1. Introduction
The vast majority of malignant tumors have epithelial origin. These tumors are termed “carcinomas” and can be further divided into adenocarcinoma and squamous cell carcinoma. The most affected anatomical sites where squamous cell carcinoma (SCC) can occur include the skin, lung, the head and neck region, esophagus, cervix and thyroid gland. SCCs share common characteristics not only in their development but also in the course of the disease [
1]. Similarities were also documented in gene expression, miRNA expression, and mutation profiles [
2].
Worldwide, lung cancer is the most frequent tumor type with more than 2 million new cases [
3], with particularly high incidence and mortality in some Central European countries [
4]. Approximately 30% of lung cancer cases originate in squamous cells (LUSC) [
5]. In the head and neck region tumors can develop in the lips, oral cavity, oropharynx, sinonasal cavities, larynx, hypopharynx, and salivary glands. Among these, head and neck squamous cell carcinoma (HNSC) is the most common type [
6]. Cervical cancer is the fourth most common cancer in women, with higher incidence in low- and middle-income countries [
7]. A common feature of these SCC tumors is the utilization of platinum compounds in their systemic chemotherapy. According to the protocols of the National Comprehensive Cancer Network [
8], this can be based on either cisplatin or carboplatin.
The response rate of platinum based treatment varies by tumor type, and the mean proportion of responder is between 20 and 40% in non-small cell lung cancer [
9], cervical cancer [
10] and in head and neck cancer [
11]. Another problem is that secondary resistance can develop in patients who were responders for the first line platinum treatment by the selection of resistant clones during treatment [
12].
MicroRNAs (miRNAs) are small noncoding RNAs with a length of 22–25 nucleotides having a crucial role in posttranscriptional gene regulation. Changes in miRNA expression were linked to the pathogenesis of a wide range of diseases [
13]. In cancer, affected miRNAs can regulate for example key cancer hallmark features including signaling pathways of cell proliferation, cell cycle and apoptosis. In addition, miRNAs can also suppress the expression of wild-type tumor suppressor genes [
14]. Altered miRNA expression can also modulate therapy response. For example, miR-128 and miR-155 have been reported as key miRNAs related to platinum response in patients with non-small cell lung cancer [
15], miR-200b and miR-155 were predictive biomarkers for the efficacy of chemoradiation in head and neck cancer patients [
16] and a total of 25 differentially expressed miRNAs have been linked to response to acetoxychavicol acetate (ACA) and/or cisplatin in human cervical carcinoma cells [
17].
The aim of this study was to identify miRNAs which could serve as predictive biomarkers in platinum-treated SCCs. Our secondary goal was to detect overlapping miRNA expression patterns in platinum-treated head and neck, cervical and lung squamous cell carcinoma samples.
2. Results
Overall, 1309 (CESC: 307, HNSC: 524, LUSC: 478) patients with squamous cell carcinomas were identified in the GDC data portal. From these patients, we excluded those with insufficient clinical data or without a platinum-based (cisplatin or carboplatin) chemotherapy. Finally, altogether 266 patients were retained for the analysis: 94 patients in the CESC subgroup (16 nonresponder and 78 responder), 105 patients in the HNSC subgroup (34 nonresponder and 71 responder) and 67 patients in the LUSC subgroup (16 nonresponder and 51 responder). Aggregate clinical characteristics of the samples are presented in
Table 1.
The expression dataset contains information for 1881 miRNAs. From these, we had retained 599 miRNAs with non-zero expression in the more than 50% of the analyzed samples.
2.1. miRNAs Associated with Treatment Outcome in Each SCC Cohort
Sixteen differentially expressed miRNAs were identified in the CESC cohort. The top three miRNAs were hsa-miR-342 (Mann–Whitney test
p value: 9.68 × 10
−5, AUC: 0.811), hsa-miR-378c (
p value: 1.29 × 10
−4, AUC: 0.805) and hsa-miR-155 (
p value: 2.01 × 10
−4, AUC: 0.796) (
Table 2 and
Figure 1A).
In the HNSC cohort 103 differentially expressed miRNAs associated with treatment outcome. The top three miRNAs were hsa-miR-326 (
p value: 9.24 × 10
−6, AUC: 0.768), hsa-miR-584 (
p value: 2.02 × 10
−5, AUC: 0.758) and hsa-miR-19a (
p value: 4.66 × 10
−5, AUC: 0.742) (
Table 3 and
Figure 1B).
There were nine differentially expressed miRNAs in the LUSC samples. The top three miRNAs were hsa-miR-130a (
p value: 1.61 × 10
−3, AUC: 0.763), hsa-miR-26b (
p value: 2.64 × 10
−3, AUC: 0.751) and hsa-miR-6781 (
p value: 1.87 × 10
−3, AUC: 0.744) (
Table 4 and
Figure 1C).
When investigating all three tumor types, there were in total 128 miRNAs with statistically significant differences in expression between nonresponder and responder cohorts in at least one subgroup. Out of these significant miRNAs, 11 miRNAs had significant differences in at least two subgroups: three in the LUSC and HNSC subgroups (hsa-miR-130a, hsa-miR-15a, and hsa-miR-877) and eight in the CESC and HNSC subgroups (hsa-miR-142, hsa-miR-150, hsa-miR-16-1, hsa-miR-181b-1, hsa-miR-378a, hsa-miR-378c, hsa-miR-378d-2, and hsa-miR-5586). We could not find a miRNA significant in both the LUSC and CESC subgroups.
2.2. miRNA Expression Comparison between Tumor Types
Based on the log2 expression fold change between nonresponder and responder cohorts the samples were divided into two clusters. CESC and HNSC formed one cluster and LUSC samples were separated into a different node. The similarity score was 0.31 between CESC and HNSC and 0.26 between LUSC and HNSC samples (
Figure 2). The correlation between HNSC and LUSC was much lower (0.13).
2.3. Logistic Regression Based Classification Model
We combined the HNSC and CESC samples to have a sufficient sample number for the development of a classification model capable of predicting chemotherapy response. The selection of HNCS and CESC was based on their higher similarity scores (due to the differences between HNSC and LUSC the overall model did not improve when including LUSC samples as well). The samples were randomly divided into a training (70%) set and a test set (30%) multiple times. Thirty differentially expressed miRNA were identified in the HNSC and CESC merged samples, with an AUC over 0.66 and an FDR below 0.05.
In the second step a stepwise logistic regression with forward selection was performed to estimate the predictive power. After this second feature selection we retained nine miRNAs with the strongest link to treatment response. The accuracy of the model was 0.88 (95% CI: 0.81–0.93), the sensitivity was 0.69, and the specificity was 0.94 in the training set. In the test set the accuracy was 0.81 (95% CI: 0.69–0.90), the sensitivity was 0.60, and the specificity was 0.89. Overall, the 10-fold cross-validated model accuracy was 0.85 (95% CI: 0.79–0.90), the sensitivity was 0.60, and the specificity was 0.94.
In the final step we simultaneously included all nine miRNAs with a significant impact. The significant remaining miRNAs were hsa-miR-5586 (Exp (B): 2.94, 95% CI for exp[B]: 1.65–5.51,
p = 0.001), hsa-miR-632 (Exp (B): 10.75, 95% CI for exp[B]: 2.34–49.37,
p = 0.002), hsa-miR-2355 (Exp (B): 0.48, 95% CI for exp[B]: 0.29–0.79,
p = 0.004), hsa-miR-642a (Exp (B): 2.22, 95% CI for exp[B]: 1.21–4.07,
p = 0.01), hsa-miR-101-2 (Exp (B): 0.39, 95% CI for exp[B]: 0.18–0.82,
p = 0.013), and hsa-miR-6728 (Exp (B): 0.21, 95% CI for exp[B]: 0.06–0.75,
p = 0.016) (
Figure 3). In this final model hsa-miR-181b-2, hsa-miR-26-2, and hsa-miR-584 were not significant (
Table 5).
2.4. Target Gene Prediction
Since some miRNAs were over-expressed and other under-expressed, the overall effect of the combination of these could neutralize each other. For this reason we investigated the over- and under-expressed miRNAs separately. The analysis of miRNAs overexpressed in nonresponder group revealed that hsa-mir-2355 and hsa-mir-6728 have two common predicted target genes which are repeated in different KEGG pathways, ENTPD5 and NT5C3A. Both genes take parts in purine (hsa00230,
p = 6.47 × 10
−10) and pyrimidine metabolism (hsa00240,
p = 5.20 × 10
−8) and NTC3A has a role in nicotinate and nicotinamide metabolism (hsa00760) and other metabolic pathways (hsa01100) as well. The underexpressed miRNAs (hsa-mir-5586, hsa-mir-642a, hsa-mir-101, hsa-mir-632) have nine common predicted target genes in four KEGG pathways. Most of the predicted genes are overlapped between pathways (
Table 6)
3. Discussion
In our study our aim was to explore the miRNAome of platinum-treated head and neck, cervical and lung squamous cell carcinoma samples. First, we determined the most significant therapy-response related miRNAs in each SCC cohort. In CESC the most significant gene was hsa-miR-342. Hsa-miR-342 is known to be downregulated in cervical cancer tissues and cell lines and suppressed proliferation, growth, invasion and migration in human cervical cells [
18]. In the HNSC samples hsa-miR-326 had the best discriminatory ability between responder and nonresponder cohorts. hsa-miR-326 was previously characterized as a tumor suppressor in breast [
19] and colorectal cancer [
20]. In our analysis responder samples had significantly elevated expression of hsa-miR-326 compared to nonresponder samples. Finally, in the LUSC subgroup hsa-miR-130a was the most significant miRNA. Elevated expression of hsa-miR-130a increased resistance to platinum chemotherapy [
21]. Our results are in line with these previous findings, as we also observed higher expression in the nonresponding cohort.
We examined the overlapping genes, e.g., genes significant in multiple cohorts. In HNSC and CESC groups we uncovered 8 common miRNAs with significant discriminatory ability between responder and nonresponder cohorts, whereas we found only three common significant miRNAs in the LUSC and HNSC samples. A potential explanation for the higher number of overlapping miRNAs in the HNSC and CESC group can be that HPV infection has a role in the etiology of both tumor types. We have found 4 significant miRNAs (hsa-miR-16, hsa-miR-145, hsa-miR-199b) which have been previously reported as HPV core miRNAs in HNSC and CESC clinical samples [
22].
The eight overlapping miRNAs between HNSC and CESC groups were previously discussed as genes related to cancer. High plasma level of hsa-mir-142 were reported as a HPV-independent prognostic marker of combined radio-chemotherapy in HNSC patients [
23], and its expression correlated with the inhibition of cell proliferation and chemoresistance in ovarian cancer cell lines [
24]. Induced overexpression of hsa-mir-150 inhibited cell invasion and metastasis in ovarian cancer [
25]. Inhibited proliferation and enhanced therapeutic effect of cisplatin was reported in osteosarcoma for hsa-mir-16 [
26]. hsa-mir-181 was overexpressed in cisplatin resistant NSCLC cells [
27], and correlated with worse survival in oral squamous cell carcinoma patients [
28]. Finally, higher expression of hsa-mir-378 reversed chemoresistance to cisplatin in NSCLC cells [
29]. In our analysis, all of these miRNAs except for hsa-mir-181b-1 were overexpressed in responders.
The three significant miRNAs shared between LUSC and HNSC (hsa-miR-130a, hsa-miR-15a, hsa-miR-877) are all upregulated in nonresponder samples. hsa-miR-130a was upregulated in esophageal squamous carcinoma cell lines [
30] and in clear cell carcinoma of ovary patients and the overexpression of the miRNA was a biomarker for disease recurrence [
31]. hsa-miR-130a was also elevated in cisplatin resistant ovarian cell lines [
32]. hsa-miR-15 was the first miRNA linked to tumor suppression in cancer [
33]. As a contrary, it has been reported that overexpression of hsa-miR-15 predicts poor disease-free survival and overall survival in colorectal cancer patients [
34], and it promotes neuroblastoma migration [
35]. hsa-miR-877 was reported as oncogene in gastric cancer [
36], but another study found that it has a tumor suppressor role in prostate cancer [
37]. Our result suggests that the higher expression of hsa-miR-130a, hsa-miR-15a, and hsa-miR-877 are all biomarkers of resistance.
In the last step we determined a miRNA signature capable of predicting the outcome of platinum-based chemotherapy in HNSC and CESC samples. Our model reached a high accuracy in the 10-fold cross-validation of 0.854. The final model includes six miRNAs with the strongest effect—the most significant of these was hsa-miR-5586, which was studied in diffuse large B-cell lymphoma patients and the elevated expression was associated with better outcome [
38]. In our study the expression of mir-5586 was elevated in responder samples both in HNSC and CESC subgroups. The second most important variable in our model was hsa-miR-632 and responders had higher expression. Earlier, its upregulation was associated with better survival in colorectal cancer [
39] and its downregulation had an inhibitory effect on hepatocellular carcinoma cell proliferation and invasion [
40]. The expression of hsa-miR-2355 was higher in nonresponder samples of HNSC and CESC subgroups. In esophageal squamous carcinoma cell line Zhang et al. reported that overexpressed hsa-miR-2355 promoted cell proliferation and invasion [
41].
In the last step target gene prediction was performed with DIANA-Tool for six miRNAs significant in the logistic regression model. For the hsa-mir-2355 and the hsa-mir-6728 we determined ENTPD5 and NT5C3A genes as potential targets. The ENTPDT gene was examined in numerous studies and association between higher expression of ENTPDT and chemotherapy response were reported in colorectal cancer cells [
42] and prostate cancer [
43]. In lung cancer cells decreased apoptosis rate was observed after knockdown of ENTPDT [
44]. Interestingly, the most significant pathway related to genes targeted by the four miRNAs with lower expression (hsa-mir-5586, hsa-mir-642a, hsa-mir-101, hsa-mir-632) was the circadian entrainment pathway with six potential target genes. This pathway was demonstrated to be associated with platinum treatment resistance in cancer [
45].
There is an important limitation of our study: we had only a restricted number of samples available for this analysis. Not only is the total number of SCC samples published low, but an important further restriction was the administration of a platinum-based protocol. Although we were able to validate the result of our final model using separate test/training cohorts in a ten-fold cross-validation, for the generalization of our results it will be necessary to repeat the analysis in the future using more clinical samples.