Candidate microRNAs as biomarkers of thyroid carcinoma: a systematic review, meta‐analysis, and experimental validation

Abstract Thyroid cancer is one of the most common carcinomas of the endocrine system with an increasing incidence. A growing number of studies have focused on the diagnostic and prognostic values of dysregulated microRNAs (miRNAs) in thyroid carcinoma. However, differences in the measurement platforms, variations in lab protocols, and small sample sizes can make gene profiling data incomparable. A meta‐review of the published studies that compared miRNA expression data of thyroid carcinoma and paired normal tissues was performed to identify potential miRNA biomarkers of thyroid carcinoma with the vote‐counting strategy. Two hundred and thirty‐six aberrantly expressed miRNAs were reported in 19 microRNA expression profiling studies. Among them, 138 miRNAs were reported in at least two studies. We also provided a meta‐signature of differentially expressed miRNAs between individual histological types of thyroid carcinoma and normal tissues. The experimental validation with qRT‐PCR analysis verified that the profiles identified with the meta‐review approach could effectively discriminate papillary thyroid carcinoma tissues from paired noncancer tissues. The meta‐review of miRNA expression profiling studies of thyroid carcinoma would provide information on candidate miRNAs that could potentially be used as biomarkers in thyroid carcinoma.


Introduction
Thyroid carcinoma represents the most frequent carcinoma of the endocrine system [1]. Most thyroid cancers originate from thyroid follicular cells (>90%) and can be subdivided into well-differentiated papillary thyroid carcinoma (PTC) and follicular thyroid carcinoma (FTC), while only less than 5% originate from C-cell, often referred to as medullary thyroid carcinoma (MTC) [2]. The most common follicular tumor is benign hyperplastic adenoma, whereas PTC represents the most frequent thyroid carcinoma (about 90%). PTC and FTC may progress to poorly differentiated carcinoma or can fully lose differentiation to give rise to anaplastic thyroid carcinoma (ATC) [3].
A large number of studies have been performed to screen candidate biomarkers for thyroid carcinoma. Quite a lot of molecular variations have been identified in thyroid carcinoma tissues [4][5][6]. miRNAs are a class of noncoding RNAs, which are between 19 to 25 nucleotides in length. They have been demonstrated to be potential early cancer detection biomarkers, prognostic indicators, and therapeutic targets [7,8]. miRNAs exert function via binding to the complementary sites in the 3′ untranslated region of target mRNAs to promote target gene mRNA degradation or inhibit translation [9]. Studies have showed that miRNAs are involved in a wide array of cellular processes, including proliferation, apoptosis, metastasis, and cellular differentiation [10][11][12].
High-throughput technologies have been employed to screen the expression of miRNAs across normal and cancer tissues. These studies could result in hundreds or thousands of aberrantly expressed miRNAs, while only a small

ORIGINAL RESEARCH
Candidate microRNAs as biomarkers of thyroid carcinoma: a systematic review, meta-analysis, and experimental validation portion of them may be of actual clinical utility. Furthermore, with respect to the identified meta-signature of miRNAs, great inconsistency existed among different studies. Finding a meaningful combination from different datasets is usually not an easy job. Differences in measurement platforms, variations in experiment protocols, limited numbers of samples studied, and low numbers of aberrantly expressed miRNAs in comparison to relatively large total numbers of miRNAs, may render miRNA expressions levels uninterpretable. Therefore, it might be better to analyze datasets separately and thereafter aggregate the miRNA list. Such a strategy has been a success in finding human gene coexpression networks [13] and in defining more accurate list of cancer-related genes [14,15] and miRNAs [8,16,17].
We could use the meta-review approach, which combines the miRNAs expression profiling results to increase the statistical power for working out the inconsistency or discrepancies. In this study, a meta-review of published miRNAs expression profiles across normal and thyroid cancer tissues was performed. Then we used the well-known meta-analysis method, the vote-counting strategy [14,15], and ranked the miRNAs based on the number of profiling studies consistently reporting this miRNA, total sample size and average fold change. The metaanalysis was first carried out in all histological types of thyroid carcinoma (PTC, follicular thyroid carcinoma (FTC), medullary thyroid carcinoma (MTC), and ATC). Then, a meta-analysis was performed in four subtypes of thyroid carcinoma, respectively.

Selection of studies and datasets
A search for thyroid carcinoma miRNA expression profiling studies was performed in PubMed using the following keywords: "miRNA" OR "microRNA" OR "miR", "thyroid carcinoma", "profiling" OR "microarray". The latest search was performed on 25 February 2016. Titles and abstracts of the obtained articles were screened, and full texts of the articles of interest were further evaluated. Original articles published in English that analyzed miRNA expression between thyroid carcinoma and noncancerous thyroid tissue in humans were included. Exclusion criteria: (1) articles published in non-English language; (2) case reports or review articles; (3) studies with the method of qRT-PCR for initial screening; (4) studies using serum or plasma of thyroid cancer patients; (5) studies not using the method of miRNA microarray or sequencing platform for initial screening; (6) profiling of histological subtypes other than the predetermined histological subtypes (PTC, FTC, MC and ATC); (7) studies not including noncancerous normal tissues; (8) detailed information of platforms were not available; (9) profiling of benign thyroid tumor samples; (10) profiling across metastatic and nonmetastatic, recurrent and nonrecurrent, aggressive and nonaggressive thyroid carcinoma tissues; and (11) profiling studies not across malignant thyroid carcinomas and normal thyroid tissues.

Data extraction
The two authors (YH and YW) performed the online search, evaluation and extraction of data utilizing the standard protocol independently, with the discrepancies resolved by discussion with the third author (EC). The information listed below were retrieved from the full texts and supplemental materials: author, time of publication, country of subjects, year of sample analysis, clinical characteristics of the enrolled thyroid carcinoma patients, characteristics of measurement platforms, list of dysregulated miRNA features, cut-off criteria of statistically differentially expressed miRNAs, and fold changes. miRNA annotation were standardized to miRBase Release 21.

Ranking
MiRNAs were ranked according to the order of importance below: (1) number of studies reporting the same miRNAs with a consistent direction of aberration; (2) total number of profiling samples in the same direction of change; and (3) average fold changes for the same miRNAs reported consistently. We consider total sample size to be more important than average fold change as fold changes were not available in many studies. Average fold change was calculated with the method of weighted mean, mean = (x 1 f 1 + x 2 f 2 + … x k f k )/(f 1 + …f k ), x k stands for fold change of a single study, f k stands for sample size. In studies where fold changes were not reported, the 2 − ΔΔ Ct method was used to determine fold change between two groups. The relative expression of miRNA was calculated with reference to expression of house-keeping genes and expressed as fold changes.

Sample collection
Twenty-five PTC samples and paired noncancer thyroid tissue samples were collected between October 2014 and May 2016 after radical surgical section at the Department of Thyroid and Breast Mininally Invasive Surgery, Ningbo Yinzhou People's Hospital (Ningbo, China). The diagnoses were finally made by skilled pathologists. Once the surgical specimens were removed, research personnel instantly transferred the PTC tissues to the lab. Pathology faculty evaluated the specimen grossly and selected the thyroid tissues that most was likely to be cancerous. Matched noncancer thyroid tissues were isolated at least 2 cm away from the tumor border and were shown to be free of tumor cells by microscopy. Each tissue samples were frozen in liquid nitrogen immediately and stored at −80°C in a refrigerator for RNA isolation.

Statistical analysis
The statistical analysis were performed utilizing SAS 9.2 software (SAS Institute Inc. NC, USA). Data are presented as means ± standard deviation. Student's t-test was utilized for comparison between two independent groups. A P < 0.05 (two-sided) was considered to be statistically significant.
The number of thyroid cancer patients measured in the 19 reports ranged from 2 to 30. These studies used various kinds of microarray platforms, and the number of miRNAs assayed ranged from 158 to 1205 (mean 778; data were missing in four studies [22,24,34,36]). Among them, three studies [18,19,23] presented the whole list of aberrantly expressed miRNAs in the supplemental materials, whereas the other studies provided a part of the profiling data. Thus, we directly contacted the corresponding authors and obtained the whole data lists from corresponding authors of seven studies [20,21,25,[28][29][30]32]. The aggregated dataset included a total of 241 tumor samples and 170 noncancerous tissue samples.
In total, 19 studies reported 486 aberrantly expressed miRNAs across thyroid carcinomas and paired normal tissues. Among them, 273 were reported to be upregulated and 213 downregulated; 138 were reported in more than one study; 90 (65.22%) miRNAs were consistently reported (Tables 2 and 3) and 48 (34.78%) were reported with an inconsistent direction (Table 4). Among the consistently reported 90 miRNAs, 37 were upregulated ( Table 2) and 53 were downregulated (Table 3). In the group of consistently reported microRNAs, miR-221-5p and hsa-miR-222-5p was reported to be upregulated in 16 studies followed by miR-146b-5p upregulated in eleven studies. miR-138-5p and miR-486-5p were found to be downregulated in eight studies. We also provided a metasignature of differentially expressed miRNAs between individual histological type of thyroid carcinoma tissues and normal tissues (Tables 5-14). PTC: Tables 5-7, FTC:  Tables 8-10, MTC: Table 11, ATC: Tables 12-14. According to the results from our meta-analysis, the top lists varied among the studies.

Discussion
The lack of agreement among studies is a common drawback of miRNA profiling studies. Variations in experiment protocols, differences in measurement platforms, limited numbers of samples studied, and low numbers of aberrantly expressed miRNAs in comparison to relatively large total numbers of miRNAs, may render miRNA expressions levels uninterpretable. It was demonstrated that each platform is comparatively stable with respect to its own intrareproducibility. Yet, the interplatform reproducibility is relatively low among different platforms [41,42]. Furthermore, the small sample size and large numbers of features have resulted in high numbers of false negative results due to low statistical power [43].
Although the ideal method of miRNA analysis is working on the aggregated raw profiling datasets; however, it is usually unrealistic to perform this rigorous approach as the raw data are often unavailable and the interplatform result concordance is low. To overcome these obstacles, it may be a preferred solution to analyze datasets separately and thereafter aggregate the resulting miRNA list. The meta-analysis approach was used to analyze thyroid cancer specific miRNAs obtained from independent reports. The key element of this method was searching for the most recognized miRNAs in the profiling studies. Microarray remains the most used assay for high-throughput screening [44,45]. Due to the fact that qRT-PCR can only detect the preselected miRNAs and the interplatform result concordance between microarray and qRT-PCR remains low [45], we concentrated on reports that screened miRNA expression with microarray platforms.
We need to consider some factors when identifying candidate diagnostic miRNAs in thyroid cancer. In the first place, the average fold change of the candidate miRNA should be big enough to discriminate cancer samples from benign tissues. As demonstrated in Tables 2 and 3, the mean fold changes of the identified, consistently reported miRNAs from microarray platform-based studies were all more than 2. Furthermore, we carried out a meta-analysis in four histological subtypes of thyroid carcinoma, respectively. We observed that the meta-signature of different subtypes of thyroid carcinoma varied considerably.
In the second place, further research on the biological functions of miRNAs are required. One miRNA may have dozens or hundreds of target genes, and one mRNA may be modulated by multiple miRNAs [7]. For example, miR-221 regulated gastric carcinoma cell proliferation by targeting phosphatase and tensin homolog deleted on chromosome ten (PTEN) [46] and could enhance growth and invasion of gastric cancer cells by targeting RECK [47]. Though the interaction between miRNA and mRNA could be tumor-specific, a deeper understanding of the molecular mechanism could contribute to advancements in clinical applications.
Thirdly, there should be adequate information about their pattern of expression in various kinds of tissues. It has been suggested that serum-obtained miRNAs are more tissue-specific than tumor-specific [48,49]. In view of the fact that there are only three studies [50][51][52] on plasma-based miRNAs, we included only studies that analyzed miRNA expression across thyroid cancer and normal tissues.
External experimental validation in an independent cohort of patients is often required to confirm the metaanalysis results. We determined the expression of the eight identified miRNAs with qRT-PCR analysis and verified that the eight miRNAs were indeed differentially expressed between PTC samples and normal thyroid tissues.
The results of the systematic review might add some information to the candidate miRNA biomarkers in thyroid carcinoma. The identified microRNAs, which are most consistently reported, may be potential diagnostic/prognostic biomarkers and therapeutic targets.