Characterization of stemness features and construction of a stemness subtype classifier to predict survival and treatment responses in lung squamous cell carcinoma

The workflow of our study is shown in Figure S1. First, the correlation between the stemness index and clinicopathological characteristics was analyzed based on the mRNAsi scores of LUSC patients from the TCGA dataset. Patients aged < 65 years had higher mRNAsi scores than patients aged ≥ 65 years (Fig. 1A). There was no significant difference among patients of different gender, T stages, N stages and clinical stages (Fig. S2A). The mRNAsi scores in normal samples were lower than those in tumor tissues (Fig. 1B). A positive correlation between mRNAsi scores and TMB was observed (Fig. S2B). The waterfall plots displayed the 20 frequently mutated genes in these two groups (Fig. S2C). Next, we systematically examined the correlation between the stemness index and patient prognosis in LUSC. We found that patients in the high-mRNAsi group exhibited better OS than those in the low-mRNAsi group (Fig. 1C). Patients in the high-mRNAsi group presented a lower probability of death than those in the low-mRNAsi group with an increase in the stemness index (Fig. 1D). We further validated the association between the stemness index and prognosis in three GEO independent datasets (GSE30219, GSE37745 and GSE73403). As expected, patients in the high-mRNAsi group had longer OS than those in the low-mRNAsi group (Fig. 1E). Altogether, these results indicated that the increase in mRNAsi scores was closely associated with a better prognosis in LUSC.

An increasing number of studies have reported that cancer stemness is closely associated with immune cell infiltration [29]. We first investigated the relationship between the stemness index and tumor immune status in the TCGA dataset by the ESTIMATE algorithm. Correlation analysis demonstrated that the mRNAsi score was positively correlated with tumor purity but negatively correlated with the immune score, stromal score, and ESTIMATE score (Fig. 2A). These results suggested that the levels of tumor-infiltrating immune cells decreased with increasing mRNAsi score in LUSC patients. Next, the ssGSEA algorithm based on 29 immune gene sets was used to quantify the immune-related pathways and relative abundance of immune infiltrating cell subpopulations. Consistent with the ESTIMATE analysis, we observed that the low-mRNAsi group was enriched in most immune-related pathways, including T-cell costimulation/inhibition, APC costimulation/inhibition and I/II-IFN response, and most types of immune cells were elevated in the low-mRNAsi group (Fig. 2B). In addition, we utilized multiple algorithms, including TIMER, QUANTISEQ, MCP-COUNTER and XCELL, to estimate the relationship between immune cells and the mRNAsi score. We found that the mRNAsi score was negatively correlated with the main types of immune cells (Fig. 2C). Taken together, these data indicated that the stemness index was negatively correlated with low tumor immune status in LUSC.

Because the stemness index was closely associated with OS and the tumor immune microenvironment status of LUSC, we conducted further investigations into the difference between the high mRNAsi and low mRNAsi groups. The GSVA enrichment analysis revealed a significant enrichment of pathways related to DNA damage repair in the high mRNAsi group. These pathways included homologous recombination, mismatch repair and nucleotide excision repair pathways (Fig. S3A). A total of 190 stemness-related DEGs were screened based on the selected thresholds (|log2-fold change (FC)| >1 and adjusted p value < 0.01), including 15 upregulated genes and 175 downregulated genes (Fig. 3A). Next, functional enrichment analysis was used to evaluate the biological functions and signaling pathways of these 190 stemness-related DEGs. KEGG pathway analysis showed that these DEGs were enriched in pathways related to the tumor microenvironment, including protein digestion and absorption, ECM-receptor interaction and complement and coagulation cascade pathways. Several classical cancer-related pathways also emerged, including the PI3K-AKT signaling and TGF-beta signaling pathways (Fig. 3B). The GO enrichment analysis indicated that these DEGs participated in (BP terms) external encapsulating structure organization (Fig. 3C), (CC terms) collagen containing extracellular matrix (Fig. 3D) and (MF terms) extracellular matrix structural constituent (Fig. 3E).

To further analyze the heterogeneity of stemness characteristics, an unsupervised consensus clustering method was applied to identify a new molecular cluster of LUSC patients based on 190 stemness-related DEGs. According to the consensus heatmap and the relative change in the area under the CDF curve, k = 2 was the optimal value for clustering (Fig. 4A). All LUSC patients were classified into 2 stemness subtypes, including stemness subtype A (259 patients, 54.2%), which tended to have a lower mRNAsi score, and stemness subtype B (219 patients, 45.8%), which had a higher mRNAsi score (Fig. 4B). The Kaplan‒Meier curve demonstrated that LUSC patients with stemness subtype B exhibited a longer OS than patients with stemness subtype A (Fig. 4C). The mRNAsi score and TMB value in the stemness subtype A group were lower than those in the stemness subtype B group (Fig. 4D). The mutation statuses of TP53, TTN and CSMD3 were statistically significant between the two subtypes, and waterfall plots displayed high-frequency mutations of different genes between the two stemness subtypes (Fig. 4E). Furthermore, patients in stemness subtype A were older than those in stemness subtype B. No significant difference was found in gender, TNM stage, clinical stage or smoking status between the two stemness subtypes (Fig. S3B).

To elucidate the potential differences in molecular pathways related to the two stemness subtypes, GSVA enrichment analysis was implemented to assess the cluster-specific signaling pathways. We found that pathways related to DNA damage repair, such as homologous recombination and nucleotide excision repair pathways, were enriched in the stemness subtype B group. However, the stemness subtype A group mainly correlated with immune-related pathways, including the chemokine signaling pathway, cytokine‒cytokine receptor interaction and leucocyte transendothelial migration pathway, and pathways related to the TME, such as ECM-receptor interaction, cell adhesion molecules and focal adhesion pathways (Fig. 5A). The results demonstrated that genomic differences might generate distinct immune infiltration statuses between the two stemness subtypes.

Subsequently, we performed ESTIMATE and ssGSEA analyses to investigate the immune infiltration status between the two stemness subtypes. The results of ESTIMATE analysis showed that ESTIMATE scores, immune scores and stromal scores were higher in stemness subtype A, while a higher tumor purity score was observed in stemness subtype B (Fig. 5B). ssGSEA revealed that almost all immune cell types were more abundant in stemness subtype A (Fig. 5C). Furthermore, the aforementioned immune classification showed that stemness subtype A consisted of most of the high-immunity patients (Fig. 5D). These findings suggested that the stemness subtype A group had a relatively higher immune infiltration status than the stemness subtype B group.

Herein, we further explored the expression profile of multiple immune checkpoint genes between the two stemness subtypes. The expression of most of the immune checkpoint genes, including PD1, PD-L1, CTLA-4, TIM3 and Lag3, was significantly higher in stemness subtype A (Fig. 6A), illustrating that patients with stemness subtype A might be more sensitive to ICIs. Subsequently, the IPS and TIDE algorithms were applied to estimate immunotherapy sensitivity in LUSC patients. As expected, patients with stemness subtype A had a higher IPS than patients with stemness subtype B, suggesting a positive response to both PD-1 and CTLA-4 inhibitors (Fig. 6B). The TIDE results also showed that patients in the stemness subtype A group had a lower TIDE score but higher MSI and T-cell dysfunction scores than those in the stemness subtype B group (Fig. 6C). These data indicated that patients with stemness subtype A were more likely to benefit from ICI administration.

Additionally, we further investigated whether these two stemness subtypes might respond differently to chemotherapeutics and targeted drugs. According to the pRRophetic algorithm, we found that the standard first-line chemotherapy drugs, including cisplatin, gemcitabine and vinorelbine, had lower IC50 values in stemness subtype A than in subtype B, indicating a higher chemotherapy sensitivity in patients with stemness subtype A (Fig. 7A). Epidermal growth factor receptor tyrosine kinase inhibitors (EGFR-TKIs) are an effective treatment for EGFR-mutant non-small cell lung cancer (NSCLC). We estimated the IC50 of EGFR-TKIs, which showed that the IC50 values of both first-generation and second-generation EGFR-TKIs were significantly lower in patients with stemness subtype B (Fig. 7B). However, the IC50 values of other targeted inhibitors, such as VEGFR, PARP1 and PI3K inhibitors, were lower in stemness subtype A (Fig. 7C).

Finally, we tried to develop a clinically applicable stemness subtype classifier that could easily discriminate the stemness subtype of LUSC patients. LUSC patients from the TCGA dataset were randomly split into a training set (n = 335) and a testing set (n = 143) at a 7:3 ratio. LASSO and RF analyses were applied to select the most hub genes associated with the stemness subtypes based on the expression of 190 stemness-related DEGs (Fig. S4A-B). A total of 51 and 31 stemness-related DEGs were selected by these two machine learning algorithms. After the intersection, nine hub genes were shared by these two feature selection methods (Fig. 8A). Next, multivariate logistic regression analysis was employed to develop a stemness subtype classifier by incorporating these nine genes. The formula of the stemness subtype classifier was 25.426 + (-0.473 × expression of AXL) + (-0.667 × expression of EFEMP2) + (-1.088× expression of VIM) + (-0.1227 × expression of EHD2) + (-0.397 × expression of COL3A1) + (-0.379 × expression of FSTL3) + (-0.557 × expression of ALOX5) + (-0.552 × expression of TNFRSF12A) + (-0.229 × expression of HSPB8). The optimal cutoff score for the classifier to discriminate these two subtypes was 0.49, indicating that patients with scores < 0.49 were assigned to stemness subtype A, while the others were assigned to stemness subtype B. ROC curves illustrated that the stemness subtype classifier was very reliable in distinguishing these two stemness subtypes with an AUC of 0.967 (Fig. 8B). The sensitivity, specificity and accuracy were 88.16%, 91.26% and 89.9%, respectively, in the training set. Furthermore, the stemness subtype classifier also showed good performance in the classification of these two stemness subtypes in the testing set, with an AUC of 0.956, and the sensitivity, specificity, and accuracy were 86.15%, 91.03% and 88.9%, respectively (Fig. 8C). Principal component analysis (PCA) validated that the stemness subtype classifier could split LUSC patients into two stemness subtypes in both the training and testing sets (Fig. 8D).

To further validate the significance of our stemness subtype classifier, three GEO datasets (GSE30219, GSE37745 and GSE73403) were enrolled as the external validation set. We calculated the score of each patient based on the same formula and then classified these patients into two subgroups according to the cutoff score of the classifier (Fig. 8E-F). Patients in three GEO validation datasets were classified into 2 stemness subtypes, including 104 patients in stemness subtype A (53.1%) and 96 patients in stemness subtype B (46.9%). Similar to TCGA results, the Kaplan–Meier curve revealed that patients with stemness subtype B had a better OS than those with stemness subtype A (Fig. 8G). Taken together, these results indicated that the stemness subtype classifier was a reliable tool in discriminating the stemness subtypes and could serve as an independent predictor for the prognosis of LUSC patients.

To validate the results of the stemness subtype classifier, we used qRT‒PCR to validate the expression levels of these hub genes in clinical specimens. As shown in Fig. 9A, seven hub genes (AXL, EFEMP2, VIM, EHD2, FSTL3, ALOX5, HSPB8) were downregulated in tumor samples, while COL3A1 was upregulated in tumor samples in the TCGA dataset. We further validated these 8 genes in 10 paired LUSC tumor and adjacent normal tissues using qRT-PCR. We found that the expression levels of EFEMP2, EHD2, FSTL3 and ALOX5 were decreased in tumor tissues, whereas COL3A1 was significantly increased in tumor samples (Fig. 9B). We did not detect the expression of AXL, VIM or HSPB8. These laboratory data were consistent with the results of previous bioinformatics results, which partially supported the credibility of the classifier.