MicroRNAs-208b-3p, 204-5p, 129-2-3p and 3065-5p as predictive markers of oral leukoplakia that progress to cancer

Abstract

Leukoplakia is the most common precursor lesion of oral squamous cell carcinoma (OSCC). Currently, the risk of progression to OSCC is assessed based on histopathologic examination alone. However, this method fails to identify the subset of microscopically innocuous leukoplakia that ultimately transforms to OSCC. The aim of this study was to determine if microRNAs (miRNAs) can be utilized to identify non- and low-grade dysplastic oral lesions at risk for cancer progression. A retrospective study of genome-wide miRNA expression level analyses was performed in the training cohort (n=20) using deep sequencing formalin-fixed paraffin embedded incisional biopsy tissues from patients with oral leukoplakic lesions diagnosed with non- or low-grade dysplasia and known clinical outcome. The promising miRNA candidates were then evaluated in the validation cohort (n=80) using quantitative real-time PCR (qRT-PCR). Four promising miRNAs-208b-3p, 204-5p, 129-2-3p and 3065-5p were identified. Combining these four miRNAs as a panel with age and histologic diagnosis (p<0.004), our final model had a predictive value for the area under the receiver operating characteristic (ROC) curve (AUC) of 0.792, sensitivity of 76.9% and specificity of 73.7% to accurately identify non- and low-grade dysplastic lesions at risk of cancer progression, which is a significant improvement over histopathologic examination alone (AUC of 0.645). While further investigation is needed, discovery of predictive markers that can accurately identify histologically innocuous oral lesions at high risk for progression to OSSC will significantly improve clinical outcome by means of early intervention.

Introduction

Oral squamous cell carcinoma (OSCC) is the 6^th leading cause of cancer related deaths worldwide [1,2]. Over 30,000 people in the United States are diagnosed with OSCC each year [1-4]. Half of those newly diagnosed with oral cancer will die of the disease [1-4]. Since the majority of OSCC develop from a precursor lesion [4,5], accurate identification, followed by complete surgical removal of the precursor lesion will effectively halt its malignant progression, offering the best hope at reducing OSCC associated morbidity and mortality.

Most precursor lesions of OSCC clinically present as a white patch known as leukoplakia [4-16]. Upon microscopic examination, most leukoplakia (80%) appear to be non-dysplastic or low-grade dysplastic lesions, whereupon the patients are diagnosed with epithelial hyperplasia and/or hyperkeratosis and placed under ‘close observation’ without further treatment [14]. Moderate to high-grade epithelial dysplasia is seen in ~17% of leukoplakia [15], in which case the lesion is completely removed by surgery [16]. The remaining 3% of cases are OSCC [14]. While moderate to high-grade epithelial dysplasia will most likely progress to OSCC within 4-5 years, some of the non-dysplastic and low-grade dysplastic lesions also undergo malignant transformation [8-16]. Hence histopathologic examination alone is insufficient for identification of leukoplakic lesions that will progress of OSCC, especially if the leukoplakia lacks atypical cellular features at the time of initial histopathologic assessment.

It is these histologically innocuous lesions that are at greatest need for intervention since without the presence of significant histologic dysplasia, these lesions could be erroneously interpreted as being harmless and not managed properly. Currently there are no established guidelines for managing non- and low-grade dysplastic lesions. Treatment options vary from no treatment to complete surgical excision to laser ablation [17]. Moreover, there is no established guideline for patient follow-up duration or frequency [17]. Accurate identification of non-dysplastic and low-grade dysplastic lesions at risk for malignant transformation will ensure proper clinical management of these premalignant lesions.

We assessed microRNA (miRNA)-based markers that can increase the predictive value when combined with histopathologic examination to accurately identify non- or low-grade dysplastic lesion that will ultimately progress to OSCC. miRNAs are small non-coding RNA molecules that function as post-transcriptional regulators capable of causing both gene silencing and activation [18]. They are involved in multiple critical biological processes, including cellular proliferation, apoptosis, differentiation, and carcinogenesis [19-22]. A limited number of studies have examined differential miRNA expression levels in precancerous lesions in the esophagus, cervix, lung and stomach [23-28]. While there are reports of miRNA expression profiles in oral leukoplakia that progressed to OSCC [29-36], only a couple of these studies focused on low-grade dysplastic lesions that progressed to cancer [34-36]. Moreover, while the majority of other studies of oral leukoplakia utilized TaqMan low density arrays, we employed genome wide deep sequencing to determine a miRNA expression profile unique to histologically innocuous (non- and low-grade dysplastic lesions) leukoplakias that ultimately progress to cancer. The advantage of this next-generating sequencing technology is that it is capable of detecting all miRNAs present in the samples, thereby increasing testing coverage [37]. Development of a clinically applicable method to identify histologically innocuous precancerous lesions will be a valuable modality for clinicians in guiding patient management.

Materials and methods

Subjects

Following institutional review board approval at the Columbia University Medical Center, we conducted a retrospective search of our pathology database to identify 100 adult patients ≥21 years old, with a clinical leukoplakia diagnosed as ‘epithelial hyperplasia’, ‘epithelial hyperplasia with hyperkeratosis’, ‘epithelial atypia limited to the basal cell region’ or ‘mild epithelial dysplasia’ prior to 2008 and which also had a minimum of 5-year follow-up information available. Only the patients who had an incisional biopsy of the leukoplakia were selected. As an excisional biopsy could account for the reason why the leukoplakia did not progress to cancer, the patients who had excisional biopsy were excluded from the study. Identified patients were stratified into the following two groups; Group 1 ‘Progressive Group’ (patients with leukoplakia that progressed to OSCC within 5 years) and Group 2 ‘Non-Progressive Group’ (patients with leukoplakia that did not progress to OSCC within 5 years). The age and gender of the patient, histologic diagnosis, and the location of lesion were recorded. Archived formalin-fixed paraffin-embedded (FFPE) tissue blocks were retrieved for all subjects.

Total RNA extraction

Ten 10-μm sections were obtained from archived FFPE precursor tissue samples for all subjects. For each sample, a representative section was stained with H&E and reviewed by a pathologist to confirm the histologic diagnosis. Total RNA was isolated from tissues using RNeasy FFPE kit (Qiagen Inc., Valencia, CA) following the manufacturer’s protocol and yield was quantitated by Nanodrop, and samples were stored at -80°C.

microRNA quality control, library preparation and sequencing

Deep sequencing analysis was performed as previously described [38] in the training set of 20 patients; 10 from Group 1 (progressive group) and 10 from Group 2 (non-progressive group). For quantification and quality control, total RNAs were tested using the Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA) and the Qubit 2.0 Fluorometer (Life Technologies, Grand Island, NY). Total RNA (minimum of 2 μg) ranging from 18 to 30 nt were gel-purified and ligated to 3’ and 5’adaptors. Ligation products were reverse transcribed, then amplified for 16 cycles using the adaptor primers, and the fragments around 150 bp were isolated from PAGE-gels using a TruSeq Small RNA Sample Prep Kit (Illumina, San Diego, CA). Libraries were sequenced on an Illumina HiSeq 2500 platform-50 SR that allows for 1x50 base-pair single-end reads in the New York Genome Center. The data was deposited in the publicly available Gene Expression Omnibus (GEO) database (GSE62809).

microRNA mapping and differential expression analysis

Adaptors were removed and low quality tags were filtered with FASTX-Toolkit (http://hannonlab.cshl.edu/fastx_toolkit/). Reads were processed with the pipeline miraligner, which mapped them to the miRBase v.20 sequences. The number of reads per miRNA was first assessed and analyzed using the bioconductor package DESeq to compare miRNA expression between the two groups. In addition to normalizing between the samples, DESeq performed a statistical test of differential expression under the hypothesis of a negative binomial distribution of the reads. To select the top miRNAs with prognostic value, those with significant (unadjusted p<0.05) and >1.25 log2-fold change of the different mean normalized expression levels between the two groups were identified. The association between selected miRNAs and their role in carcinogenesis was also investigated through a literature search.

qRT-PCR for microRNAs

Selected top-ranked miRNAs were quantified using TaqMan MicroRNA Assays Kits (Applied Biosystems, Foster City, CA) as described previously [38], in the validation set with 80 patients, consisting of 40 patients in Group 1 (progressive cases) and 40 patients in Group 2 (non-progressive cases). Input RNA was reverse transcribed using the TaqMan miR Reverse Transcription Kit and miR-specific stem-loop primers for the selected four miRs in a small-scale RT reaction. For quantification, diluted RT product was combined with PCR assay reagents and real-time PCR carried out on an ABI7900HT thermocycler. The endogenous control (RNU48) showed even expression levels between the two groups. Test samples were assayed in duplicate with the laboratory blinded to survival status and with 5% triplication after relabeling. The coefficient of variation was calculated and values <5% was considered acceptable. Data was analyzed with SDS Relative Quantification Software version 2.2.2 (Applied BioSystems) to determine the threshold cycle (Ct). The fold change was determined by the 2^-ΔΔCt method [39]. A two-sample t test was used to compare the normalized expression levels between the two groups.

microRNA target prediction

MicroRNA target analysis was performed by the TargetScanHuman release 7.0 (www.targetscan.org) to compare potential target genes affected by the top four miRNAs with the NCBI reference (human genome build 36). The biologic processes, molecular functions, cellular components, and protein classes of these genes were examined utilizing the Panther web site (Supplemental Table 1).

Statistical analysis

Univariate logistic regression models were first used to test if the selected miRNAs were independent factors of 5-year progression status. A multiple logistic regression model was then used to build a prediction model. Histologic findings were incorporated into the model by assigning a numeric value of ‘1’ to diagnoses of hyperkeratosis, ‘2’ to epithelial hyperplasia with hyperkeratosis, ‘3’ to epithelial atypia and ‘4’ to mild epithelial dysplasia. Using the multiple logistic regression model, we constructed receiver-operating characteristic (ROC) curves and calculated the area under the curve (AUC) and the sensitivity and specificity with 0.5 predicted probability of progression to OSCC as the cutoff point. P<0.05 was considered statistically significant. Statistical analyses were conducted using SAS 9.3 (SAS Institute).

Results

Deep sequencing, mapping of miRNAs and selection of predictive miRNAs

Age, gender, histopathologic diagnosis and location of lesion are listed in Table 1 for the 20 subjects in the training set and 80 subjects in the validation. For deep sequencing, the total number of reads obtained ranged from 3,229,855 to 39,216,964 for the 20 tissue samples. After removing adaptors and filtering out low quality tags, 973,424 - 31,446,445 clean reads were obtained (~30.14-80.19%). From these reads, the mapping rate to miRNA ranged from 0.89 to 12.71, and the rate to miRalligner, which distinguish -3p and -5p sequences, ranged from 1.48 to 14.75. The length distribution analysis revealed a peak at 22 nt, which is the size of most known microRNAs. For each sample, ~30 thousand to 11 million sequence reads mapped to the human genome were obtained, which included miRNA, rRNA, Mt_rRNA, snoRNA, snRNA, tRNA. A total of 1755 miRNAs were detected in each sample using the miRaligner and the number of reads of each miRNA ranged from 0 to 138,335.

Table 1.

Demographic and clinicopathologic characteristics of training and validation sets

	Training Set (n=20)		Validation Set (n=77)
	Non-progressive Group (n=10)	Progressive Group (n=10)	Non-progressive Group (n=39)	Progressive Group (n=38)	P-value*
Age mean (SD)	58.9 (11.1)	63.2 (23.2)	59.62 (12.6)	66.5 (17.5)	0.050
Gender % male	40%	50%	33%	35%	0.813
Histologic Diagnosis					0.058
Hyperkeratosis	1 (10%)	2 (20%)	17 (43%)	11 (28%)
Epithelial hyperplasia and hyperkeratosis	3 (30%)	3 (30%)	9 (23%)	6 (15%)
Epithelial atypia of the basal cell layer	4 (40%)	2 (20%)	10 (25%)	9 (23%)
Mild epithelial dysplasia	2 (20%)	3 (30%)	4 (10%)	14 (35%)
Location					0.256
High risk site (tongue, floor of mouth)	7 (70%)	7 (70%)	21 (53%)	26 (65%)
Low risk site (buccal mucosa, vestibule, gingiva, palate, lip mucosa)	3 (30%)	3 (30%)	19 (47.5%)	14 (35%)

^*Two-sample t-test was used to compare means between non-progressive and progressive groups and Pearson’s Chi-square test was used to compare proportions between non-progressive and progressive groups within the validation set (n=80).

A total of 4 miRNAs (miRs-129-2-3p, 204-5p, 208b-3p and 3065-5p) were identified to be differentially expressed in the progressive group (Group 1) compared to non-progressive group (Group 2) with at least 1.25 log2 fold change with unadjusted p<0.05 and to be expressed in close to (at least 80%) if not all samples. The mean expression levels and the fold changes of the 4 selected miRNAs are listed in Table 2. There were no miRNAs with false discovery rate adjusted p<0.05. One of the miRNAs (miR-208b-3p) was overexpressed in the progressive group and 3 were underexpressed (miRs-129-2-3p, 204-5p and 3065-5p) in the progressive group as shown in Figure 1.

Table 2.

Mean expression levels of the 4 selected miRNAs and the fold change between the two groups by deep sequencing in the testing set with 20 patients and qRT-PCR in the validation set with 77 patients

miRNAs	Methods	Non-progressive Group mean (SD)	Progressive Group mean (SD)	Fold-change	P-value**
miR-129-2-3p	Deep Seq (n=20)	13.28 (9.67)	2.75 (2.78)	0.21	0.0001
	qRT-PCR (n=77)	6.81 (1.98)	6.21 (1.42)	1.36	0.361
miR-204-5p	Deep Seq (n=20)	848.03 (1117)	239.16 (537.9)	0.28	0.007
	qRT-PCR (n=77)	3.55 (1.83)	2.88 (1.70)	1.60	0.250
miR-208b-3p	Deep Seq (n=20)	51.86 (58.46)	132.73 (169.1)	2.56	0.003
	qRT-PCR (n=77)	6.24 (1.89)	5.43 (1.99)	1.73	0.049
miR-3065-5p	Deep Seq (n=20)	14.03 (9.36)	5.01 (3.47)	0.36	0.005
	qRT-PCR (n=77)	6.23 (1.78)	6.61 (1.79)	0.86	0.564

*Mean count for deep sequencing and mean normalized Ct value for qRT-PCR.

^** p-value is unadjusted for multiple comparisons and is based on two-sample t-test comparing the means of the normalized miRNA expression levels.

Figure 1.

Scatterplot demonstrating overexpression of miRNA-208-3p and underexpression of miRNAs-204-5p, 3065-5p and 129-2-3p in the Progressive Group (Group 1) compared to the Non-Progressive Group (Group 2) in the training set (n=20) as measured by deep sequencing.

Putative target genes for the selected microRNAs

The biologic characteristics of 8, 74, 59, and 96 conserved genes potentially targeted by miRNAs-129-2-3p, 204-5p, 208b-3p and 3065-5p, respectively, were evaluated using TargetScanHuman release 7.0 (www.targetscan.org) and NCBI database (human genome build 36). Enriched genes were associated with multiple biologic pathways including pathways involving regulation of transcription, VEGF activity, Activin, IL-6, Notch signaling pathways, p53 regulation, Wnt/beta-catenin signaling cascade, cell proliferation, cell differentiation, cellular adhesion, apoptosis, cell migration, tumor suppression, and oncogenic signaling (Supplemental Table 1).

Evaluation of the top four miRNA expression profiles by RT-qPCR in the validation set

To confirm overexpression of miR-208b-3p and underexpression of miRs-204-5p, 129-2-3p and 3065-5p, these four miRNAs were quantified using qRT-PCR in the validation set with an additional 40 patients that progressed to OSCC (Group 1) and 40 patients that did not progress to OSCC (Group 2).

Only two of the four selected microRNAs showed changes in expression level consistent with that of the deep sequencing; miR-208b-3p was overexpressed and miR-3065-5p was underexpressed in the progressive group (Table 2). This finding is consistent with the putative role of these miRNAs; miR-208b-3p with oncogenic function, and miR-3065-5p with tumor suppressor role (Table 3). Although statistically insignificant, the other two miRNAs, miR-129-2-3p and miR-204-5p, showed overexpression in the progressive group (significant underexpression of these two microRNAs were demonstrated in the deep sequencing analysis), despite their presumptive role as tumor suppressors.

Table 3.

Univariate analysis of microRNAs associated with cancer progression in the validation set (n=77)

miRNAs	OR (95% CI)	P-valuea	Chromosomal Location	Putative function	Expression in Progressive Group
miR-129-2-3p	1.785 (1.686-1.694)	0.361	11p11.2	Tumor suppressor	increased
miR-204-5p	2.184 (1.78-2.243)	0.250	9q21.12	Tumor suppressor	increased
miR-208b-3p	1.472 (1.236-2.061)	0.049	14q11.2	Oncogenic	increased
miR-3065-5p	1.605 (1.569-1.469)	0.564	17q25.3	Tumor suppressor	decreased

^a p<0.05 is considered to be significant.

Multiple logistic regression analysis was performed to construct a predictive model of cancer progression using the top 4 miRNAs together with age, gender, histologic diagnosis, and location of the lesion. After model selection procedures, the final predictive model included the four miRNA marker panel (miRs-208b-3p, 3065-5p, 129-2-3p and 204-5p), age and histologic diagnosis. In the final model the endogenous control was undetectable in three samples. Since normalization could not be performed for miRNA expression level in those samples, they were eliminated (n=77). The AUC of the ROC curve of the final model (p<0.004 for the 6 predictors combined) was 0.792, with a sensitivity of 76.9% and specificity of 73.7% (Figure 2). Interestingly, this predictive value was stronger than that histology alone (AUC of the ROC curve = 0.646 with sensitivity of 57.5% and specificity of 65.0%), which is utilized in current clinical practice as a sole source of oral cancer progression predictor. As an exploratory study, we performed Maximum Likelihood Estimates analysis for the 6 predictors included in the final model in an attempt to generate a meaningful risk score formula that can be utilized in a clinical setting. The following formula was constructed from the analysis:

Figure 2.

ROC curve for miRNA levels combined with age and histologic diagnosis that can identify progressive oral precursor lesions in the validation set (n=77).

Score of cancer progression = -1.837 - (0.147 x miR-129b-3p) - (0.320 x miR-204-5p) - (0.312 x miR-208b-3p) + (0.348 x miR-3065-5p) + (0.048 x age) + (0.160 x histologic diagnosis level2) + (0.259 x histologic diagnosis level3) + (1.803 x histologic diagnosis level4)

Risk of cancer progression = exp(Score)/{1+exp(Score)}

From this analysis, -1.837 is the intercept, and the higher risk of cancer progression was associated with over expression of miR-3065-5p and underexpression of miRs-129-2-3p, 204-5p and 208b-3p, as well as older age. For the histologic diagnosis, increase in assigned number [1=epithelial hyperplasia, 2=epithelial hyperplasia with hyperkeratosis, 3=cellular atypia, 4=epithelial dysplasia] was associated with increased risk of cancer progression. When the risk score of cancer progression was calculated for all subjects in the validation set (n=77), the mean risk score in the Progressive group was 0.773 (63% likelihood) versus those in the Non-Progressive group -0.680 (38% likelihood). Using the mean (-0.282) as a cutoff point, 31 of the 39 progressive cases (80%) were correctly identified as high risk for cancer progression. In the non-progressive group, 24 of the 38 cases (63%) were properly identified as minimal risk of cancer progression.

Discussion

The deep sequencing analyses in the training set of 20 subjects revealed the overexpression of miR-208b-3p and underexpression of miR-204-5p, 129-2-3p and 3065-5p in the non- and low-grade dysplastic lesions that ultimately progressed to cancer. Indeed, the biologic characteristics of conserved genes potentially targeted by these miRNAs are associated with multiple critical biological processes related to carcinogenesis and tumor suppression. miR-208-3p has a presumptive oncogenic role and its overexpression has been linked to increased cellular proliferation, cell cycle progression, and tumorigenicity in hepatic tissue [40] and esophageal squamous cells [41]. SOX6 tumor suppressor gene is directly targeted by miR-208, and hence, overexpression of miR-208 leads to downregulation of SOX6 protein, which results in downregulation of p21, upregulation of cyclin D1, and deregulation Rb via phosphorylation [41].

miR-3065-5p is a relatively novel microRNA with a limited number of reports. Expression of miR-3065-5p was much lower in metastatic prostate cancer cells compared to those that did not metastasize. Hence, miR-3065-5p is thought to have a tumor-suppressive effect, contributing to reduction of cell migration and tissue invasion [42]. miR-129-2-3p is a negative regulator of SOX4, an oncogene [4]. Hypermethylation of its promoter region and subsequent gene silencing of miR-129-2-3p leads to overexpression of SOX4 and Cdk6, an event reported in tumorigenesis of endometrium [43], stomach [44], and liver [45]. Epigenetic alteration, such as hypermethylation of the promoter region, has been reported to be an early event in oral carcinogenesis [46]. Similarly, miR-204-5p is thought to have a tumor suppressor function. Its downregulation via hypermethylation is associated with increased metastasis and decreased overall survival in breast cancer [47], endometrial carcinoma [48], and colorectal cancer [49]. Researchers have shown that miR-204 is located at the genomic imbalanced 9q21.1-22.3 locus associated with genetic predisposition for head and neck cancer [50].

In the validation cohort (n=77), miR-208-5p expression levels showed a significant increase in the progressive group compared to that of the non-progressive group, concordant with the deep sequencing data. There was underexpression of miR-3065-5p in the progressive group, as observed in the deep sequencing analysis, although it was not statically significant. In contrast to the deep sequencing data, overexpression of miR-129-2-3p and miR-204-5p were detected in the progressive group in the validation cohort. Similar discrepancies between the deep sequencing and qRT-PCR data have been reported by Schee et al [50]. In theory, lack of specificity of the qRT-PCR assay for miR-129-2-3p and miR-204-5p might explain such discrepancies [50]. Alternatively, the small sample size in our training set may have led to false positives.

Despite the discrepancies in expression profile of microRNAs between the training and validation cohorts, when the four miRNAs were combined as a panel together with age and histologic diagnosis, the AUC of the ROC curve was 0.792, with a sensitivity of 76.9% and specificity of 73.7% (p<0.004) to identify non- and low-grade dysplastic lesions that progress to cancer. As previously mentioned, histopathologic examination alone has limitations in identifying leukoplakic lesion that will progress of OSCC, especially if the leukoplakia lacks significant atypical cellular features at the time of initial histopathologic assessment. Limitations of histopathology were also observed in our study; the predictive value of the histopathology alone was 0.645 (AUC of the ROC curve). If the four miRNA marker panel is added to the histopathologic examination, which can be obtained via qRT-PCR with relative ease, together with patient’s age, the predictive power to identify non- and low-grade dysplastic lesions at higher risk for malignant transformation will increase from 0.645 to 0.792.

For clinical practicality, we conducted an exploratory study to assess the feasibility of developing a parsimonious risk score formula from our final model. The main purpose of risk score formula is to translate miRNA expression levels assessed by qRT-PCR at the time of initial biopsy into a score that reflects the patient’s risk of cancer progression. Using the risk score formula, 31 of the 39 progressive cases (80%) were accurately identified as high risk for cancer progression. Conversely, 24 of the 38 non-progressive cases (63%) were properly identified as minimal risk of cancer progression. Those cases identified as high risk would have been treated aggressively initially which would potentially have prevented carcinoma development. The 63% of the low cancer progression risk group could have potentially been spared from unnecessary surgical excision. Further investigation is being planned to determine if our model can be utilized as a predictive modality for progression in early oral leukoplakia. This study is limited by small sample size. Also there are limitations in terms of generalizability as this is a single institutional study. However, our study is unique in that it assesses genome-wide miRNA expression profiles to identify a panel of miRNAs that may be utilized as a cancer progression predictive modality in non- and low-grade dysplastic lesions. We have plans to perform an internal validation and test the repeatability of the final risk score model in a larger set of samples. We will then test the model in a multicenter setting prior to conducting a large-scale prospective study. If validated, we will have obtained a useful predictive modality that can be applied in the clinic, which will guide appropriate management of patients at risk of developing oral cancer.

Disclosure of conflict of interest

None.