From 2014 to 2018, 102 SCLC surgery specimens were collected from the Shanghai Pulmonary Hospital, China. Two independent reviewers screened the pathological types and surgical histology reports (Chunyan Wu and Liping Zhang). The tumor-node-metastasis staging system version 8th was applied. Samples were obtained following written informed consent from all participants.

After dewaxing by xylene and alcohol, all formalin-fixed, paraffin-embedded tissue slides were rinsed with distilled water. Then, the target retrieval solution kit (DM828 or DM829, Dako) was used for antigen repairing. In order to reduce the background staining, we used 3% hydrogen peroxide. Primary antibodies (Galectin-9, NBP2-45619, Novusbio), and secondary antibodies which were goat-anti-Mouse/Rabbit IgG that labeled with horseradish peroxidase were applied standardly.

Two independent pathologists (Chunyan Wu and Liping Zhang) reviewed all clinical samples (online supplemental table S1). Once discrepant evaluations were obtained, they reviewed together to arrive at consensus results. More than 30% staining was the cut-off of Gal-9 on TILs. On tumor cells, all positive stains of Gal-9 were regarded as positive. The screening process to find the best cut-off point was completed by survival analysis.19 29

We adapted the eXtreme Gradient Boosting (XGBoost) algorithm to construct XGBoost predictive models by various immune biomarkers and clinical features.30 As a machine-learning technique, XGBoost algorithm could work with the data of first and second derivatives to discovery non-linear relationship. It also could employ regularization item to control the overfitting and overly complex of predictive model, and provide the contribution of each feature to the outcome. To be specific, the corresponding formula of regularization item is as followed:

Ω(fk)=γT+12λ‖w‖2

where T represents the number of leaves, W is defined as the magnitude of leaf weights. Both γ and λ are two penalty parameters that could respectively control penalty for T and W. By means of cross-validation, the penalty parameter is chosen. In the process of pruning, the threshold value of γ helps restrict the internal nodes of tree. During the process of smoothing, coefficient λ was added, thus finally avoid overfitting.

In the study, the whole cohort was randomly divided into the calibration subset and training subset which accounted for 70%. The final XGBoost survival models were composed of the top three predictive features and limited to the maximum depth of 6. Further, the process of model construction was repeated for 1000 times so as to fully use the sample information. The predictive value of the XGBoost model was visualized by the log-rank test. Then, we combined results of XGBoost model which ranked the relative importance of each signature and Cox multivariate model which offered coefficients of selected features to construct risk score models for patients with SCLC. The prognostic risk score equation of immune biomarkers was: immune risk score=(–0.550*Gal-9 on TILs)−(0.295*CD4)–(0.407*PD-L1 on TILs). The performance of risk score model was assessed by the areas under time-dependent receiver-operating characteristic (ROC) curves (AUCs). The larger the AUCs, the higher the quality of prognostic prediction of XGBoost predictive model.

In order to further evaluate the predictive value of Gal-9 and immune risk score in advanced patients, we used the cBioportal Database (https://www.cbioportal.org). The enrolled cBioportal dataset must meet the following inclusion criteria1: mRNA sequencing for tumor tissues from patients with clinical stage IV SCLC,2 complete mRNA expression data,3 prognostic information from patients with clinical stage IV SCLC.

In order to verify the expression of Lgals9 mRNA which encodes Gal-9 protein in SCLC cell lines and tumor tissues, we used the Cancer Cell Line Encyclopedia (CCLE) Database (https://portals.broadinstitute.org/ccle)31 and the Gene Expression Omnibus (GEO) Database (https://www.ncbi.nlm.nih.gov/geo/). As one of cancer-related databases, CCLE currently summarizes the expression level of more than 80,000 genes in total 1457 cancer cell lines. GEO is a public genome database which provides various kinds of gene expression data of corresponding study. The mRNA expression data from GEO were identified according to the following inclusion criteria1: mRNA sequencing for tumor tissues from patients with clinical SCLC,2 mRNA sequencing for normal tissue,3 complete mRNA expression data. The following exclusion criteria were considered1: insufficient data were available to compare gene expression,2 mRNA sequencing for animals or cell lines. After downloading suitable expression profiling from GEO, limma R package was used for screening differently expressed genes (DEGs) between tumor and controlled group.

In order to explore different biological pathways between high Gal-9 expression group and low Gal-9 expression group, the Gene Set Enrichment Analysis (GSEA) software (V.4.0.3) was employed.32 We divided the mRNA expression dataset of GEO into two groups evenly based on the Gal-9 expression level and kept all parameters in GSEA set at their defaults. The network between Gal-9 and Gal-9-related genes with strong correlation (>0.6) was visualized by Cytoscape software (V.3.7.1; https://cytoscape.org/).33

To investigate the landscape of immune infiltration in patients with SCLC with high and low immune risk, we applied CIBERSORT method to the mRNA expression profile. As one of online databases for immune-infiltration analysis, CIBERSORT provides relative proportion of 22 human immune cell types in tumor tissues on the basis of deconvolution method.34 LM22 is a leukocyte gene signature matrix with high sensitivity and specificity for estimating 22 human immune phenotypes, including naive B cells, memory B cells, CD8 T cells, different CD4 T cell types, Tregs, NK cells, plasma cells, monocytes, three macrophages types, and dendritic cells. According to the calculation results of immune risk score based on the expression level of Lgals9, CD4 and CD274 that encode PD-L1, 23 SCLC clinical samples were divided into high-risk and low-risk group. Then, by combining CIBERSORT with LM22, the assessment of the component of 22 immune cells in each clinical sample of high-risk and low-risk group was obtained.

We evaluated the correlation analysis between Gal-9 status and clinical factors or program death-1 (PD-1)/PD-L1 by Χ² tests. Through taking multiple characteristics into account, we used univariate and covariate logistic regression analysis for predicting Gal-9 expression. We also performed Cox regression analysis and Kaplan-Meier method, which helped compare the prognosis conditions of different groups. All statistical examines were two-sided, and the p value smaller than 0.05 was defined as statistical significance. By means of X-tile software (V.X86, Yale University, USA), we picked the best cut-off value of immune risk score. The statistical tool SPSS (V.22.0) and the R Programming Language (V.4.0.1) for Windows were installed for the data analysis.

There are 102 patients in total, with a median age of 62. The majority of patients were under 70 years old (79/102, 77.5%). Among all enrolled participants, men (84/102, 82.4%) were more than women (18/102, 17.6%). There were 58 (56.9%) non-smokers, and 44 (43.1%) were smokers. All patients were stage I–III. In the cohort as a whole, stage I–II accounted for a little more than half (60, 58.8%) (table 1).

More than half of patients received treatment after surgery, including chemotherapy alone (40/102, 39.2%), radiotherapy alone (1/102, 0.01%), and chemotherapy plus radiotherapy (26/102, 25.5%). All enrolled patients had pulmonary nodules which were highly suspected as malignant tumor by imaging examination, thus receiving surgery to further confirm pathological types and follow-up care. However, a total of 35 patients with SCLC were treated with surgery alone because of some practical reasons, such as the contraindication of chemotherapy, financial stress, and personal willingness of refusal of postoperative treatment. In addition, a small group of patients relapsed and died within a month, thus losing the chance of postoperative treatment.

In all specimens, 32 (31.4%) were positive Gal-9 expression on tumor cells and 28 (27.5%) were positive Gal-9 expression on TILs (figure 1). There was no significant correlation among clinical factors and Gal-9 level on tumor cells when gender, age, smoking status, metastasis status and SCLC staging were taken into consideration (p>0.05). Similarly, negative results were obtained in the Gal-9 level on TILs (p>0.05) (table 2).

Through summarizing the correlation between Gal-9 on TILs and other immune biomarkers, we detected that the status of Gal-9 on TILs had widespread contacts with other immune checkpoints or immune cell level including PD-1 on TILs (p=0.001), PD-L1 on TILs (p<0.001), CD3 (p<0.001), CD4 (p<0.001), CD8 (p<0.001), and FOXP3 (p=0.047). However, results among different PD-L1 status on malignant cells, there was no significance of the TILs’ Gal-9 status in statistics (p=0.182). The p value, which was higher than 0.05, illustrated that the degree of Gal-9 expression on malignant cells was not significantly related to immune biomarkers that were taken into consideration (table 3).

By modifying relevant parameters, ORs and corresponding 95% CIs were summarized in online supplemental tables 2 and 3. On TILs, logistic regression analysis identified that the OR for Gal-9 status was 11.581 (95% CI, 2.093–64.083; p=0.005) when samples revealed CD3 positive compared with those revealed negative. Regretfully, none of other variables included had a statistically significant effect on SCLC cancer cells’ Gal-9 status.

In this study with 102 patients enrolled, the median recurrence-free survival (RFS) was 18.0 months, 56 (54.9%) patients had relapsed by the end of 2018. In all 60 patients in stage I–II, 25 (41.7%) reached the end event for RFS (median 19.0 months). The median RFS for all 42 patients with stage III SCLC was 15.0 months, among whom 31 (73.8%) had relapsed. With Kaplan-Meier method for time to relapse as the criterion standard, we analyzed the differences between positive Gal-9 and negative Gal-9 on TILs or tumor cells. We found that the positive Gal-9 on TILs demonstrated better RFS (RFS 30.4 months, 95% CI: 23.8–37.1 vs 39.4 months, 95% CI: 31.6–47.3, p=0.009). For the status of Gal-9 expression on tumor cells, the median time of RFS was 32.0 months (95% CI, 22.2–41.8）in the positive group, and 35.1 months (95% CI, 28.0–42.3) for patients with SCLC with negative Gal-9 status. In spite of a difference between two datasets, no significance was showed in statistical terms (p=0.714; figure 2).

We carried out the subgroup analysis based on Gal-9 level on TILs (figure 3 and online supplemental figure S1). It is worth noting that both Gal-9 and PD-1 on TILs positive (vs either Gal-9 and PD-1 on TILs positive or both Gal-9 and PD-1 on TILs negative; 39.4 months, 95% CI: 29.9–48.9 vs 33.1 months, 95% CI: 25.0–41.2 vs 28.8 months, 95% CI: 21.1–36.4, p=0.040), both Gal-9 and PD-L1 on TILs positive (vs either Gal-9 and PD-L1 on TILs positive or both Gal-9 and PD-L1 on TILs negative; 42.2 months, 95% CI: 33.5–50.9 vs 28.3 months, 95% CI: 21.3–35.3 vs 28.9 months, 95% CI: 21.7–36.1, p=0.014), both Gal-9 on TILs and CD3 positive (vs either Gal-9 on TILs and CD3 positive or both Gal-9 on TILs and CD3 negative; 40.4 months, 95% CI: 32.7–48.2 vs 35.4 months, 95% CI: 23.7–47.2 vs 27.9 months, 95% CI: 20.1–35.7, p=0.012), both Gal-9 on TILs and CD4 positive (vs either Gal-9 on TILs and CD4 positive or both Gal-9 on TILs and CD4 negative; 44.2 months, 95% CI: 35.9–52.5 vs 24.5 months, 95% CI: 16.7–32.3 vs 30.4 months, 95% CI: 23.5–37.4, p=0.017), either Gal-9 on TILs or CD8 positive (vs both Gal-9 on TILs and CD8 positive or both Gal-9 on TILs and CD8 negative; 45.5 months, 95% CI: 33.1–57.9 vs 41.3 months, 95% CI: 31.9–50.6 vs 27.9 months, 95% CI: 21.1–34.7, p=0.005) was notably correlated with longer RFS in SCLC. In spite of the significantly prognostic differences in the subgroups of Gal-9 on TILs in combination with PD-L1 on cancer cells (p=0.045) or FOXP3 (p=0.014), the double immune biomarkers positive group failed to fully reflect the objective fact for its limited sample size. The subgroup analysis of Gal-9 level on cancer cells in combination with PD-1, PD-L1, CD3, CD4, CD8, and FOXP3, respectively, showed no significant difference among different groups, which indicated the failure of Gal-9 on tumor cells in predicting the RFS in SCLC (online supplemental figure S2).

Univariate and multivariate Cox regression models with categorical variables were established in sequence, for the purpose of adjusting for potential confounding characteristics and identifying prognostic factors. The HRs and their 95% CIs were calculated for assessment. By univariate Cox regression, SCLC staging (p=0.006) and Gal-9 level on TILs (p=0.012) were considered as the meaningfully predictive biomarkers for RFS. Multivariate Cox regression analysis further indicated that positive Gal-9 on TILs (vs negative GAL-9 on TILs; p=0.024, HR 0.436, 95% CI: 0.212–0.897), and SCLC stage I–II (vs SCLC stage III; p=0.014, HR 1.951, 95% CI: 1.146–3.322) were significantly related to better prognosis (table 4).

Given the significant significance of subgroup analysis which all included Gal-9 level on TILs, we proposed the hypothesis that Gal-9 level on TILs had the meaningful interrelation with other immune biomarkers. To confirm this conjecture, XGBoost was first used for features selection. The diagram of feature importance to outcome which comprised all immune biomarkers illustrated that Gal-9 on TILs ranked first, CD4 second, and PD-L1 on TILs third (figure 4A). By incorporating top three variables, XGBoost results showed that the predictive curve was fitted well with the actual one (21.0 months, 95% CI: 16.3–25.7, vs 17.0 months, 95% CI: 9.7–24.3, p=0.300; figure 4B).

Therefore, Gal-9 on TILs, CD4, and PD-L1 on TILs were chosen to construct the prognostic risk score model for SCLC. The survival analysis demonstrated that the high-risk group contributed poorer prognosis (RFS, high risk 43.7 months, 95% CI: 35.9–51.5 vs low risk 29.9 months, 95% CI: 23.5–36.4, p=0.002; figure 4C). The significant difference between immune risk score and RFS was also suggested by univariate Cox regression (p=0.004, HR 3.874, 95% CI: 1.541–9.741; table 4). Considering that risk score covered the feature of Gal-9 status on TILs, the multivariate Cox regression model that included SCLC risk score and staging was established. Both risk score (p=0.009, HR 3.424, 95% CI: 1.351–8.676) and staging (p=0.026, HR 1.836, 95% CI: 1.077–3.131) were considered as independent prognostic features in SCLC (table 4). Furthermore, by means of time-dependent ROC analysis, immune risk score model obtained the best AUC value when compared with single immune biomarker. The AUC value for risk score model, Gal-9 on TILs, CD4, and PD-L1 on TILs were 0.671, 0.622, 0.621, 0.644, respectively (figure 4D), which highlighted that the risk score model performed better than other immune biomarkers in the prediction of prognosis in stage I–III SCLC.

We downloaded the suitable SCLC dataset which contained 81 clinical SCLC samples from cBioportal Database.35 After data screening, the RNA sequencing (RNA-Seq) data of patients with stage IV SCLC were available for nine specimens (online supplemental table S4). For patients with SCLC in early and extensive stage, similar results of correlation analysis between Gal-9 and clinical factors or immune markers were obtained (online supplemental figure S3A and table S5). Gal-9 also especially showed significant correlation with PD-1 (p<0.001), PD-L1 (p<0.001), CD3 (p<0.001), CD4 (p<0.001), CD8 (p=0.04), and FOXP3 (p=0.002) in advanced SCLC. In addition, no significant correlation exhibited between Gal-9 and clinical features, including age (p=0.408) and sex (p=0.359).

Then, we also investigated the prognostic value of Gal-9 in nine patients with stage IV SCLC from public dataset. The survival analysis indicated that patients with extensive SCLC with higher Gal-9 expression level showed better overall survival (OS) than patients with lower Gal-9 expression (16.0 months, 95% CI: 7.4–24.6 vs 7.0 months, 95% CI: 2.1–11.9; p=0.122; online supplemental figure S3B). For better evaluating the performance of the risk score model, we applied it to patients with SCLC in extensive stage IV. The result supported that patients with advanced SCLC with higher risk score had shorter OS (vs lower risk score; 16.0 months, 95% CI: 7.4–24.6 vs 7.0 months, 95% CI: 2.1–11.9; p=0.122; online supplemental figure S4B).

We further verified the relative expression level of Gal-9 in both SCLC cell lines and tissues. The CCLE Database collected the mRNA expression level of Gal-9 coding gene, Lgals9, in 54 SCLC cell lines and 136 non-SCLC (NSCLC) cell lines. As the online supplemental figure S4A showed, Lgals9 was lowly expressed in the SCLC cell lines when compared with the NSCLC cell lines. In the GEO Database, both GSE43346 dataset which included expression profiles of 23 clinical SCLC tissues and 43 normal specimens, and GSE6044 dataset which was composed of 9 SCLC samples and 5 control subjects without cancer met the inclusion criteria, thus being included in this study. The significant differences of Lgals9 expression between SCLC and normal tissues were testified in the above two datasets (online supplemental figure S4B). Specifically, the expression level of Lgals9 in SCLC was lower than that of the controlled group (GSE43346, p=0.014; GSE6044, p=0.028).

For the sake of better understanding the biological pathways that significantly participated in group with high Gal-9 expression and investigating latent Gal-9-related genes, we performed GSEA and Cytoscape in GSE43346 dataset. Among a total of 174 gene sets which were upregulated or downregulated between two SCLC groups, 119 upregulated gene sets in the high Gal-9 expression group accounted for the largest proportion (119/174, 68.4%). Figure 5 demonstrated that the top four high Gal-9 expression-related pathways with enrichment scores >0.6 and false discovery rate <0.25 were as follows: “KEGG_PRIMARY_IMMUNODEFICIENCY”,“KEGG_SYSTEMIC_LUPUS_ERYTHEMATOSUS”, “KEGG_ALLOGRAFT_REJECTION”, and “KEGG_GRAFT_VERSUS_HOST_DISEASE”. A close relationship and high overlapping rate between the above four Gal-9 expression-related pathways were found (online supplemental figure S5A). A total of 15 genes overlapped in three of these pathways, indicating that they might played crucial roles in the high Gal-9 level. Among more than 20,000 genes, CD4 was also involved in the pathway which significantly enriched between different Gal-9 expressions. The results of leading edge analysis revealed that Jaccard values of numbers of the occurrences mainly concentrated on the range of 0–0.40 (online supplemental figure S5B). Further, we verified that 18 genes participated in mentioned high Gal-9-related pathways were differentially expressed between normal and SCLC tissues online supplemental figure S6A. In addition, the expression level of Gal-9 was moderately to highly related to that it of DEGs in the tumor microenvironment (online supplemental figure S6B). The correlation matrix also displayed that the expression level of all 18 DEGs had moderate correlations with each other. The network by Cytoscape visualized the strong correlation (>0.6) between all 18 DEGs and Lgals9 in SCLC (online supplemental figure S6C).

In the GEO SCLC cohort, we further analyzed the difference of immune infiltration condition between high-risk and low-risk score group by CIBERSORT and LM22. Two heatmaps separately depicted the detailed immune characteristics of 22 immune cells in patients with SCLC with high and low immune risk score (figure 6A,B). The relative percentage of TILs varied from sample to sample and summed up to 100%. Online supplemental figure S7A,B summarized the relationship between all immune cell proportions in two SCLC groups, while the correlation with each other was fairly modest. Then, we explored the significant differences of activated memory CD4 T cells between high-risk and low-risk group when the expression level of three prognostic biomarkers were all taken into account (p=0.048, online supplemental figure S7C). The low-risk group displayed considerable higher enrichment of activated memory CD4 T cells in comparison with high-risk group, indicating that the heterogeneity of immune cells in SCLC might act as a meaningful feature for outcome prediction. These results further verified the importance of the immune risk score model in terms of the tumor microenvironment and immune infiltration in SCLC.