Identification and validation of an immune cell infiltrating score predicting survival in patients with lung adenocarcinoma

We adopted the public datasets from the Gene Expression Omnibus (GEO) (https://​www.​ncbi.​nlm.​nih.​gov/​geo/​) as the training cohort. Four GEO datasets (GSE30219, GSE37745, GSE50081 and GSE68465, the latter also known as the Director’s Challenge Consortium) were included in this study. All of the four datasets were derived from the genechips of Affymetrix^® (Santa Clara, California, USA). GSE30219, GSE37745 and GSE50081 were based on the GPL570 genechip, while GSE68465 were from the GPL96 microarray. Then, raw data and genechip files were downloaded directly. A total of 85, 106, 127 and 443 samples of lung adenocarcinoma from GSE30219, GSE37745, GSE50081 and GSE68465 datasets were enrolled. We used a robust multichip average method via RMAExpress for background adjustment, quantile normalization and summary to process the gene profiles [17‐19].

Level 3 RNA sequence data from lung adenocarcinoma samples were also downloaded from The Cancer Genome Atlas (TCGA) before June 5, 2018 (https://​portal.​gdc.​cancer.​gov/​). A total of 594 samples were obtained, comprising 535 adenocarcinoma and 59 normal lung samples. Baseline clinicopathological factors and treatment information were also acquired from TCGA. Primary lung adenocarcinoma samples with complete follow-up and baseline information (age, sex and pathological stage) were included as the validation cohort in this study. Due to the original records of all included datasets, the seventh edition of the TNM staging method was utilized in our study. Information regarding histologic subtypes of lung adenocarcinoma was intergraded with records from the University of California Santa Cruz Xena database (http://​xena.​ucsc.​edu/​) and previous literature [20‐22]. According to previous studies, tumors were classified into three subgroups (low-, intermediate- and high-grade adenocarcinoma) regarding the survival data for each predominant lung adenocarcinoma subtype [23, 24].

The CIBERSORT method was adopted to quantify the proportions of the immune cell in both training and validation cohorts [25, 26]. Common methods, like flow cytometry and immunohistochemistry, rely on few phenotype markers and tissue disaggregation prior to cytometry, may cause damaged or lost cells, altering analyses [26]. The CIBERSORT was developed to accurately evaluate the relative levels of the 22 immune cell phenotypes, especially closely related types, using signature gens within complex expression mixtures [26]. Such mixtures could derive from both patients’ solid tissues and blood profiled by array or RNA-sequencing [26]. The 22 immune cells are mainly composed of B cells, T cells, macrophages, dendritic cells, plasma cells, natural killer cells and mast cells. To process data by the CIBERSORT, the targeted gene profile was uploaded to its website (https://​cibersort.​stanford.​edu). Then, a heat-map table with levels of 22 immune cell subtypes returned as the final results. As previously described, we set the threshold P-value < 0.05 and excluded samples without immune infiltration in training (5 patients) and validation (78 patients) cohorts [25]. The optimal cut-off values of the proportions of different immune cells in the training cohort were calculated on the basis of the prognostic significance using X-Tile software [27]. Each immune cell fraction level of samples from both cohorts was then divided into two groups based on the cut-off point. The immune cell fraction level divided by the cut-off value was valued as 0 or 1 in the subsequent scoring formula.

All statistical analyses in this study were performed using R version 3.5.0 (R Foundation for Statistical Computing, Vienna, Austria) and IBM SPSS Statistics 22.0 (IBM, Inc., Armonk, NY, USA). We used the LASSO Cox regression model to select the ideal prognostic model among the estimated immune cell subtypes using the glmnet package in R [28]. The optimal model parameter λ and corresponding coefficients were determined by tenfold cross validations. We chose λ via 1 − SE (standard error) criteria as previously recommended [8, 9]. Based on the final immune infiltrating score model and corresponding coefficients, a formula for the score was constructed using the training cohort. The sensitivity and specificity of the prognosis prediction of the immune cell infiltrating score were evaluated by a time-dependent receiver operating characteristic (ROC) curve in R software. The nearest neighbor estimation method in the ROC curve was adopted with cut-off times of 1, 3 and 5 years. Survival curves were estimated using the Kaplan–Meier method and the log-rank test was utilized for comparing survival curves. Univariate and multivariate Cox proportional hazard models were adopted to identify all significant prognostic factors. The prognostic value of the included variables was evaluated by Harrell’s concordance index (c-index) [29]. The score stratified by subgroups were compared using the one-way ANOVA method, if normal distribution and homogeneity of variance were followed. Correlation analysis was also conducted to explore the potential relationship between the score and the expression of immune checkpoint regulators. In this study, a two-tailed P-value of < 0.05 was considered statistically significant.

After the initial selection process, 751 and 418 patients were enrolled as the training and validation cohorts, respectively. The characteristics of the patients are listed in Table 1 and Additional file 1: Table S1. Optimal cut-off values for the fraction of each immune cell type in the training cohort were generated as described in “Methods” section (Additional file 2: Table S2). LASSO Cox regression analysis was performed, based on the fraction of immune cells in the training cohort (Fig. 1a, b). Finally, the formula of the immune cell infiltrating score for lung adenocarcinoma in the training cohort was built (Additional file 3). The prognostic accuracy of the scoring model was evaluated using time-dependent ROC analysis in the training cohort. The predictive time points were set at 1, 3 and 5 years, respectively (Fig. 2a). The area under the ROC curve for the immune infiltrating score at each time point was 0.674, 0.684 and 0.675, respectively.

Survival analyses were performed stratified by the score, in the training cohort (Fig. 2b). Patients were separated into low- and high-score groups by the median score of the cohort. We observed that patients with a high score had significantly worse survival than those with a low score (P < 0.001, Fig. 2b). A high score was also associated with significantly shorter progression-free survival (P < 0.001, Additional file 4: Figure S1).

As a continuous variable, the immune cell infiltrating score was found to have significantly prognostic value using univariate Cox regression analysis (P < 0.001, Table 2). Multivariate analysis confirmed that the immune infiltrating score was an independent prognostic factor in the training cohort, along with the TNM staging system (P < 0.001, Table 2). According to the corresponding c-index, the combination of the scoring model and TNM stage showed a significantly better predictive value than the TNM stage alone in this cohort (P < 0.001, Table 3).

We adopted a validation group based on the dataset from TCGA to evaluate the prognostic value of the proposed scoring model. The same formula for the immune cell infiltrating score and optimal cut-off point for each immune cells were applied to the validation group. Likewise, ROC analysis was adopted to assess the prognostic value of the scoring model. The area under the curve was 0.650, 0.616 and 0.567 at the predictive time of 1, 3 and 5 years, respectively (Fig. 2c). In the validation group, a high score was associated with significantly worse prognosis (P = 0.013, Fig. 2d).

Consistent with the previous findings, the score was an independent prognostic factor in the validation cohort, based on the univariate and multivariate Cox regression model (P = 0.001 and 0.002, Additional file 5: Table S3). Furthermore, the combination of the scoring model and TNM stage improved the prognostic model of lung adenocarcinoma in this cohort (P < 0.001, Table 3).

In the training group, data regarding adjuvant chemotherapy were documented only in the GSE68465 and GSE37745 datasets. Patients were divided into low and high score groups based on the score median value. Given the differences in staging methods and initial study design, patients with stage I disease were excluded from this analysis. The survival advantage of the low-score group was observed in patients with or without adjuvant chemotherapy (P = 0.004 and 0.002, Additional file 6: Figure S2A, B). Moreover, there was a trend for better prognosis in patients with a low score who received adjuvant chemotherapy, although the difference was not statistically significant (P = 0.909, Fig. 3a). No significant prognostic difference regarding adjuvant chemotherapy was observed in high score group (P = 0.764, Fig. 3b). We also extracted data on adjuvant chemotherapy from the validation group. There was a trend for better survival in low score group who received chemotherapy (P = 0.084, Additional file 6: Figure S2C). No difference was found in patients without chemotherapy (P = 0.761, Additional file 6: Figure S2D). Furthermore, patients with a low score may benefit from adjuvant chemotherapy (P = 0.010, Fig. 3c), while no difference was observed in high-score group (P = 0.213, Fig. 3d).

Smoking status was included in two datasets (GSE50081 and GSE68465) for the training cohort. Patients with unknown smoking status were excluded. We observed that patients who are current smokers had significantly higher score than those who do not smoke (P < 0.001, Fig. 4a). In the validation cohort, patients who are current or reformed smokers (< 15 years) had higher scores than non-smokers or reformed smokers (> 15 years), although the difference was not statistically significant (P = 0.932, Fig. 4b).

Furthermore, we explored potential relationships between the immune cell infiltrating score and histologic subtypes of lung adenocarcinoma. In the validation cohort, 186 patients with information on histologic subtypes were included. We found that high-grade adenocarcinoma subtypes were associated with significantly higher score than low-grade tumors (P = 0.003, micropapillary/solid vs acinar/papillary vs lepidic predominant lung adenocarcinoma, Fig. 4c).

Owing to the difference of the microarray platforms used for the training cohort, correlation analyses between the immune infiltrating score and the expression of immune checkpoint regulators or inflammatory mediators were performed in the validation cohort. We found that there was a significant positive correlation between the score and some of the regulators and mediators, including PD-L1 (P = 0.002), IDO1 (P = 0.003) and LAG3 (P = 0.017), IFNB1 (P = 0.002), IL-1A (P = 0.005), TNFA (P = 0.010) and IL-6 (P = 0.007) (Additional file 7: Table S4).