Preliminary study using a small plasma extracellular vesicle miRNA panel as a potential biomarker for early diagnosis and prognosis in laryngeal cancer

Abstract

Purpose

Plasma extracellular vesicle (EV) miRNAs are important biomarkers for body fluid biopsy. The purpose of this study was to screen and construct a plasma small EV (sEV) miRNA panel as a biomarker for diagnosis and prognosis in laryngeal squamous cell carcinoma (LSCC).

Methods

Plasma sEV miRNAs from 6 LSCC patients with three typical anatomical sites and 3 normal controls (NCs) were analyzed by next-generation sequencing. The aberrant expression profile of sEV miRNAs was compared with the online databases of LSCC to construct and verify the diagnostic and prognostic panel by machine learning. Additionally, quantitative real-time polymerase chain reaction (qRT‒PCR) was performed to validate the diagnostic efficacy of the screened miRNAs in an independent clinical cohort.

Results

A plasma sEV miRNA panel (consisting of hsa-miR-139-3p, hsa-miR-486-5p, hsa-miR-944, hsa-miR-320b and hsa-miR-455-5p) was successfully constructed for the early diagnosis and prognosis of LSCC and showed good predictive potential with AUCs of 0.782, 1.000, 0.716, and 0.875 by an artificial neural network (ANN) panel in independent datasets. This panel was further validated in an independent cohort consisting of 84 clinical cases (48 LSCC and 36 NCs). In the validation cohort, the AUC of the 5 individual miRNAs ranged from 0.721 to 0.837. The accuracy was further increased by the logistic model, which further increased the AUC to 0.959 by adjusting for the number of miRNAs. The miRNA‒mRNA regulatory network and immune function analysis revealed the possible underlying pathogenesis of LSCC.

Conclusion

Plasma sEV miRNA panels can be promising plasma biomarkers for predicting early diagnosis and prognosis in LSCC.

Supplementary information

The online version contains supplementary material available at 10.1007/s13402-023-00792-y.

Introduction

Laryngeal squamous cell carcinoma (LSCC) is one of the most common malignant tumors in otolaryngology, with an incidence of approximately 1.86/100,000 in China [1]. Approximately 40% of LSCC patients are diagnosed with stage III/IV disease because of the occult tumor site, leading to the expansion of surgical scope and poor prognosis [2]. Early diagnosis of LSCC greatly improves the survival rate. However, there is still no effective method for early screening and diagnosis. Although many diagnostic approaches, including fiberoptic endoscopy, computed tomography (CT), magnetic resonance imaging (MRI), positron emission tomography (PET) scan, p16 immunohistochemistry (IHC), and tumor programmed death ligand 1 (PD-L1), have been extensively utilized [3–5], their clinical benefits are still limited by high cost, inconvenience, invasiveness or uncertainty. As a result, noninvasive biomarkers with easy access, high sensitivity and specificity are urgently needed for early diagnosis [6].

Benefiting from their noninvasive, convenient, and efficient characteristics, plasma extracellular vesicle (EV) detection has been increasingly explored in the early diagnosis of cancer, among which miRNAs are more stable because of their specific sorting mechanism [7] and may offer an efficient, noninvasive or minimally invasive marker alternative. Recently, plasma EV miRNAs have been shown to be early diagnostic markers for gastric cancer [8, 9], epithelial ovarian cancer [9], pancreatic cancer [10], and lung adenocarcinoma [11], and they can be used to assess treatment response and prognosis in esophageal squamous cell carcinoma [12], non-small cell lung cancer [13], breast cancer [14], etc. In LSCC, serum EV hsa-miR-941 has been identified as a promising oncogenic marker, and upregulated levels of hsa-miR-941 promote cell proliferation and invasion [15]. In addition, serum EV hsa-miR-21 can also be used as a valuable diagnostic and prognostic biomarker for LSCC [16]. Tumor-derived EVs are highly abundant in the plasma of HNSCC patients, nonetheless, there is still a lack of studies regarding the relationship between plasma EV miRNAs and LSCC.

In this study, we collected peripheral blood specimens from 6 LSCC patients and 3 normal controls (NCs) and performed miRNA differential analysis by next-generation sequencing. We further screened and built a plasma small EV (sEV) miRNA diagnostic panel, combining the miRNA expression profiling data of LSCC and normal tissue samples from The Cancer Genome Atlas (TCGA) database, and verified it through plasma sEV samples and independent datasets from the Gene Expression Omnibus (GEO) database. Furthermore, we constructed a miRNA‒mRNA regulatory network, performed functional enrichment analysis, and analyzed the prognostic predictive value of plasma sEV miRNAs for LSCC. Finally, quantitative real-time polymerase chain reaction (qRT‒PCR) was conducted to measure the expression level of the screened miRNAs in an independent clinical cohort. In brief, our results indicated that plasma sEV miRNAs may be used for early diagnosis and prognosis evaluation of LSCC.

Materials and methods

Study design

Figure 1 depicts all phases of this study. Six LSCC patients with three typical anatomical sites and 3 patients with vocal cord polyps were enrolled, and 5 miRNAs were then identified and validated in the TCGA and GEO databases. All LSCC patients were treatment-naïve upon enrollment in the study. The differentially expressed miRNAs (DEMs) were intersected with the DEMs obtained from the TCGA database, and 8 DEMs were screened for subsequent analysis. Five DEMs were screened by random forest (RF), and the gene score was calculated. Furthermore, the TCGA database was used to build a diagnostic panel through an artificial neural network (ANN), and then plasma sEV samples and other independent datasets from the GEO database were used to verify the accuracy. Then, the CIBERSORT algorithm was used to analyze immune cell infiltration in patients with LSCC and NCs. Furthermore, we verified the expression of the above 5 miRNAs through the CancerMIRNome database [17]. After evaluation of the diagnostic efficiency of hsa-miR-455-5p, hsa-miR-320b, hsa-miR-944 hsa-miR-139-3p and hsa-miR-486-5p, functional exploration and prognostic correlation of the 5 verified miRNAs were also conducted, and a prognostic model based on the above 5 miRNAs was also successfully constructed by least absolute shrinkage and selection operator (LASSO) Cox regression analysis, demonstrating the value of this miRNA panel in the prognosis of LSCC. Finally, quantitative real-time polymerase chain reaction (qRT‒PCR) was performed to measure the expression level of the screened miRNAs in an independent clinical cohort consisting of 48 LSCC and 36 NC samples.

Fig. 1.

Flow chart of the study

Patient information and ethical approval

Peripheral blood specimens were collected from 6 LSCC patients (glottic, supraglottic, and subglottic) diagnosed by pathology. Three patients with vocal cord polyps served as controls. All samples were collected from the Eye, Ear, Nose and Throat Hospital, Fudan University between January 2021 and October 2021. Clinicopathologic data of the patients were also collected (Table 1).

Table 1.

Clinical features of 6 LSCC patients

	Cancer (n = 6)
Age at surgery(years)
≤ 65	2(33.3%)
> 65	4(66.7%)
Sex
Male	6(100.0%)
Female	0(0.0%)
Clinical stage
I-II	1(16.6%)
III-IV	5(83.3%)
Site
Supraglottic	2(33.3%)
Glottic	2(33.3%)
Subglottic	2(33.3%)
Lymph node metastasis
Negative	2(33.3%)
Positive	4(66.7%)
Histology grade
SCC I-II	6(100.0%)
SCC III	0(0.0%)

All specimens were obtained with the approval of the Ethics Committee of the Eye, Ear, Nose and Throat Hospital, Fudan University, Shanghai, China. Signed informed consent was obtained from all participants included in this study.

Isolation and identification of plasma sEVs

Collected blood samples were stored in heparinized vacuum tubes prior to being centrifuged at 4 °C (3000 g, 15 min) and at 4 °C (12,000 g, 10 min) to remove cell debris. The Ribo™ Exosome Isolation Reagent (for plasma or serum) (RiboBio Co., Ltd., Guangzhou, China) was used to isolate plasma sEVs in accordance with the operating instructions. Isolated sEV pellets were identified by transmission electron microscopy (TEM), nanoparticle tracking analysis (NTA) and western blotting, as described in our previous study [18]. As previously described, TEM images showed the typical size range and morphology of membrane‑bound sEV particles, which were homogeneous in appearance. NTA analysis showed a size distribution (mean diameter at 116.8 ± 35.9 nm) for the sEV particles. Western blotting revealed positive expression of CD63 and TSG101, two well-established surface markers for sEV. Due to their size distribution, sEVs were named based on the Nomenclature in the MISEV2018 guidelines [19].

Extraction of sEV RNAs and sequencing of DEMs

The Magen HiPure Serum/Plasma miRNA Kit (RiboBio) was used to extract high-quality total RNA from sEVs obtained from each specimen in accordance with the operating instructions. sRNA next-generation sequencing technology was used to detect the differences in the miRNA profiles between plasma sEVs from LSCC and control patients. The sRNA library preparation and sample sequencing were performed by RiboBio Co., Ltd. (Guangzhou, China) using an Illumina HiSeq™ 2500 device, as described previously [20]. The differential expression analysis was performed using the “EdgeR” package of R software version 4.0.1 with appropriate correction (| log2(fold change) | >1, P < 0.05) for multiple testing.

Identification of candidate DEMs in combination with TCGA database

The miRNA expression data of LSCC (150 samples of LSCC and 14 normal samples, type: STAR - Counts) was obtained from the TCGA database, and the miRBase website (http://www.mirbase.org/) was used to retrieve all mature miRNA (mature.fa) sequences. Using Perl language, the ID of each miRNA was converted to the mature form. Then, the DEMs were screened using the “limma” package in R software and visualized in the heatmap plotted with the “pheatmap” package.

Construction, validation, and evaluation of the candidate miRNA diagnostic panel

The DEMs of the plasma sEV samples and TCGA database were intersected. Then, the “randomForest” package of R software was applied to further screen out DEMs that contributed the most to the diagnosis of LSCC patients. Genes with an importance score greater than 2 according to the Gini coefficient method were selected as candidate DEMs for subsequent panel building. Subsequently, the expression data of the candidate DEMs were converted into a score table named “Gene Score”. The specific conversion rules were as follows: Expression values of DEMs were compared to the median of all sample expression values, with a value of 1 if the expression value of the upregulated DEM was greater than the median and 0 otherwise. Similarly, a downregulated DEM had a value of 0 if its expression value was greater than the median value and 1 otherwise.

The ANN panel was constructed using the “neuralnet” and “neuralnettools” packages with the disease attributes set to y and the gene score set to X. The ANN consisted of an input layer, a hidden layer, and an output layer with 5 hidden nodes included in the hidden layer and 2 nodes included in the output layer (LSCC and normal). The LSCC classification panel was constructed by the gene weight information obtained in the ANN panel. The weight score was multiplied by the sum of the products of the specific DEM expression levels as the disease classification score.

Time receiver operating characteristic (ROC) curve analysis was conducted using the “time ROC” package, and the area under the curve (AUC) was calculated to evaluate the model’s diagnostic ability. Then, plasma sEV samples and other independent datasets (38 LSCC samples and 5 normal samples from GSE79493, GSE106280 and GSE137308) from the GEO database were selected as the external validation sets to validate the panel accuracy. The raw data were downloaded from the GEO database and then processed and standardized using Perl language. The “limma” and “sva” packages of R software were used to remove the batch effect of the 3 datasets for further analysis. The classification scores of specific DEMs were calculated according to the ANN panel, and then the AUC was calculated to verify the classification efficiency.

In addition, we also tested the diagnostic efficiency of a single miRNA for LSCC in the TCGA database and plasma sEV samples and evaluated the expression of 5 miRNAs in LSCC and normal tissues through the CancerMIRNome database.

Target gene prediction and functional enrichment analysis

To predict the potential biological functions of the miRNA panel, we used the online analysis tools TargetScan, miRWalk, miRTarBase and miRDB to predict miRNA target genes. These are all currently mainstream miRNA related interaction (MTIs) databases [21]. Then, the “VennDiagram” package of R software was used to identify the overlapping target genes. If there were three or more target gene analysis tools supporting the gene as the target gene, we hypothesized that the gene might be the target gene of the miRNA. Then, the target genes were intersected with plasma sEV differentially expressed mRNAs, and the miRNA‒mRNA regulatory network was constructed using Cytoscape 3.9.1. Furthermore, gene ontology (GO) analysis and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis were performed to obtain the enrichment of candidate target genes related to biological processes, molecular functions, cellular components and signaling pathways. The protein‒protein interaction (PPI) network of target genes was obtained from STRING online software to identify the core genes. In addition, the CIBERSORT algorithm was used to filter 22 kinds of immune cells, which was a digital cytometry which enables estimation of cell type abundances from bulk tissue transcriptomes without cell separation or prior knowledge [22]. Then the “vioplot” package was used to visualize the proportions of infiltrating immune cells between patients with LSCC and healthy controls.

Construction and validation of the DEM prognostic model

The expression levels of DEMs in LSCC samples from the TCGA database were combined with survival information to determine whether DEM expression might be associated with prognosis. After excluding patients without survival data, LSCC cases from 150 patients were analyzed using Lasso Cox regression analysis with the “glmnet” package of R software to construct a prognostic model. The penalty parameter (λ) was determined using the minimum standard, and the risk score of the DEM prognostic model was computed with the formula ΣX_iY_i, where X represents the expression level of each miRNA, and Y represents the corresponding regression coefficient. According to the median risk scores calculated after centralizing and normalizing the TCGA expression data, patients with LSCC were divided into high- and low-risk groups. Kaplan‒Meier survival curves were drawn using the “Survival” and “survminer” packages of R software, and principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) analysis were performed using the “Rtsne” package and “ggplot2” package, respectively. To evaluate the model’s prognostic ability, ROC analysis was performed using the “time ROC” package, and the AUC was calculated at 1, 3, and 5 years. Finally, the clinical data of LSCC patients were extracted from the TCGA database, and the risk score of the prognostic model was analyzed using univariate and multivariate Cox regression analyses to explore the predictive ability of the prognostic model for clinical characteristics.

Validation of the expression of 5 DEMs in an independent clinical cohort

Plasma sEV samples were obtained from patients at the Eye & ENT Hospital of Fudan University. All patients (48 LSCC and 36 NC) signed an informed consent form, and the research was approved by the Ethics Committee of the Eye & ENT Hospital of Fudan University. All 36 NC patients were male and diagnosed as vocal cord polyps. The AJCC eighth edition of the TNM staging system was used. Clinicopathologic data of the 48 LSCC patients were shown in Supplementary Table 1. Total RNA was isolated from plasma sEVs with TRIzol reagent (Invitrogen, Thermo Fisher Scientific) before being reverse-transcribed using a miRNA 1st strand cDNA synthesis kit (AG11717, Accurate Biology, Hunan, China). The primers of five candidate miRNAs were designed and synthesized by Sangon Biotech (Shanghai). The sequences of the primers used are listed in Table 2. qRT‒PCR was performed using the SYBR® Green Premix Pro Taq HS qPCR Kit (AG11718, Accurate Biology) with the ABI 7500 Real-Time PCR System (Life Technologies, Shanghai, China) using the 2^-ΔΔCt method.

Table 2.

Sequences of miRNA primers used for qRT‒PCR

miRNA name	Primer sequence (5’→3’)
hsa-miR-139-3p-F	CGGCCCTGTTGGAGTAAAAA
hsa-miR-320b-F	AAAAGCTGGGTTGAGAGGGCAA
hsa-miR-455-5p-F	TATGTGCCTTTGGACTACATCG
hsa-miR-486-5p-F	TACTGAGCTGCCCCGAGAA
hsa-miR-944-F	CGGGATTATTGTACATCGGATGAG
U6-F	CTCGCTTCGGCAGCACA
U6-R	AACGCTTCACGAATTTGCGT

Abbreviations: qRT‒PCR, quantitative real-time polymerase chain reaction; F, forward primer; R, reverse primer; U6, U6 small nuclear RNA

Statistical analysis

All statistical analyses were accomplished with Perl language, R software 4.1.0 and GraphPad Prism 8.0.2. The diagnostic accuracy of the diagnostic panel and miRNAs was assessed by ROC curve analysis, and the AUC was also calculated. The prognostic factors were determined using univariate and multivariate Cox regression analyses. The prognostic risk model was constructed using Lasso Cox regression analysis. Measurement data were expressed as the mean ± standard deviation ($\stackrel{-}{x}$ ± s) and that which met normal distribution and homogeneity of variance were analyzed by one-way ANOVA. Counting data were analyzed by χ2 test. All tests were two-tailed, and the false discovery rate (FDR) was controlled for multiple comparisons. P < 0.05 was considered significant.

Results

Laryngoscopy and contrast-enhanced CT

Laryngoscopy and enhanced CT showed the typical clinical features of 6 LSCC patients (Fig. 2).

Fig. 2.

Clinical features of 6 LSCC patients revealed by laryngoscopy and enhanced CT. a, b Two patients with glottic LSCC. c, d Two patients with supraglottic LSCC. e, f Two patients with subglottic LSCC

Aberrant expression profiles of plasma sEV miRNAs in LSCC patients

Next-generation sequencing technology revealed 136 DEMs between plasma sEVs from LSCC and control patients. Aberrant expression profiles of sEV miRNAs were illustrated by scatter and volcano maps (Fig. 3a, b). The scatter plot revealed the differential expression of sEV miRNAs, and the volcano plot revealed the same expression profiles of sEV miRNAs. Of those, 93 were upregulated and 43 were downregulated (|log2(fold change) | >1, P < 0.05). To verify the accuracy of the sequencing results, qRT‒PCR was used to detect the expression of hsa-miR-139-3p, hsa-miR-486-5p, hsa-miR-944, hsa-miR-320b and hsa-miR-455-5p in plasma samples. As anticipated, the results of qRT‒PCR detection revealed that the expression levels of these selected sEV miRNAs were consistent with the sequencing results (Supplementary Fig. 1). These findings indicated that the results of the small RNA sequencing were reliable.

Fig. 3.

Investigation of the expression levels of 38 DEMs and construction of the diagnostic panel based on 8 DEMs. a Scatter plot of all DEMs in the plasma sEV samples. Green and red spots represent downregulated and upregulated genes, respectively. b Volcano plot of all DEMs in the plasma sEV samples. c Venn diagram of 38 critical DEMs in plasma sEV samples and samples from the TCGA database. d Heatmap for critical DEMs shows 29 upregulated and 9 downregulated in the plasma sEV samples. e The influence of the number of decision trees on the error rate. The x-axis represents the number of decision trees, and the y-axis indicates the error rate. f The importance of 8 DEMs ranked by mean decrease Gini of accuracy. g Heatmap for 5 DEMs in the TCGA database. h Results of ANN visualization

Identification of critical DEMs combined with TCGA database

A total of 150 LSCC samples and 14 normal samples were obtained from the TCGA database, and 246 DEMs were obtained through differential analysis. After intersecting with plasma sEV DEMs, 38 DEMs were identified (Fig. 3c, Supplementary Table 2), and their expression levels in plasma sEV samples are shown in Fig. 3d.

Random forest and ANN-based establishment of the diagnostic panel

Based on previous studies, we selected 8 miRNAs with consistent trends for subsequent analysis. To find the optimal parameters, circular random forest classification of the 8 miRNAs was performed using the random forest algorithm, and the average error rate was calculated. Then, the panel with the best stability was selected for subsequent analysis. Five miRNAs with significance greater than 2 according to the Gini coefficient were identified as candidates (Fig. 3e). Of the 5 genes, hsa-miR-139-3p, hsa-miR-486-5p, and hsa-miR-944 were most important, followed by hsa-miR-320b and hsa-miR-455-5p (Fig. 3f). In the heatmap (Fig. 3g), 5 candidate miRNAs could be used to distinguish LSCC from normal samples, among which hsa-miR-139-3p and hsa-miR-486-5p were expressed at low levels in LSCC and highly expressed in normal tissues, and hsa-miR-944, hsa-miR-320b and hsa-miR-455-5p were highly expressed in LSCC tissues.

Then, the ANN panel was constructed based on candidate DEMs from random forest screening (Fig. 3h). The expression data of the 5 DEMs were first converted into a “gene score”, and then a neural network algorithm was used to calculate the weight value of each DEM. The panel classification score was calculated by “gene score” and “gene weight”, with the following formula: LSCC ANN panel score = ∑ (gene expression level × gene weight). Next, the panel score of 154 samples was set as the predictive value, and the LSCC and normal samples were set as the outcome value. The classification ability of the panel was validated using the “pROC” package, and it was found that the AUC of the model was 0.782, suggesting that the model had an accurate predictive effect (Fig. 4a).

Fig. 4.

Validation of the diagnostic panel and individual miRNAs as biomarkers of early LSCC. Validation of the diagnostic panel as a biomarker of early LSCC in the training group (a, samples from the TCGA database), test group 1 (b, plasma sEV samples) and test group 2 (c, samples from GEO database). Validation of hsa-miR-139-3p (d), hsa-miR-486-5p (e), hsa-miR-455-5p (f), hsa-miR-944 (g), and hsa-miR-320b (h) as biomarkers of early LSCC in the training group (samples from TCGA database), test group (miRNA sequencing of 9 samples) and validation group (miRNA qRT‒PCR of 84 samples). Validation of miRNA expression of hsa-miR-139-3p (i), hsa-miR-486-5p (j), hsa-miR-455-5p (k), hsa-miR-944 (l), and hsa-miR-320b (m) in LSCC and normal samples from the CancerMIRNome database

Validation of the diagnostic panel

To further verify the panel accuracy, plasma sEV samples and another dataset (38 LSCC samples and 5 normal samples from GSE79493, GSE106280 and GSE137308) were used to verify whether the panel had the same predictive ability in the independent dataset. Using the same method, the “gene score” and “gene weight” of the 5 DEMs were calculated, and then the scores were calculated by the ANN panel and verified by the ROC curve. The AUC in plasma sEV samples was 1.000 (Fig. 4b), and the AUC in the dataset from the GEO database was 0.716 (Fig. 4c), which further proved the validity and stability of the panel. In addition, we also verified the diagnostic value of individual miRNAs from the diagnostic panel using data from the TCGA as the training group, sequencing data from plasma sEV samples as the test group and PCR results as the secondary validation. The AUC of hsa-miR-139-3p was 0.957 in the training group, 0.944 in the test group and 0.944 in the validation group (Fig. 4d); that of hsa-miR-486-5p was 0.853 in the training group, 1.000 in the test group and 1.000 in the validation group (Fig. 4e); that of hsa-miR-455-5p was 0.816 in the training group, 0.889 in the test group and 0.944 in the validation group (Fig. 4f); that of hsa-miR-944 was 0.713 in the training group, 0.750 in the test group and 0.889 in the validation group (Fig. 4g); and that of hsa-miR-320b was 0.847 in the training group, 0.667 in the test group and 0.667 in the validation group (Fig. 4h). In addition, we verified the expression of miRNAs through the CancerMIRNome database, and the results showed that the 5 miRNAs were differentially expressed in LSCC and normal tissues, which was consistent with the sequencing results (Fig. 4i-m).

Functional enrichment analysis of the DEMs

The DEMs were separately input into the online miRNA target gene prediction websites TargetScan, miRWalk, miRTarBase and miRDB for target gene prediction, and the candidate target genes were considered those predicted by three or more databases that intersected (207 target genes of hsa-miR-139-3p, 134 target genes of hsa-miR-320b, 389 target genes of hsa-miR-455-5p, 324 target genes of hsa-miR-486-5p and 1359 target genes of hsa-miR-944). Figure 5a-e). In total, 2207 target genes were obtained after removing 206 duplicate genes jointly regulated by multiple miRNAs. The intersection of target mRNAs for upregulated miRNAs (hsa-miR-455-5p, hsa-miR-320b, hsa-miR-944) and downregulated mRNAs, and target mRNAs for downregulated miRNAs (hsa-miR-139-3p, hsa-miR-486-5p) and upregulated mRNAs were taken. The results were performed on a total of 61 genes, including 4 upregulated genes and 57 downregulated genes. The miRNA‒mRNA regulatory network between the 5 miRNAs and their 61 target genes is shown in Fig. 5f.

Fig. 5.

Target gene prediction and functional enrichment analysis of DEMs. Target gene prediction of hsa-miR-139-3p (a), hsa-miR-320b (b), hsa-miR-455-5p (c), hsa-miR-486-5p (d), and hsa-miR-944 (e). f The miRNA‒mRNA regulatory network between the 5 miRNAs and their 61 target genes. g GO analysis of target genes. Biological process (BP), cellular component (CC), and molecular function (MF). h KEGG enrichment analysis of target genes corresponding to DEMs. i PPI network of the relationship between target genes. j Violin diagram of the proportions of 22 immune cell types between LSCC and healthy controls by CIBERSORT

To explore the functions and pathways associated with the target genes, we performed GO and KEGG pathway analyses. GO analysis showed that target genes were mainly enriched in negative regulation of protein phosphorylation, negative regulation of protein kinase activity, cell leading edge, distal axon, cadherin binding, and enzyme inhibitor activity (Fig. 5g). KEGG analysis showed that target genes were mainly enriched in complement and coagulation cascades, pancreatic secretion, peroxisomes, and diabetic cardiomyopathy (Fig. 5h). In addition, the PPI network demonstrated that ACTR3, ACTR2, YWHAZ, EIF3A, and SYNCRIP were the core genes in the network (Fig. 5i). In addition, we found significant differences in immune cell infiltration in patients with LSCC and healthy controls. Patients with LSCC had lower proportions of resting NK cells (P < 0.001), monocytes (P < 0.001), resting mast cells (P < 0.001), and eosinophils (P = 0.007) and higher proportions of plasma cells (P = 0.039), CD8 T cells (P = 0.02), follicular helper T cells (P = 0.011), M1 macrophages (P = 0.014), activated mast cells (P = 0.015), and neutrophils (P = 0.029) than healthy controls (Fig. 5j).

Prognostic analysis of the DEMs

We examined the prognostic predictive value of individual miRNAs in the diagnostic panel for patients with LSCC, and although capable of being predictive, the results were less favorable (Supplementary Fig. 2). Then, we explored the prognostic prediction ability by miRNA joint analysis. In the TCGA database, samples from patients with unknown survival times were excluded, and the remaining 150 LSCC samples were randomly divided into training and test cohorts on average. We found no significant difference between the training cohort and the test cohort (Supplementary Table 3). Then, LASSO Cox regression analysis was performed to construct the DEM-based risk score model, and cross-validation was performed to avoid overfitting to construct a 5-miRNA prognostic model based on the optimum λ value (Fig. 6a, b). The risk score was calculated with the established formula: risk score = (0.1528×hsa-miR-139-3p exp.) + (0.1133×hsa-miR-486-5p exp.) + (0.5263×hsa-miR-455-5p exp.) + (-0.4418×hsa-miR-320b exp.) + (0.2768×hsa-miR-944 exp.). The median risk score divided the 75 LSCC patients in the training cohort equally into high- and low-risk groups (Fig. 6c). Kaplan‒Meier survival curves revealed that the low-risk group had a significantly higher survival probability than the high-risk group (P < 0.001, Fig. 6d). Furthermore, a ROC curve was used to check the accuracy of the model, and the AUC was calculated as 0.674 at 1 year, 0.792 at 3 years, and 0.813 at 5 years (Fig. 6e), indicating the reliability and feasibility of the prognostic model. Patients in the high-risk group had a greater probability of death and a shorter lifespan than those in the low-risk group (Fig. 6f). PCA and t-SNE analysis confirmed that patients with different risks were sorted well into the different categories (Fig. 6g, h).

Fig. 6.

Construction and validation of the prognostic model based on 5 miRNAs. a Cross validation of the 5 DEMs for tuning the parameter selection in the Lasso regression. b Lasso coefficients of the 5 DEMs in LASSO Cox regression. c Distribution of patients based on the risk score. d Kaplan–Meier survival curves. e ROC curve analysis. f The survival status of patients in the high (on the right side of the dotted line) and low (on the left side of the dotted line) risk groups. g PCA plot based on the risk score. h t-SNE plot based on the risk score. i Distribution of patients based on the risk score in the test cohort. j Kaplan–Meier survival curves in the test cohort. k ROC curve analysis in the test cohort. l The survival status of patients in the high- and low-risk groups in the test cohort. m PCA plot in the test cohort. n t-SNE plot in the test cohort. o Univariate and p multivariate analyses for the training cohort. q Univariate and r multivariate analyses for the test cohort

The robustness of the prognostic model was determined by a test cohort containing 75 LSCC samples, which was divided into high- and low-risk groups according to the median risk score calculated by an established formula constructed from the training cohort, with 47 patients in the high-risk group and 28 patients in the low-risk group (Fig. 6i). The Kaplan‒Meier survival curve showed the same trends (P = 0.043. Figure 6j), suggesting that the model had the same prognostic risk assessment ability for patients with LSCC. Furthermore, the AUC was calculated as 0.591 for 1 year, 0.667 for 3 years, and 0.603 for 5 years (Fig. 6k). As the risk score increased, the survival time gradually decreased (Fig. 6l), and PCA and T-SNE analyses also obtained the same trend as the training cohort (Fig. 6m, n).

Furthermore, univariate and multivariate Cox regression analyses were also conducted to identify the independent factors associated with survival status. The univariate Cox regression analysis showed that the risk score was an independent prognostic factor for the survival of patients in the training cohort (P = 0.002, HR = 4.179, 95% CI: 1.666 − 10.478, Fig. 6o) and test cohort (P = 0.042, HR = 1.768, 95% CI: 1.020 − 3.065, Fig. 6p). The multivariate Cox regression analysis also indicated that the risk score was a prognostic factor independent of other clinical factors for LSCC patients in the training cohort (P = 0.003, HR = 4.644, 95% CI: 1.692–12.748, Fig. 6q) and test cohort (P = 0.014, HR = 1.896, 95% CI: 1.140–3.152, Fig. 6r). Moreover, univariate and multivariate Cox regression analyses revealed that sex and N stage were independent prognostic factors of poor survival in patients with LSCC in the test cohort. (Fig. 6q, r).

Validation of the expression of 5 DEMs in an independent clinical cohort by qRT‒PCR

An independent clinical cohort consisting of 48 LSCC and 36 NC samples was used to measure the expression status of the screened DEMs by qRT‒PCR. As expected, hsa-miR-139-3p and hsa-miR-486-5p were downregulated in LSCC patients. In contrast, hsa-miR-320b, hsa-miR-455-5p and hsa-miR-944 were upregulated in LSCC patients (Fig. 7a-e). To determine if plasma miRNAs from an independent clinical cohort could show the same effect, we evaluated the diagnostic efficacy of the above selected 5 DEMs. Consistent with previous results, individual miRNAs did not provide accurate diagnostic efficiency (Fig. 4d-h), and the ANN panel increased the AUC to 0.875 (Fig. 7v), further demonstrating the stability of the ANN panel. After that, we tried to construct a logistic model to further test the diagnostic efficacy of the 5 DEMs (Fig. 7f-u), and the AUC reached a maximum of 0.906 when two miRNAs were combined (Fig. 7l, hsa-miR-139-3p + hsa-miR-486-5p). The AUC was further improved when increasing the number of miRNAs from the logistic model, and it ranged from 0.923 to 0.949 for the logistic model containing four miRNAs (Fig. 7p-t). Then, the AUC reached 0.959 when the number of miRNAs increased to five (Fig. 7u). In summary, our diagnostic panel built with ANN was validated with a logistic model and still achieved good diagnostic efficiency. All 5 miRNAs provided promising AUC values for distinguishing LSCC patients from controls.

Fig. 7.

Validation of the expression of the 5 DEMs in an independent clinical cohort. The miRNA levels of hsa-miR-139-3p (a), hsa-miR-486-5p (b), hsa-miR-455-5p (c), hsa-miR-944 (d), and hsa-miR-320b (e) were measured in plasma samples from LSCC and NC patients. Validation of logistic model (f, hsa-miR-139-3p + hsa-miR-455-5p), (g, hsa-miR-486-5p + hsa-miR-944), (h, hsa-miR-486-5p + hsa-miR-320b), (i, hsa-miR-455-5p + hsa-miR-944), (j, hsa-miR-455-5p + hsa-miR-320b), (k, hsa-miR-139-3p + hsa-miR-320b), (l, hsa-miR-139-3p + hsa-miR-486-5p), (m, hsa-miR-455-5p + hsa-miR-486-5p), (n, hsa-miR-139-3p + hsa-miR-944), (o, hsa-miR-486-5p + hsa-miR-320b), (p, hsa-miR-455-5p + hsa-miR-320b + hsa-miR-486-5p + hsa-miR-944), (q, hsa-miR-139-3p + hsa-miR-455-5p + hsa-miR-486-5p + hsa-miR-944), (r, hsa-miR-139-3p + hsa-miR-455-5p + hsa-miR-320b + hsa-miR-944), (s, hsa-miR-139-3p + hsa-miR-455-5p + hsa-miR-320b + hsa-miR-486-5p), (t, hsa-miR-139-3p + hsa-miR-320b + hsa-miR-486-5p + hsa-miR-944), (u, hsa-miR-139-3p + hsa-miR-320b + hsa-miR-486-5p + has-miR-455-5p + hsa-miR-944) as biomarkers of early LSCC. The ROC curves of the sEV group (miRNA sequencing of 9 samples) are shown in red, while the ROC curves of the validation cohort (miRNA qRT‒PCR of 84 samples) are shown in blue. (v) Validation of the ANN diagnostic panel as a biomarker of early LSCC in the validation cohort

Discussion

LSCC accounts for a large portion of the worldwide cancer burden and is the second most common type of HNSCC [23]. Despite great strides on LSCC treatment in recent years, patients derive limited benefit from common therapeutic approaches, such as surgical resection, chemotherapy, and radiotherapy. At present, physical examination of the throat and biopsy remain the routine diagnostic procedures for the diagnosis of LSCC patients [24]. Nevertheless, this detection method is not sufficiently convenient and specific owing to the similarity with throat inflammatory disease and complexity of the operation. Therefore, seeking reliable diagnostic and prognostic markers for LSCC patients remains an effective way to improve the current status of diagnosis and treatment.

Since the discovery of EVs and their high correlation with cancers, EVs have gained growing interest in cancer research. EVs have high bioavailability, biological stability, targeting specificity, low toxicity and immune characteristics, indicating their potential in cancer diagnosis and treatment [25]. MiRNAs are a class of evolutionary endogenous small noncoding RNAs that bind to target mRNAs and negatively regulate gene expression through posttranscriptional inhibition or target mRNA degradation [25, 26]. MiRNAs are also reported to be involved in the occurrence and development of a variety of cancers [27]. Due to the specific sorting mechanism, the concentration of miRNA in EVs is higher and more stable, which has higher research application value. However, there is still a lack of understanding regarding the diagnosis, prognosis and related mechanisms of EV miRNAs in LSCC.

In this study, we performed miRNA expression screening analysis using plasma sEV specimens from 6 LSCC patients and 3 patients with vocal cord polyps. After intersection with the miRNA expression data from the TCGA database, 5 miRNAs were successfully screened to construct a prognostic panel for early diagnosis using RF and ANN, which was well validated in the independent datasets from the GEO database. The ANN diagnostic panel obtained an AUC of 0.782 in the training group and AUCs of 1.000, 0.716, and 0.875 in the three independent validation groups, which demonstrated the stability of the panel. The above 5 DEMs also achieved high AUC values in the logistic model, with AUCs of 1.000 and 0.959 when 5 DEMs were used to construct the logistic model, demonstrating the important potential of the screened DEMs in the early diagnosis of LSCC. Subsequently, we further verified the prognostic prediction ability of the above 5 miRNAs and successfully constructed a DEM-based independent prognostic model through LASSO Cox regression analyses, which was well validated in the test cohort. Finally, qRT‒PCR validated the expression of DEMs in an independent clinical cohort consisting of 48 LSCC and 36 NC samples, further demonstrating the accuracy and stability of the diagnostic panel.

Hsa-miR-139-3p has been confirmed to be downregulated in LSCC [28, 29], which inhibits LSCC development by targeting rab5a [30], the KDM5B/SOX2 axis and the Wnt/β-catenin pathway [31]. Based on the above data and our findings, we believe that hsa-miR-139-3p can be used as a reliable marker for the diagnosis and prognosis of LSCC. For hsa-miR-486-5p, it has been found that hsa-miR-486-5p in salivary EVs can discriminate HNSCC from healthy samples with 85% accuracy [32], and it has been shown to be downregulated in LSCC [33, 34]. Additionally, it may play an important role in the pathogenesis and development of LSCC. Hsa-miR-944 is a newly discovered cancer-related miRNA that is involved in regulating the proliferation, invasion and metastasis of oral squamous cell carcinoma (OSCC), hypopharyngeal cancer, and nasopharyngeal cancer [35–38]. However, the role of hsa-miR-944 in LSCC is still unknown. Our results suggest that hsa-miR-944 may be a promising biomarker that is worthy of further study. Similarly, there is a lack of relevant research on the role of hsa-miR-320b in LSCC. However, recent research has found that EVs and intracellular hsa-miR-320b promote lymphatic metastasis in esophageal squamous cell cancer (ESCC) and are closely associated with prognosis [39, 40]. Although there is no direct evidence that hsa-miR-455-5p is associated with plasma EVs of LSCC, studies have confirmed that it is highly expressed in laryngeal cancer, the expression level is more significant than that of adjacent noncancerous counterparts and benign laryngeal tissues [41] and is associated with prognosis [42]. Furthermore, hsa-miR-455-5p isolated from plasma has been proven to be an early diagnostic indicator of OSCC [43]. It may regulate OSCC cell proliferation and promote OSCC tumorigenesis by affecting the PTPRS-related pathway and transforming growth factor-β (TGF-β) pathway. These pathways are also common in HNSCC [44]. Each of these miRNAs seems to have a certain impact on tumorigenesis, and combining these miRNAs into a panel can more effectively predict cancer outcomes.

In the miRNA‒mRNA regulatory network, we identified ACTR3, ACTR2, YWHAZ, EIF3A, SYNCRIP and hsa-miR-944 as hub genes related to LSCC. Actin-related protein 3 (ACTR3) and ACTR2 are important subunits of the ARP2/3 complex, and the proteins they encode are known to be the major constituents of the ARP2/3 complex. The ARP2/3 complex has been proven to be involved in the regulation of cell motility and the epithelial mesenchymal transition (EMT) process and is closely associated with cancer development in gastric cancer, squamous cell carcinoma, and colorectal cancer [45, 46]. Tyrosine 3 monooxygenase/tryptophan 5-monooxygenase activation protein zeta (YWHAZ), belonging to the 14-3-3 protein family, is a central hub protein involved in many signal transduction pathways, plays a key role in tumor progression, and participates in a wide range of cell activities, including cell growth, cell cycle, apoptosis, and migration/invasion, in multiple types of cancers [47]. EIF3A is recognized as a proto-oncogene and has been widely reported in recent years, which is correlated with cancer occurrence, metastasis, prognosis, and therapeutic response [48]. EIF3a appears to be essential for LSCC to maintain a malignant phenotype, which may be potentially associated with disease development and progression of LSCC [49]. Synaptotagmin-binding Cytoplasmic RNA-Interacting Protein (SYNCRIP) is related to mRNA splicing, mRNA translation and pri-miRNA processing. It affects miRNA localization by mediating the distribution between intracellular compartments and EVs and has become a poor prognostic marker for a variety of diseases and cancers [50]. We also used GO and KEGG analyses to predict the biological functions and enrichment pathways in which the target genes may be involved, with GO analysis showing that the target genes were primarily involved in protein metabolism and cellular communication. KEGG analysis showed that the target genes were associated with immune responses and cellular metabolism. These aberrant GO classes and KEGG signaling pathways may be relevant to the pathogenesis and cancer biology of LSCC.

There were limitations in this study. First, sEV miRNAs are secreted not only by the tumor itself but also by the tumor microenvironment. Due to the lack of plasma sEV-related data from LSCC, we chose the TCGA and GEO databases for further analysis, which have data on the miRNAs derived from the tumor itself, which may have an impact on the number of DEMs. Second, immune-related research has not been well validated in this model, which will certainly inspire future research studies. In addition, the sEV miRNA data were obtained from a small cohort, and although they were subsequently verified in 207 samples from online databases and 84 clinical samples, this study is still preliminary. It is still necessary to further verify its accuracy in a large enough clinical cohort, as well as carry out follow-up mechanistic studies.

Conclusion

In summary, we have created a novel diagnostic panel based on 5 plasma sEV miRNAs for the early diagnosis of LSCC, which can also be used for prognosis prediction. Plasma sEV miRNAs may be used as convenient, noninvasive, and effective clinical markers to guide the clinical management of LSCC.

Supplementary information

Below is the link to the electronic supplementary material.

Author contributions

Chunping Wu and Mei Wang contributed to the study conception and design. Material preparation, data collection and analysis were performed by Haopeng Zhang, Zhengxun Chen, Qiang Huang, and Yang Guo. The first draft of the manuscript was written by Haopeng Zhang and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.

Funding

This study was supported by grants from the Special Fund for Clinical Research in Health Industry of Shanghai Municipal Health Commission (No. 202240320) and the clinical three-year action plan of Shanghai Shenkang Hospital Development Center (No. SHDC2020CR6011). The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Data availability

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

The authors have no relevant financial or non-financial interests to disclose.

Ethics approval

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committee of the Eye, Ear, Nose and Throat Hospital, Fudan University, Shanghai, China.

Consent to publish

 The authors declare that they agree to submit the article for publication.

Competing interests

The authors declare that they have no conflicts of interest.