. S4).

Figure 1. Construction of the IS_NSCLC by the LASSO model. LASSO coefficient profiles of 20 significant immune features with the top 20 smallest P-values in LUAD training cohort (a) and LUSC training cohort (d). For tuning parameter selection in the LASSO model, by 10-time cross-validation via 1-SE criteria, λ = 0.194 was chosen in LUAD training cohort (b), and λ = 0.283 was chosen in LUSC training cohort (e). Representative multiplex fluorescent immunohistochemistry images show the final features enrolled in IS_LUAD (c) and in IS_LUSC (f). Bar, 100 µm. IS, immunological score; NSCLC, non-small cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma; CT, the core of tumor; IM, invasive margin.

The distribution of the clinicopathological characteristics had no significant differences between the high-IS_NSCLC and low-IS_NSCLC groups in each cohort (Supplementary Tables S3 and S4). At different follow-up times, time-dependent ROC analyses were carried out to assess the prognostic accuracy of the IS_NSCLC in the training/validation cohort, which showed large area under the curve (AUC) values (Figure 2). Within the LUAD training cohort, upon comparing the low-IS_LUAD patients with the high-IS_LUAD patients, the 5-y DFS rates were 54.7% vs. 8.1%, and the 5-y OS rates were 82.4% vs. 36% (all P< .0001; Figure 2(a)). Within the LUSC training cohort, upon comparing the low-IS_LUSC patients with the high-IS_LUSC patients, the 5-y DFS rates were 74.0% vs. 14.7%, and the 5-y OS rates were 78.2% vs. 17.6% (all P< .0001; Figure 2(c)). We further applied the IS_NSCLC scoring system to the LUAD/LUSC validation group, and the results were found consistent (Figure 2(b,d)), which confirmed that the IS_NSCLC has excellent prognostic power in various populations.

Figure 2. IS_NSCLC performance in time-dependent ROC curves and Kaplan–Meier survival analyses in different cohorts. (a) LUAD training cohort. (b) LUAD validation cohort. (c) LUSC training cohort. (d) LUSC validation cohort. AUCs at 1, 3, and 5 y were used to assess prognostic accuracy. P-values were calculated by the log-rank test. DFS, disease-free survival; OS, overall survival; ROC, receiver operator characteristic; AUC, area under the curve.

Univariate, multivariate analysis, and stratified analysis of IS_NSCLC

The univariate association of IS_NSCLC and clinicopathological characteristics with DFS or OS in the training/validation cohort was tested (Supplementary Table S5). Significant variables were then analyzed by multivariate Cox proportional hazard regression models. After adjusting clinicopathological variables, the multivariate analysis suggested that the IS_NSCLC was an independent negative prognostic indicator of DFS/OS in the LUAD/LUSC cohorts (Table 1). In addition, in some cohorts, T stage, N stage, TNM stage, and tumor volume are also independent prognostic factors; sometimes their HR are more accurate to discriminate between patients for survival (Table 1). This is consistent with the importance of these clinicopathological features in tumor prognosis.

Table 1. Multivariable Cox regression analyses of the IS_NSCLC, clinicopathological characteristics, and survival.

LUAD cohort			LUSC cohort
Variable	HR (95% CI)	P	Variable	HR (95% CI)	P
Disease-free survival			Disease-free survival
Training Cohort (n= 102)			Training Cohort (n= 50)
ISLUAD	2.766 (2.154, 3.550)	<0.001	ISLUSC	2.866 (1.898, 4.329)	<0.001
T stage (T3+ T4 vs. T1+ T2)	2.328 (1.139, 4.759)	0.021	N stage (N1+ N2+ N3 vs. N0)	2.947 (1.234, 7.039)	0.015
Validation Cohort (n= 102)			Validation Cohort (n= 50)
ISLUAD	1.952 (1.501, 2.540)	<0.001	ISLUSC	3.638 (2.190, 6.044)	<0.001
TNM stage (III vs. I+ II)	1.353 (1.059, 1.728)	0.016
Tumor volume (≥10 cm^3 vs. <10 cm^3)	1.910 (1.172, 3.112)	0.009
Overall survival			Overall survival
Training Cohort (n= 102)			Training Cohort (n= 50)
ISLUAD	2.153 (1.642, 2.824)	<0.001	ISLUSC	2.575 (1.728, 3.836)	<0.001
Validation Cohort (n= 102)			Validation Cohort (n= 50)
ISLUAD	1.930 (1.402, 2.656)	<0.001	ISLUSC	2.502 (1.644, 3.805)	<0.001
T stage (T3+ T4 vs. T1+ T2)	2.930 (1.503, 5.711)	0.002

IS, immunological score; NSCLC, non-small cell lung cancer; LUAD, lung adenocarcinoma; LUSC, lung squamous cell carcinoma.

Stratified analyses were performed based on the TNM stage and other crucial clinicopathological risk factors in the LUAD/LUSC cohorts. In both LUAD and LUSC groups, regardless of stage I, II, or III, low-IS_NSCLC patients always had a prolonged DFS/OS compared with that of the high-IS_NSCLC patients (Figure 3). When stratified with other clinicopathological risk factors in LUAD and LUSC, the IS_NSCLC was still a significant prognostic marker for DFS or OS. It should be noted that the IS of patients with stage T3/T4 LUAD and patients with LUSC volume smaller than 10 cm³ did not always have prognostic significance. However, poor prognostic trends of the IS could be found in these groups, and the non-significance maybe a result of the small size of the sample. The subgroup Kaplan–Meier analyses are shown in Supplementary Figs. S6-9. The forest plots present the stratified COX regression analysis results (Supplementary Fig. S10).

Figure 3. Kaplan–Meier survival analysis stratified by the TNM stages. Kaplan–Meier survival analysis of DFS and OS for all the NSCLC patients according to the IS_NSCLC stratified by the TNM stages. P-values were calculated by the log-rank test. (a) Stage I LUAD. (b) Stage I LUSC. (c) Stage II LUAD. (d) Stage II LUSC. (e) Stage III LUAD. (f) Stage III LUSC.

The ROC analysis showed that the prognostic accuracy of the IS_NSCLC was relatively higher than that of the TNM stage, both in LUAD and LUSC (Figure 4(h,i)), revealing that the IS_NSCLC exhibited a superior prognostic value compared with the TNM stage. We also calculated the IMMUNOSCORE® (IMMUNOSCORE® is a registered trademark owned by INSERM), which is obtained by the quantification of tumor-infiltrating CD3/CD8 in the CT/IM areas. Low to high densities of CD3/CD8 cells in these two areas equal IS 0 (I0) to IS 4 (I4).¹³ We then performed Kaplan–Meier analyses on the IMMUNOSCORE® (Supplementary Fig. S11). Similar to previous studies,^13–15 patients with high scores tend to have better prognosis than patients with low scores, but sometimes the opposite result can occur. Both in LUAD and LUSC, the ROC analysis revealed that IS_NSCLC exhibited a better prognostic value than that of the IMMUNOSCORE® (Figure 4(h,i)).

Figure 4. From TNM to TN-I. DFS curves of LUAD/LUSC patients according to IS_NSCLC (a, b) and TNM stage (c, d). Patients were grouped into three groups (e, f) according to the combination of TNM stage (t, n) and IS_NSCLC (i) (T-N-I risk score) by survival (3-y DFS, <20%, 20%-70%, and >70%; 5-y OS, <30%, 30%-70%, and >70%), resulting into a TN-immunological score (TN-I) staging table (g), adding significant prognostic impact to each TNM stage. ROC curves show the comparisons of the area under the curve of TN-I stage, IS_NSCLC, and the IS_NSCLC classifier (high vs. low) with TNM stage and the IS proposed by SITC in LUAD cohort (h) and LUSC cohort (I). SITC, the Society for Immunotherapy of Cancer.

TN-I staging and nomograms construction

To provide a better prognostic classification of NSCLC, we supplemented our IS_NSCLC into TNM staging system. Low-IS_NSCLC equaled immune stage I0, meaning the early immune stage, while high-IS_NSCLC equaled immune stage I1, representing the advanced immune stage. As we did not have stage IV patients, we utilized multivariable COX regression analyses to obtain the risk score according to the T, N, I stage (T-N-I risk score). X-tile was used to find two optimal cutoff values to divide the risk score into three groups based on the DFS (Supplementary Fig. S12). Next, we used the risk score value to find the corresponding combinations of T, N, I, which were then divided into three prognostic groups, which were defined as the three stages of our new TN-I staging system. Although the cutoffs were 1.5/2.4 in LUAD while they were 1.0/2.4 in LUSC, the LUAD and LUSC cohorts obtained the same division of T, N, I combinations. The final TN-I staging system is shown in Figure 4. Both in LUAD and LUSC, the ROC analysis revealed that the TN-I stage exhibited a higher prognostic accuracy than that of the TNM stage (Figure 4(h,i)).

Finally, nomograms were constructed in both LUAD and LUSC as a quantitative method to predict the 3-y recurrence rate and the 5-y mortality rate for stage I–III LUAD/LUSC patients (Figure 5(a)). The nomograms integrated the IS_NSCLC and clinicopathological factors on the basis of the logistics regression model. We put all the important clinicopathological factors (age, gender, tumor volume, differentiation status, T stage, N stage, Smoking, Chemotherapy, TNM stage) and IS_NSCLC into the multivariate logistic regression, which adopts a stepwise backward method with an inclusion criterion of P < .05 and an exclusion criterion of P > .1. The finally screened factors (shown in Figure 5(a)) and TNM stage were used to construct the nomograms. The ROC curve analyses revealed that the nomograms had larger AUC value when compared with that of IS_NSCLC or TNM stage alone (Figures 4 and 5(b)). Calibration plots showed excellent agreement between the predicted and observed probabilities, indicating a superior performance of nomograms (Figure 5(c)). DCA was conducted to assess the nomograms’ net clinical benefit (Figure 5(d)).

Figure 5. Nomograms for predicting the 3-y recurrence and 5-y mortality rate after surgery in NSCLC patients. Nomograms (a) are showed at the top. Receiver operator characteristic curves (b) exhibit the prediction accuracy of each nomogram model. Model performance is shown by calibration plots (c) depicting the agreement between predicted and observed probabilities of each model. Decision curve analysis (d) shows the net clinical benefit of the nomograms.

Discussion

In our study, we constructed a new IS system named IS_NSCLC consisting of two different models. One was designated as IS_LUAD comprising three immune features for LUAD patients; the other was designated as IS_LUSC comprising two immune features for LUSC patients. The IS_NSCLC is a powerful prognostic tool independent of TNM staging, with high IS_NSCLC indicating potential risky patients. TNM stage was established according to the anatomical information. Patients with the same TNM stages have a significantly different prognosis, indicating that this stage classifier provided inadequate prognostic information.^14,15,35 Many efforts had been made to establish the prognostic model combining TNM stage with the clinicopathological factors (such as the three-class risk model established in LUSC) and obtain good results.^7,8 These models pay more attention to the characteristics of the tumor itself. On the contrary, the IS_NSCLC is providing biological and immunological information and could function as a powerful supplement to the TNM stage. Compared with the IS of NSCLC proposed by previous studies, ^13–15 our IS_NSCLC system discussed LUAD and LUSC separately, with complete consideration of the differences of immune characteristics between LUAD and LUSC. Besides, our model contained myeloid cells, checkpoints, and multi-labeled markers, and was established based on the LASSO-COX regression model, which significantly improved the accuracy of prediction.^25,26,28 Through classifying the patient into low-IS_NSCLC or high-IS_NSCLC group, the prognostic risk information is easily available. After stratified by the clinicopathological factors, IS_NSCLC was still a good predictor of prognosis. Our data also suggested that the IS_NSCLC seemed unsuitable for patients with stage T3/T4 LUAD and LUSC volume smaller than 10 cm³. However, considering the small number of patients in those groups, the results need further verification.

Multivariate analyses indicated that IS_NSCLC was an independent indicator for prognosis in all cohorts. However, besides IS_NSCLC, T stage, N stage, TNM stage, and tumor volume may also have important prognostic significance. Thus, we combined the IS_NSCLC with these clinicopathologic factors in two ways to better predict the prognosis: the establishment of TN-I staging system and the construction of nomogram, which improved the clinical practicality of the IS_NSCLC. TN-I staging system is a concise supplemental version of the TNM staging system, using IS staging consisting of “0” and “1” to suggest the degree of immunosuppression. Nomogram is a precise calculation to predict the possibility of 3-y recurrence and 5-y mortality. ROC analyses revealed that the prognostic efficacy of IS_NSCLC, nomograms, and the TN-I staging system was better than that of the IMMUNOSCORE® and the TNM staging system. The IS_NSCLC enhances the prognostic value of TNM stages and becomes a powerful supplement to TNM staging. Finally, based on the TN-I staging system, NSCLC patients in the same TNM stage could be classified into two distinct groups with different risks, which could help the decision-making in clinical practice. For patients with early TN-I stage, overtreatment should be avoided, where immunotherapies that function through enhancing existing immunity might have a better effect. For patients with advanced TN-I stage, we need to evaluate more effective and comprehensive approaches to enhance the effect of treatment and improve the prognosis based on the main immunosuppressive mechanism in this subtype of lung cancer. As immunotherapies such as PD-1/PD-L1 pathway blockades are playing a critical role in cancer therapy, ³⁶ further investigations are needed to explore whether the TN-I stage can predict the prognosis and responses of patients receiving immunotherapy.

In this study, for LUAD, our formula had three variables: IDO⁺_IM and PD-L1⁺_CT were associated with poor outcomes, while CD4⁺CD73⁺_CT was associated with favorable outcomes. IDO is an enzyme catalyzing the degradation of the amino acid tryptophan and is critical in immunosuppression. In microenvironment, the expression levels of IDO are diverse and are usually correlated with worse clinical outcomes.³⁷ In our study, IDO located in LUAD IM is mostly expressed on immune cells and was correlated with poor prognosis. The prognostic significance of PD-L1 remained controversial in NSCLC. Various researches revealed that PD-L1 expression levels could be relevant to various clinical outcomes, including good, poor, or no significance to prognosis.^38–41 In our study, our findings suggested that PD-L1 had a negative prognostic impact on LUAD. Nevertheless, as most of our patients received chemotherapy rather than immune checkpoint blockade treatment, the effects of PD-L1 in patients receiving adjuvant immunotherapy need to be further studied. CD73 is an ectoenzyme generating adenosine. The clinical role of CD73 in tumors is not yet well understood. One prostate cancer study reported that in the tumor stroma, highly expressed CD73 was related to prolonged biochemical recurrence (BCR)-free survival. The mechanism might be that the stroma CD73 were playing an anti-tumor role by reducing NF-κB signaling in the tumor cell.⁴² In this study, CD73 had a positive prognostic value and the favorable prognostic CD73 was mainly expressed in the stroma of LUAD CT, especially on CD4⁺ T cells. In contrast, CD73 in LUSC did not have prognostic significance.

The formula had two components for LUSC: CD8⁺CD39⁻CD73⁻_CT was a favorable prognostic factor, while CD8⁺Tim-3⁺_IM was negatively associated with tumor prognosis. CD39, also known as ecto-nucleoside triphosphate diphosphohydrolase (NTPDase 1), collaborates with CD73 in breaking down adenosine triphosphate (ATP) into adenosine.^32,43 Recent literature revealed that CD39 and CD73 were new potential “immune checkpoint mediators” as the extracellular adenosine they generate could suppress the anti-tumor immune response.³² In our research, we found that LUSC CT CD8⁺ T cells that did not express CD39 and CD73 were strongly associated with favorable outcomes, indicating that CD8⁺CD39⁻CD73⁻_CT cells might have the most potent cytotoxicity. Meanwhile, IM CD8⁺ T cells expressing Tim-3 were associated with worse outcomes, indicating a tumor-promoting effect of CD8⁺Tim-3⁺IM cells. It was reported that Tim-3 is a promising checkpoint.⁴⁴ Several studies revealed that highly expressed Tim-3 on CD8⁺ T cells had a negative prognostic significance, which is consistent with our result in LUSC.^45,46 Nevertheless, we found that CD8⁺ T cells played a double-faced role in LUSC.

To sum up, the strong prognosis-correlated immune features were entirely different between LUAD and LUSC. In LUAD, metabolism-associated factors like CD73, IDO, and checkpoint PD-L1 were playing more important roles. On the other hand, in LUSC, cytotoxic CD8 cells played a critical and double-faced role. Immune-related metabolism changes may play an important role in LUAD, but in the LUSC, the role of cellular immunity may be more outstanding. These findings partially harmonize with recent studies, ^47–49 and the results in this study maybe a result of the significant biological differences between LUSC and LUAD. LUSC has more tumor mutational burden and tumor-specific antigens, whereas LUAD has more targetable EGFR, ALK, or KRAS mutations.^5,47,48,50 In addition, Rictor (an important subunit of the mTORC2 complex) has a higher rate of gene amplification in LUSC, which may promote tumor development and metastasis.⁵¹ To our knowledge, this is the first study wherein LUAD and LUSC have been discussed separately in IS establishment, which is also the highlight of this study. Completely different results have been obtained; thus, respective IS models have been established. LUSC and LUAD have shown significant differences in genomic and histological heterogeneities; therefore, different results can be expected.⁵ Our findings suggested that the prognostic significance of various immune components from different subtypes were differential or even opposite. It is consistent with several previous studies on the immune microenvironment in LUSC vs. LUAD patients.^24,48,50 This study also suggests that the use of anti-PD agents or IDO inhibitors in the treatment of LUAD may have a better result, whereas in LUSC, treatments that increase T cell infiltration, promote lymphocyte activation, or reverse T cell exhaustion (such as anti-Tim-3 agents, etc.) may be more efficacious. Therefore, in view of the differences in these immune characteristics, future clinical trials need to be designed separately for LUAD and LUSC.

There are also some limitations to this study. All the patients enrolled in this study are from China, which may not represent all the races. However, on a global scale, China has a very high morbidity and mortality rate of lung cancer.^1,52 The investigation of IS_NSCLC in Chinese patients could provide many valuable information for oncologists and patients around the world. Besides, our training and validation cohorts were collected from the same institution. External validation should be done in the future to validate the universality of this model. In addition, this study was retrospective. Therefore, further international multicenter prospective clinical trials need to be carried out to validate the results and conclusions of this study.

In conclusion, the newly established IS_LUAD and IS_LUSC were completely different, and both the scores were excellent indicators for the prediction of recurrence and survival. The TN-I stage and nomograms constructed based on the IS_NSCLC could effectively improve the prognostic accuracy and facilitate the clinical application. With the help of the TN-I staging system, in the future, the prognosis prediction, clinical management, and individualized treatment decisions of lung cancer will be more precise and more reliable.

Authors’ contributions

Conception and design: XR; development of methodology: ZZ, FY, HZ, FW; acquisition of data (acquired and managed patients, provided facilities, etc.): ZZ, FY, YW; analysis and interpretation of data: ZZ, FY, YW; writing, review, and/or revision of the manuscript: ZZ, HZ, FW; administrative, technical, or material support: PZ, XZ; study supervision: XR, FW; all authors read and approved the final manuscript.

Conflicts of interest

The authors have declared no conflicts of interest.

Acknowledgments

The authors sincerely appreciate all the physicians and patients who participated in this study.

Supplementary material

Supplemental data for this article can be accessed on the publisher’s website.

Additional information

Funding

This study was supported by the National Key R&D Program [2018YFC1313400]; National Major Scientific and Technological Special Project for “Significant New Drugs Development” [2018ZX09201-015]; National Natural Science Foundation of China [81872166, 81974416, 81672697]; and key projects of Tianjin Health Industry [No.15KG145].