Four microarray expression profiles (i.e., GSE6044, GSE43346, GSE60052, and GSE149507) were obtained from the GEO database. The selected expression datasets satisfied the following criteria: 1) Studies were of human small-cell lung cancer tissues and corresponding para-cancerous tissues/normal lung tissues, 2) the number of samples included in each dataset was at least greater than 10, 3) the SCLC patients were naïve (untreated), and 4) the dataset search was limited to those studies written in English. The characteristics of the four selected datasets are shown in Table 1. For microarray data not shown in the form of log2-transformed values, log2 conversion was performed. If the data were not quantile-normalized, then the normalizeBetweenArrays method in the limma package of R was used for quantile normalization. The expression data were averaged for cases where multiple probes were mapped to one gene. Four GSE datasets were used to perform DEG identification, functional enrichment analysis, PPI network, and immune infiltration analysis in SCLC and normal lung tissue.

DEGs were identified using the limma package (version 3.42.2) with the empirical Bayes method (Ritchie et al., 2015). Key parameters that selected DEGs conformed to the following criteria: |log2 fold change (FC)|>1 and adjusted p < 0.05. The RobustRankAggreg package (version 1.1) was used to integrate DEGs from four selected gene profiles. Heatmaps (package: pheatmap, version 1.0.12) and volcano plots (package: ggpubr, version 0.4.0) were used to visualize gene expression patterns. The common DEGs among these four datasets were obtained by using “Venn Diagram.”

A GO analysis including biological process (BP), cellular component (CC), and molecular function (MF) was performed to better explore potential biological function of DEGs. In addition, KEGG analyses were performed on the up- and downregulated genes using the Bioconductor package “clusterProfiler” (version 3.14.3). Notably, q value (adjusted p value) < 0.05 was used as a cutoff, and only GO terms and signal pathways passing this threshold were considered significant.

To better decipher the connections between the identified DEGs, the STRING biological database was utilized to construct the original PPI network. Only those DEGs with interaction scores >0.9 could be mapped into the network (Szklarczyk et al., 2019). Cytoscape 3.8.1 was utilized for the generation and visualization of the PPI network (Smoot et al., 2011). The top thirty genes with the maximum interactions were defined as hub genes using the cytoHubba plug-in of Cytoscape software (Chin et al., 2014). In addition, the highly interconnected clusters were extracted from the PPI network using another plug-in, MCODE.

We procured SCLC RNA sequencing as well as corresponding clinical data from the cBioPortal database for the construction of a prognostic model; FPKM values were subsequently transformed into TPM values. Samples were included when both mRNA-sequencing data and corresponding survival data of an SCLC patient were accessible (77 samples). For platinum-related genes (that were obtained from functional enrichment analysis) and HLA class I genes, we performed LASSO regression analysis (ten-fold cross confirmation and p < 0.05) using R package glmnet (version 4.1.1). Then, those selected genes were utilized to generate a prognostic model for SCLC using multivariate regression analysis (package: rms; version 6.2.0).

Next, we observed the survival differences between high- and low-risk subgroups (the median value of risk score was used as a cutoff) via Kaplan–Meier analysis by using the “survive” and “survminer” R packages. Next, the 1-, 3-, and 5-year receiver operating characteristic (ROC) curves of the proposed model for SCLC in comparison to other clinicopathological factors were drawn. Besides, univariate and multivariate Cox regression analyses were implemented to show whether our proposed model had predictive value for the prognosis of SCLC. A nomogram was generated to predict 1-, 3-, and 5-year OS of SCLC patients.

In addition, a chi-square test was conducted to disclose relationships between the model and other clinicopathological characteristics by using the “ComplexHeatmap” R package (version 2.2.0). Furthermore, a scatter diagram was used to visualize the Wilcoxon signed-rank test analysis results, which revealed differences in risk scores across distinct groups of clinicopathological characteristics.

To show the independent prognostic potential of platinum resistance–related genes and genes in the proposed model (i.e., MSH2, TOP2A, BIRC5, CDKN2A, and HLA-ABC), we used survival and survminer packages to perform Kaplan–Meier analysis. Patients were clustered into two groups (high- or low-expression groups) based on the corresponding optimal cutoff value for each gene.

The immune infiltration analysis in SCLC and normal lung tissue was performed by using CIBERSORT (https://cibersort.stanford.edu), which could provide mRNA expression profiling of 22 immune cells. Notably, GSE6044 and GSE149507 only have 9 and 18 tumor samples, respectively. Given that analysis of few tumor samples may provide unreliable outcomes, we integrated four datasets to increase the total number of samples. The sva package (version 3.20) in R software (version 3.6.3; 64-bit) was applied to adjust potential batch effects. Besides, Spearman correlation analysis of the expression of the key genes (i.e., TOP2A, CDKN2A, BIRC5, MSH2, and HLA-ABC) and infiltrating immune cells was conducted by using the “ggpubr” package in the cBioPortal cohort.

Paraffin-embedded specimens were obtained from 151 patients with primary SCLC confirmed by surgery or needle biopsy between April 2013 and October 2019 at the Department of Clinical Oncology, Shengjing Hospital of China Medical University. In addition, none of patients had secondary tumors or other severe diseases, nor did they receive preoperative chemotherapy or radiotherapy. Overall survival time of individuals was assessed from the surgery date to the date of event occurrence (death or last follow-up). The last follow-up was on 19 December 2021. All patients provided informed consent before surgery.

The specific steps for the immunohistochemistry (IHC) procedure were performed strictly according to the instructions. Paraffin-embedded sections underwent xylol dewaxing and rehydration routinely, and then endogenous catalase was inactivated with 3% H₂O₂ solution (15 min). For better exposing antigenic sites, the sections were immersed in a pre-heated citrate buffer (95–96°C, pH 6.1) for 20 min and then blocked in 5% normal goat serum to avoid unspecific binding (15 min). The sections were then incubated with anti-TOP2A antibodies (1:100, Proteintech, United States) or anti-HLA-ABC (1:5000, Proteintech, United States) antibodies at 4°C overnight and then incubated with biotin-labeled goat anti-mouse/rabbit IgG (Zhongshanjinqiao, Beijing, China) at 37°C for 120 min. DAB (Zhongshanjinqiao, Beijing, China) was used for detection.

Three random images of each SCLC section were taken at high-power fields (×200), and Image-Pro Plus (version 6.0) was used to measure the average optical density (IOD/area). Thus, the staining index of SCLC tissues was quantified. For HLA-ABC analysis, the assessment was primarily based on the presence of membranous staining, while TOP2A staining was mainly based on intracytoplasmic or intranuclear staining. A time-dependent ROC curve was used to better ascertain the cutoff value of TOP2A and HLA-ABC (package: survivalROC, version 1.0.3).

The χ² test was utilized to disclose associations between TOP2A, HLA-ABC, and categorical variables (SPSS, Chicago, IL, United States, version 3.6.3). Survival data of patients with SCLC were analyzed using the Kaplan–Meier method, and comparison of different survival times between groups was drawn using the log-rank test (package: survminer, version 0.4.8). Correlation analysis between TOP2A and HLA was conducted using the Pearson method (SPSS) in datasets obtained from GSE datasets and the cBioPortal website, while Spearman rank correlation was used in our own cohort (two-sided p values and p < 0.05 was considered statistically significant).

Four microarray datasets (GSE6044, GSE43346, GSE60052, and GSE149507) were used to analyze DEGs, and volcano plots of DEGs for each dataset are presented in Figures 2A–D. The overlap of these DEGs was examined using a Venn diagram (Figure 2E). Then, these DEGs were integrated using the robust rank aggregation algorithm of the RobustRankAggreg package to obtain 313 DEGs (tumor vs. normal; 133 upregulated DEGs and 180 downregulated DEGs, respectively). A heatmap of the DEGs is shown in Figure 2F. The top 10 upregulated genes were CDC20, NOL4, INSM1, INA, NUSAP1, BIRC5, UCHL1, MAD2L1, TOP2A, and RRM2, while the top 10 downregulated genes were AQP1, MSLN, SLP1, SFTPD, SFTPC, PTGDS, FOLR1, CYP4B1, ADH1B, and SCGB1A1.

To shed light on the biological roles of DEGs in patients with SCLC, we performed GO and KEGG enrichment analyses for both upregulated and downregulated genes. Results of the GO analysis revealed that upregulated DEGs mainly focused on the cell mitosis process and microtubule motor activity (Figure 3A), and the KEGG analysis revealed that those genes were primarily mapped to the process related to cell cycle, DNA replication, p53 pathway, and platinum drug resistance (Figure 3C). In contrast, among the downregulated DEGs, GO terms showed significant enrichment in neutrophil-mediated immunity (Figure 3B), while KEGG analysis also showed enrichment in cytokine–cytokine receptor interaction and activation of transcription and the interleukin-17 signaling pathway (Figure 3D).

As depicted in Figure 4. 313 nodes and 1,548 edges were involved in the PPI network. The top 30 hub genes were screened on the basis of connectivity degree in this network (Figure 4F). The MCODE plug-in was then optimized to identify clusters in the network; four clusters were identified using k-core = 2 (Figures 4B–E). Cluster 1 had 32 nodes and 453 edges; it was the highest scoring cluster among those identified (Figure 4B). Core genes were also screened based on functional enrichment analysis and candidate hub genes. Our KEGG analysis further investigated four platinum resistance–related genes, that is, TOP2A, CDKN2A, BIRC5, and MSH2. Among these, TOP2A and BIRC5 were included in module 1 and with 37 and 35 connections, respectively, suggesting that they potentially play critical roles in SCLC. Notably, HLA-DMA was downregulated in SCLC tissues in contrast with its expression in normal tissues (logFC = −1.7849; p = 0.018). HLA genes are known to have an important function in immune response; therefore, to further investigate the expression of HLA family genes in SCLC, we searched each GEO dataset to find neglected information. We found that HLA-E was downregulated in GSE6044; HLA-E, HLA-DQB1, HLA-DPB1, and HLA-DMA were downregulated in GSE149507; and HLA-B, HLA-DQB1, HLA-DQB2, HLA-E, HLA-DMA, HLA-DPA1, HLA-DPB1, HLA-DRA, HLA-DRB1, HLA-DRB6, and HLA-DOA were downregulated in GSE60052; in contrast, HLA-related DEGs were not found in GSE43346. In summary, TOP2A and BIRC5 were upregulated in SCLC, whereas HLA family genes were possibly downregulated in SCLC.

We screened out five genes (i.e., TOP2A, BIRC5, CDKN2A, MSH2, and HLA-ABC) for further LASSO regression in light of the above findings. According to the optimal value of λ, the three-gene–based signature had superior predictive value (Figure 5B), and coefficients for each gene are presented in Figure 5A. Finally, a prognostic prediction panel was established. For each patient, the risk score was computed as follows: Risk   score = TPM   value   of   TOP 2 A × 0.002566 − TPM   value   of   HLA − B × 0.00022 − TPM   value   of   MSH 2   ×   0.0154

According to the median levels of risk score, patients were dichotomized into high- and low-risk groups. As shown in Figure 5C, patients with higher risk had significantly worse survival. Besides, our model outperformed other variables including patient age, gender, smoking, and metastasis status (Figures 5D–F), with an AUC = 0.851, 0.824, and 0.752 for 1, 3, and 5 years, respectively. In addition, survival analysis was employed to analyze platinum resistance–related genes (i.e., HLA-B, TOP2A, MSH2, BIRC5, and CDKN2A) and prognosis (Figures 5G–K). Patients with high expression of HLA-B, MSH2, and CDKN2A expression had better survival. Although p value for TOP2A was greater than 0.05, it can be seen from Figure 5G that the long-term survival of patients with high expression of TOP2A was worse than those with lower expression. Furthermore, results showed that a high risk score denoted worse prognosis, and this proposed model was predictive of outcomes in SCLC (Figure 5L).

Correlations between the model and various clinicopathological characteristics were further investigated by using the chi-square test; however, results indicated that the risk score was not significantly associated with clinicopathological factors (Figure 6A). Then, the Wilcoxon signed-rank test was performed to analyze possible discrepancy in risk score among subgroups stratified by clinicopathological characteristics. As indicated in Figures 6B–D, stage IV had a higher risk score than stage I–III. Similarity, N0–2 also showed no differences, but they had a much lower risk score than N3. Besides, distant metastasis was positively correlated with risk scores. A nomogram based on the panel and other traditional clinicopathological characteristics was constructed, as shown in Figure 7A. The calibration curve showed adequate fit of the proposed nomogram model in predicting 1-, 3-, and 5-year OS of SCLC (Figures 7B–D).

Given that the downregulated DEGs showed enrichment in immune-related pathways, immune infiltration analysis was performed in SCLC and normal samples. Interestingly, we found that SCLC tissues had a higher proportion of activated memory CD4⁺ T cells, T-follicular helper cells, regulatory T cells, M1 macrophages, and resting dendritic cells than that detected in normal lung tissues. In contrast, the proportion of monocytes, activated dendritic cells, resting mast cells, and neutrophils was much lower in SCLC tissues (Figure 8).

A lollipop plot was drawn to show correlations between expression of those key genes (i.e., TOP2A, CDKN2A, BIRC5, MSH2, and HLA-ABC) and immune-related cells. As shown in Figure 9A, TOP2A expression was negatively correlated with CD8+ T cells; HLA-A, HLA-B, and HLA-C expression had significant positive correlations with M1 macrophages, memory B cells, and CD8+ T cells. Besides, they all were inversely associated with M2 macrophages (Figures 9B–D). In addition, both HLA-A and HLA-C expression were positively correlated with M0 macrophages, while HLA-B expression was positively correlated with resting dendritic cells. However, CDKN2A, BIRC5, and MSH2 expression showed no correlations with immune cells (Supplementary Tables S1–S3).

To investigate the association between TOP2A and HLA-ABC, Pearson correlation analysis was applied in GEO datasets and the cBioPortal cohort. Of note, the correlation coefficients of TOP2A with HLA-ABC were all negative (Table 2). However, statistically significant results were only found in GSE60052, which may have happened because of the large sample size of this dataset (Table 2). Notably, in our cohort, TOP2A expression was significantly inversely correlated with HLA (Table 3).

The patient group (n = 151) comprised 105 males and 46 females, ranging in age from 37 to 79 years (mean ± SD, 61.27 ± 8.323 years old). Additionally, there were 109 patients (72.19%) who had lymph node metastases (LNM) and 104 patients (68.87%) had distant metastasis. TOP2A and HLA-ABC expression levels were analyzed semiquantitatively using the average optical density viewed in Image-Pro Plus after immunohistochemical staining. Representative pictures of IHC staining of SCLC are shown in Figure 10A, and more immunohistochemical pictures are shown in Supplementary Figure S1. Time-dependent ROC curves showed that the best critical cutoff values for TOP2A and HLA-ABC were 0.00498 and 0.00731, respectively (Figures 10C,D). The reciprocal of HLA-ABC (1/HLA-ABC) was used to improve the comparability of the ROC curves because this gene was downregulated in SCLC and may therefore be a protective factor. Thus, according to expression levels of TOP2A and HLA-ABC, patients were classified into high- or low-expression groups. Fifty-four percent of the patients had high TOP2A expression levels, while 46% of patients had low levels of TOP2A; in contrast, high and low expression of HLA-ABC was found in 32.5 and 67.5% patients, respectively. Correlations between TOP2A/HLA-ABC expression and clinicopathological characteristics of 151 patients with SCLC are shown in Table 4. However, there was no appreciable difference in age, sex, history of smoking, or lymph node metastasis between TOP2A/HLA-ABC high- and low-expression groups (p > 0.05).

Additionally, low expression of TOP2A and high expression of HLA-ABC predicted longer OS as indicated in Figures 10E,F. Univariate Cox analysis showed that LNM (HR = 3.000; p < 0.05), distant metastasis (HR = 4.500; p < 0.05), and high expression of TOP2A (HR = 1.800; p < 0.05) portended a poor prognosis, while high expression of HLA-ABC denoted a favorable prognosis. Multivariate Cox analysis showed that the risk of death was 2.1 times higher in patients with LNM than those without LNM and 3.5 times higher for patients with M1 stage than those with M0 stage. Patients with high expression of TOP2A had 1.5-fold higher mortality risk than patients with low expression, while patients with high HLA-ABC expression harbored a 0.66-fold lower risk of death. Therefore, TOP2A, HLA-ABC, lymph node metastasis, and distance can be regarded as independent prognostic factors of SCLC (Figure 10B).