A six-gene panel to label follicular adenoma, low- and high-risk follicular thyroid carcinoma

Abstract

The distinction between follicular thyroid carcinomas (FTCs) and follicular-patterned benign lesions is almost impossible on fine-needle aspiration cytology. Furthermore, minimally invasive FTCs (MI-FTCs) with less than 4 vascular invasion foci generally have an excellent prognosis, but there are exceptions and, so far, no molecular marker appears able to identify them reliably. We aimed to distinguish benign lesions from low- and high-risk FTCs by a small-scale combination of genes. The expression analysis of 75 selected genes was performed on 18 follicular adenomas (FAs), 14 MI-FTCs and 6 widely invasive FTC (WI-FTCs). The mutational status of the RAS genes, TERT promoter and PAX8-PPARG rearrangements was also investigated. Seven samples were mutated, namely 3 MI-FTCs and 4 WI-FTCs. Twenty-five genes were differentially expressed (FDR <0.05) between FAs and WI-FTCs. Six of these (ECM1, RXRG, SDPR, SLC26A4, TIFF3, TIMP1) were also differently expressed among MI-FTCs and FAs or WI-FTCs and were considered to build a classification model, which was tested to classify samples according to their histological class. Hence, 31 out of 38 were correctly classified, and accuracy remained high after cross-validation (27/38). The 2 MI-FTCs incorrectly classified as WI-FTCs harbored both RAS and TERT promoter mutations. The capability of these six genes to stratify benign, low- and high-risk lesions appears to be promising in supporting the diagnosis of indeterminate thyroid nodules.

Introduction

The incidence of follicular thyroid carcinoma (FTC) has decreased over the last years (1), but still accounts for 10–15% of all thyroid cancers (2). The distinction between follicular adenoma (FA) and FTC, conventionally classified as minimally invasive (MI-FTC) and widely invasive (WI-FTC) (3), cannot be determined by fine-needle aspiration cytology (FNAC), since the demonstration of malignancy in these types of lesions is based on the presence of capsular and/or vascular invasion, which can be assessed only after diagnostic surgery and histological observation (4). However, in these cases, postoperative diagnoses reveal a high rate of unnecessary or inadequate surgeries (5). Several authors have suggested to further distinguish angioinvasive MI-FTCs (or moderately invasive FTCs) from MI-FTCs with tumor capsular invasion only (4, 6). However, it has been demonstrated that the extensive circumferential evaluation of the capsule not only allows to reach a higher effectiveness in the distinction between benign and malignant follicular neoplasms, but also yields a more adequate assessment of invasive vascular foci that are often underestimated by random histological section sampling. These results have highlighted that the presence of capsular invasion could be an indicator of vascular invasion (7). Although follicular lesions with vascular invasion were associated with a higher risk than those with solely capsular invasion (8, 9), only extensive vascular invasion (i.e. 4 or more vascular invasion foci) strongly correlated with a worse outcome (10, 11). Latest guidelines of the American Thyroid Association (ATA) then suggest a cutoff of 4 intracapsular vascular invasion foci to distinguish low- from high-risk MI-FTC (12). Even the degree of tumor capsule invasion for the diagnosis of MI-FTC is controversial; in fact, some authors have questioned whether a partial invasion may be sufficient for FTC diagnosis (13). However, some studies have justified FTC diagnosis in the presence of capsular invasion (even partial) without vascular involvement, since these tumors, although rarely, can develop recurrences and metastatic disease (14, 15). Despite the unresolved issues in MI-FTC diagnosis, it has been well demonstrated that MI-FTCs generally have a better outcome than WI-FTCs, with a low rate of recurrence, metastatic disease and disease-related mortality (2, 16, 17). For these reasons, the latest guidelines of the ATA suggested that lobectomy may be a sufficient treatment for low-risk FTC (i.e. MI-FTC with less than 4 vascular invasion foci) (12). However, molecular markers such as TERT promoter mutations were found to be an independent poor prognosis factor in differentiated thyroid cancer (DTC) (18, 19), especially in coexistence with BRAF or RAS mutations (20). Then, ATA guidelines have suggested that these molecular markers, alone or in combination, may be helpful to provide a more accurate risk-stratification of DTC, and in addition to the other clinico-pathological risk factors, may redefine the surgical approach to these lesions (12). In the last years, large-scale screening studies have disclosed many other potential biomarkers for the preoperative diagnostics of FTC, including both somatic mutations (21) and expression profiling (22, 23, 24), but none of them have been found to be conclusive. In the present study, we sought to find molecular biomarkers, which could help distinguish MI-FTCs from both FAs and WI-FTCs. Furthermore, the combination of gene expression analysis and genotyping could be useful for the stratification of MI-FTCs with similar clinico-pathological features.

Materials and methods

Study group

The study included 40 patients with diagnoses of FA, MI-FTC or WI-FTC, who underwent total/near-total thyroidectomy at the Department of Surgical, Medical, Molecular Pathology and Critical Area of the University of Pisa, Italy, between 2013 and 2015. Hematoxylin and eosin-stained sections of neoplasms obtained from the archives of the section of Pathology of the University Hospital of Pisa were re-evaluated independently by two pathologists (C U, F B). A diagnostic concordance rate of 98% was achieved between the two investigators. Discordant cases were eliminated. Tumors were diagnosed and classified according to the WHO 2004 histopathological criteria (3). For all FAs and MI-FTCs, the neoplastic capsules were examined in toto. Moreover, the vascular invasiveness and the infiltration of thyroid parenchyma and extra-thyroid tissues were evaluated. Hürthle cells neoplasms or neoplasms with oncocytic aspects were excluded from the study. To avoid contamination with other cell types, cases with any grade of thyroiditis were excluded, and only neoplasms with almost 50% of neoplastic cellularity were considered for the examination. This retrospective work was performed on archival material, and it was conducted anonymously; therefore, ethical committee approval was not required. The study conforms to the Principles of the Helsinki Declaration of 1975. Both informed and surgical consent were achieved one day before the operation.

Nucleic acids extraction and purification

The most representative FFPE tissue block was considered for each specimen. Two micrometer sections were taken from each block, stained with hematoxylin and eosin and histologically examined to confirm the presence of tumor tissue. The tumor area was marked and the tumor cell percentage was indicated by the pathologist. For each sample, 4 unstained 10 μm sections and 5 μm sections were used for DNA and RNA extraction, respectively. Unstained sections were deparaffinized with xylene and rehydrated in decreasing-grade ethanol solution. Manual micro-dissection was performed to maximize the amount of tumor cells. DNA was extracted and purified by using the QIAamp DNA Mini Kit (Qiagen), according to the manufacturer’s protocol. DNA was eluted in 50 μL of elution buffer. RNA was isolated using the RNeasy FFPE Kit (Qiagen) in conformance to the manufacturer’s instructions. RNA was eluted in 20 μL of RNase-free water. RNA and DNA quantification and quality were assessed by means of a spectrophotometer (Xpose Trinean, Gentbrugge, Belgium).

Detection of point mutations and rearrangements

The mutational status of NRAS (exons 2 and 3), HRAS (exons 2 and 3), KRAS (exons 2 and 3) and of the TERT promoter was tested by direct sequencing (3130 Genetic Analyzer, Applied Biosystems), according to the standard procedures (25). In addition, the presence of PAX8-PPARG rearrangements (exons 7 and 9) was evaluated by a two-step RT-PCR. Firstly, 600 ng of total RNA were used to synthesize cDNA by the RevertAid First Strand cDNA Synthesis Kit (Thermo Scientific) in a reaction volume of 20 μL. Secondly, 75 ng of cDNA were amplified in a 20-μL volume by Premix Ex Taq (Takara Bio), 4 ng of each primer (25) and 8 ng of probe (25), following the manufacturer’s suggestions. Amplification was performed in 40 cycles (denaturation at 95°C for 5 s and annealing and elongation at 56.5°C for 30 s) according to the manufacturer’s protocol. Each sample was amplified in duplicate, cDNA adequacy was tested by amplifying GAPDH housekeeping gene in double copy for each sample, and each assay included a double copy of the positive controls. Only the samples in which both GAPDH copies were amplified were considered suitable for rearrangement-positivity evaluation. The amplification of both GAPDH and PAX8-PPARG was positively considered when the cycle threshold was less than 35 cycles. The cases resulting positive for rearrangements were confirmed by direct sequencing (3130 Genetic Analyzer, Applied Biosystems).

NanoString analysis

The nCounter custom code set used in this study was designed and synthesized by NanoString Technologies (Seattle, Washington, USA). This code set consisted of reporter and capture probe pairs specific for the 75 genes and 5 housekeeping genes, which was already used by our group (26). In this assay, 150 ng of total RNA were hybridized with reporter and capture probes for 16 h at 65°C in a SensoQuest thermal cycler (SensoQuest, Gottingen, Germany). Sample clean-up and the counts of digital reports were performed as recommended by the manufacturer’s instructions. Five MI-FTCs were analyzed in triplicate to evaluate their potential intra-tumor RNA heterogeneity, by employing the RNA extracted from 3 different areas of the lesion. The first included the RNA from the entire tumor, the second from the non-invasive area and the third from the invasive foci.

Statistical analysis and bioinformatics tools

The differential expression analysis among FAs, MI-FTCs and WI-FTCs was performed using the procedures of the Limma R package (27). The design matrix contrasted all three group pairings by the voom transformation of counts (28). The Bayes function was used to compute the empirical Bayes moderated differential expression statistics. Genes with a false discovery rate (FDR) below 0.05 were considered to be differentially expressed between groups. Multiple discriminant analysis and leave-one-out cross-validation were then used to quantify the accuracy of classification of differentially expressed genes. In brief, two linear and orthogonal discriminant functions were calculated, based on the log₂ expression values of differentially expressed genes to maximize the distances between the averages (centroids) of the three groups. Specifically, the two discriminant functions were used to calculate a classification score for each sample, and centroids were calculated for each discriminant function as the average score of the samples belonging to the specific group. Each sample was then assigned to the group with the shortest distance to the closest group centroid. The accuracy of classification was calculated as the proportion of samples correctly classified within the actual group. Finally, leave-one-out cross-validation was done by computing the two discriminant functions in 37 samples (training set), and the remaining sample was used to test the accuracy of classification of the discriminant functions. Cross-validation was then run 38 times, each time with a different training set, and overall accuracy was calculated as the proportion of samples correctly classified within the actual group over all the 38 runs. Unsupervised hierarchical clustering was performed on normalized data with the nSolver 2.5 Analysis Software (NanoString Technologies). Clustering was applied independently to both samples (columns) and genes (rows) by Pearson’s correlation (r). The prediction of the effects of the newly reported amino acid substitution was performed using the PredictSNP tool (29).

Results

Clinico-pathological features

A total of 38 samples were eligible for inclusion in the study: 18 FAs, 14 MI-FTCs and 6 WI-FTCs. Every case had a complete follicular growth pattern and a total absence of oncocytic aspects and thyroiditis. Each selected tumor area had more than 60% of neoplastic cells. Seven of the 14 MI-FTCs showed capsular invasion only, while the others had both capsular and vascular invasion foci (less than 4). Furthermore, 7 MI-FTCs had only a partial invasion of the capsule. Four of the 6 WI-FTCs had extraglandular spread, whereas 2 were intrathyroidal tumors. Mean age at diagnosis was 48.1 ± 13.7 years, and mean tumor size was 29.1 ± 13.1 mm. Thirty cytological diagnoses, classified according to the Bethesda system (30), were available: 21 were category III, 8 were category IV and 1 was category V.

Genotyping results

All samples but 3 were suitable for genotyping analysis. None of the FAs harbored mutations, whereas 3 MI-FTCs were mutated (one had a PAX8-PPARG rearrangement and 2 were mutated both in TERT promoter and RAS genes, NRAS in one case and KRAS in the other); finally, 4 WI-FTCs harbored NRAS mutations (Table 1). The newly reported NRAS M72L mutation was found in heterozygosis and should not be deleterious according to the PredictSNP result.

Table 1.

Genotyping results of 18 follicular adenomas, 14 minimally invasive follicular thyroid carcinomas and 6 widely invasive follicular thyroid carcinomas.

	NRAS			KRAS	TERT promoter	PAX8-PPARG	TOTAL no. of mutated samples
	Q61R, n	Q61K, n	M72L, n	Q61R, n	C228T, n	ex.7, n
FAs	0	0	0	0	0	0	0
MI-FTCs	0	1	1	1	2	1	3
WI-FTCs	4	0	0	0	0	0	4
Total	4	1	1	1	2	1	7

Fas, follicular adenomas; MI-FTCs, minimally invasive follicular thyroid carcinomas; WI-FTCs, widely invasive follicular thyroid carcinomas.

Gene expression analysis of the 75-gene panel

All samples passed the quality control and were included in the differential gene expression analysis. Raw expression data were normalized using both housekeeping genes and positive controls (31). Samples analyzed in triplicate showed an average Pearson’s correlation of 0.90 (data not shown). Normalized expression levels were used to identify differentially expressed genes (DEG) among FAs, MI-FTCs and WI-FTCs based on a FDR <0.05. Twenty-five genes were deregulated between FAs and WI-FTCs; specifically, 14 and 11 were upregulated and downregulated in WI-FTCs compared to FAs, respectively; five were also differentially expressed between MI-FTCs and WI-FTCs; in particular, ECM1 and RXRG were upregulated and SLC26A4, SDPR and TFF3 were downregulated in WI-FTCs. Finally, TIMP1 was upregulated in MI-FTCs and WI-FTCs compared to FAs (Fig. 1 and Supplementary Table 1, see section on supplementary data given at the end of this article).

Figure 1.

Differentially expressed genes. Fas, follicular adenomas; MI-FTCs, minimally invasive follicular thyroid carcinomas; WI-FTCs, widely invasive follicular thyroid carcinomas.

Unsupervised hierarchical clustering

Unsupervised hierarchical clustering was performed on DEG normalized data using Pearson’s correlation (r). Two main clusters were obtained (Fig. 2): cluster 1 (r = 0.20) included 24 samples (18 FAs and 6 MI-FTCs), and cluster 2 (r = 0.17) consisted of 14 samples (8 MI-FTCs and 6 WI-FTCs). No correlations were found between MI-FTCs in cluster 1 and MI-FTCs in cluster 2 considering age, size, gender, presence of vascular invasion and degree of capsular invasion (partial or entire thickness). All mutated samples but one (1 MI-FTC harboring a PAX8-PPARG rearrangement) were encompassed in cluster 2.

Figure 2.

Unsupervised hierarchical clustering. Differentially expressed genes (rows) and samples (columns) were independently clustered by Pearson’s correlation (r). Red and green equal to high and low expression levels, respectively.

6-gene-based classification

The six DEGs (ECM1, RXRG, TIMP1, SLC26A4, SDPR and TFF3) between MI-FTCs and the other two types of lesions were taken into account to determine the centroids of each histological class (FAs, MI-FTCs and WI-FTCs), and each sample was classified according to the nearest class centroid. By following these criteria, 31/38 samples were appropriately classified (Table 2). A jackknife resampling was then used to cross-validate the results obtained, and 27/38 samples were correctly classified.

Table 2.

Six-gene-based classification of samples based on histology compared to age, size, capsule invasion degree, vascular invasion foci, cytological diagnosis (Bethesda system) and genotype.

Sample ID	Histology	Age (years)	Size (mm)	Capsule invasion degree	Vascular invasion foci	Bethesda category	Genotype	Prediction	Probability (%)
FA1	FA	41	35	None	None	III	WT	FA	76
FA13	FA	58	28	None	None	III	WT	FA	65
FA14	FA	19	18	None	None	IV	WT	FA	65
FA2	FA	55	25	None	None	III	WT	FA	82
FA25	FA	35	34	None	None	III	WT	FA	75
FA26	FA	34	18	None	None	IV	WT	MI-FTC	51
FA3	FA	31	44	None	None	IV	WT	FA	78
FA39	FA	50	20	None	None	III	WT	FA	54
FA40	FA	47	50	None	None	III	WT	FA	91
FA41	FA	51	15	None	None	III	WT	FA	63
FA42	FA	53	10	None	None	NA	WT	FA	90
FA70	FA	40	26	None	None	III	WT	FA	79
FA80	FA	41	30	None	None	III	WT	FA	90
FA81	FA	45	36	None	None	III	WT	MI-FTC	88
FA82	FA	49	25	None	None	NA	WT	FA	58
FA83	FA	50	42	None	None	III	WT	FA	81
FA85	FA	40	28	None	None	III	WT	FA	80
FA91	FA	35	52	None	None	NA	WT	FA	100
MI-FTC19	MI-FTC	45	15	Entire	<4	III	WT	MI-FTC	80
MI-FTC22	MI-FTC	64	26	Entire	<4	III	NRAS Q61K	WI-FTC	75
							NRAS M72L
							TERT C228T
MI-FTC30	MI-FTC	34	22	Partial	None	IV	WT	MI-FTC	91
MI-FTC33	MI-FTC	59	22	Partial	<4	III	WT	MI-FTC	58
MI-FTC36	MI-FTC	63	34	Partial	<4	NA	PAX8-PPARG ex. 7	MI-FTC	62
MI-FTC50	MI-FTC	45	50	Entire	None	III	WT	MI-FTC	67
MI-FTC51	MI-FTC	69	28	Partial	None	III	WT	MI-FTC	92
MI-FTC58	MI-FTC	55	22	Partial	None	IV	WT	MI-FTC	73
MI-FTC59	MI-FTC	28	18	Entire	<4	IV	WT	MI-FTC	83
MI-FTC60	MI-FTC	42	10	Partial	<4	NA	WT	FA	56
MI-FTC61	MI-FTC	39	18	Entire	None	NA	WT	MI-FTC	84
MI-FTC92	MI-FTC	41	22	Entire	None	III	WT	MI-FTC	74
MI-FTC93	MI-FTC	38	35	Entire	None	NA	WT	FA	63
MI-FTC94	MI-FTC	72	22	Partial	<4	III	KRAS Q61R	WI-FTC	93
							TERT C228T
WI-FTC4	WI-FTC	63	38	Entire	>4	III	WT	WI-FTC	70
WI-FTC48	WI-FTC	29	15	Entire	<4	III	NRAS Q61R	WI-FTC	100
WI-FTC49	WI-FTC	63	60	Entire	>4	IV	WT	FA	72
WI-FTC5	WI-FTC	53	NA	Entire	>4	NA	NRAS Q61R	WI-FTC	100
WI-FTC69	WI-FTC	77	NA	Entire	>4	V	NRAS Q61R	WI-FTC	100
WI-FTC75	WI-FTC	73	57	Entire	>4	IV	NRAS Q61R	WI-FTC	95

Probability refers to the reliability of classification according to the distance from histological class centroids.

FA, follicular adenoma; MI-FTC, minimally invasive follicular thyroid carcinoma; NA, not available; WI-FTC, widely invasive follicular thyroid carcinoma.

Discussion

In the present study, we performed a molecular characterization of 18 FAs, 14 MI-FTCs and 6 WI-FTCs, in order to find molecular biomarkers allowing to discriminate among these three types of lesions, which generally deserve a different surgical and therapeutic approach. Firstly, the potential intra-tumor mRNA heterogeneity of MI-FTCs was tested by analyzing five samples in triplicate: the expression profiles obtained from the RNA gathered from the entire tumor, from the invasive foci and from the non-invasive area were very similar (r = 0.90). The normalized gene expression levels were then used to determine DEGs among FAs, MI-FTCs and WI-FTCs. Twenty-five out of the 75 genes of the panel were differentially expressed (FDR <0.05) between FAs and WI-FTCs (Fig. 1), highlighting a great difference between these lesions, even from a molecular point of view. Five of these genes (ECM1, RXRG, SLC26A4, SDPR and TFF3) were also differentially expressed between MI-FTCs and WI-FTCs; in particular ECM1 and RXRG were upregulated in WI-FTCs, and SLC26A4, SDPR and TFF3 were downregulated. TIMP1, upregulated in MI-FTCs vs FAs, was the only differentially expressed gene in this comparison. These results were confirmed by an unsupervised clustering approach, which mirrored the DEG analysis: FAs and WI-FTCs were separated into two different groups, whereas MI-FTCs were split into the two main groups (Fig. 2). However, MI-FTCs that segregated in the two different clusters did not show differences considering age, size, degree of capsule invasion and even presence or absence of vascular invasion. The six genes differentially expressed between MI-FTCs and FAs or WI-FTCs (ECM1, RXRG, SLC26A4, SDPR, TFF3 and TIMP1) were then taken into account as a mini-panel, by testing the accuracy of their combination in predicting the histological class of the samples. Following this approach, 31 samples out of 38 were properly classified (Table 2), and more importantly, the prediction accuracy remained high after cross-validation (27/38). Furthermore, 2 MI-FTCs erroneously classified as WI-FTCs harbored both RAS and TERT mutations, the coexistence of which is strongly associated with recurrence and mortality (20); consequently, these 2 MI-FTCs should not be considered low-risk lesions, as suggested by the recent ATA guidelines (12). Then, the six-gene classification seemed able to identify high-risk MI-FTCs that did not show histopathological characteristics useful to stratify these lesions (e.g. diffuse vascular invasion, large size). Although the majority of inaccuracies were FA–MI-FTC misclassification, there was one exception: the WI-FTC classified as FA. This error underlines that the model needs to be further tested in order to evaluate how it could be useful in the clinical practice. Nevertheless, the faithfulness of the panel-based risk-stratification is supported by previous studies (32, 33, 34, 35, 36, 37) describing the biological role of these six genes in cancer progression and associating them with poor prognosis in many cancer models. TIMP1 protein is primarily an inhibitor of secreted and anchored metalloproteinases (38); however, its upregulation has been reported in several cancers, including melanoma, glioblastoma and breast cancer, in which it was also associated with poor prognosis (39, 40, 41). This apparent controversy could be explained by the protease-independent functions (32) and by the weak metalloproteinase-inhibitory capacity compared to other TIMPs (32). So far, TIMP1 expression has not been associated with adverse outcomes in thyroid cancer, but an increased expression with respect to normal tissue or benign lesions has already been described (23, 42). In our series, TIMP1 was also overexpressed in both MI-FTCs and WI-FTCs compared to FAs.

ECM1 is a soluble protein (43) overexpressed in many malignant epithelial tumors including thyroid cancer (33, 44). We found ECM1 to be highly expressed only in WI-FTCs and in the two MI-FTCs harboring both TERT and RAS mutations, highlighting its association with high-risk lesions. In fact, it was demonstrated that ECM1 is actively involved in promoting the Warburg effect (45) and angiogenesis (46) and, accordingly, that it is also associated with poor prognosis (33). RXRG encodes for a member of the retinoid X receptor family (47). This isoform was expressed similarly to ECM1, with few exceptions, confirming a previously reported high expression in thyroid cancer compared to normal thyroid cells (34). Interestingly, treatment with retinoids showed thyroid tumor growth suppression in vitro, especially in those expressing γ isoform (34). TFF3 is a secretory protein of the trefoil family (48). Although it is overexpressed and also correlates with poor prognosis in many cancer models (49, 50), we reported a downregulation of TFF3. However, this is in agreement with many studies that clearly demonstrated a lower expression of TFF3 in thyroid cancer with respect to normal tissue and benign lesions (22, 42, 51). Moreover, the forced expression of TFF3 in transfected anaplastic thyroid cancer cells restored an epithelial-like cell morphology and the expression of differentiation markers of follicular thyroid cells (35). SDPR is a phospholipid-binding protein that is over-expressed in serum-starved cells (52). SDPR downregulation was previously described in thyroid cancer (24) as well as in our series; the same trend was also reported in breast, kidney and prostate cancer (53). Furthermore, functional studies in breast cancer highlighted an important role of SDPR in apoptosis promotion (36) and in the suppression of the epithelial–mesenchymal transition (54). Lastly, SLC26A4 encodes for pendrin, a chloride–iodide transporter associated with the Pendred syndrome (55). SLC26A4 mRNA and pendrin levels are both considerably lower in thyroid cancer than those in the follicular cells of normal tissue or in benign thyroid tumors (24, 56). Pendrin, normally expressed at the apical pole of thyrocytes, is located in the intracellular compartments of the DTC majority (56). Despite the lack of evidence of association between SLC26A4 expression and poor prognosis in thyroid cancer, we reported a downregulation of this gene only in widely invasive carcinomas. However, low levels of SLC26A4 mRNA were associated with papillary thyroid cancer that failed to uptake ¹³¹I (37). In conclusion, even if we are fully aware that follicular-patterned tumors also encompass the follicular variant of papillary thyroid carcinoma, in this preliminary study, we focused on FA and FTC. Herein, we present a small panel of genes able to stratify benign, low- and high-risk lesions. Although the validation in a larger cohort of cases and through long-term follow-up is needed, our data provide some points that deserve further investigations: (a) since lesions carrying TERT promoter mutations have gene expression profiles very similar to those of WI-FTCs, these alterations could be useful markers for stratifying MI-FTCs prognosis; (b) gene expression and this six-gene model have the potential to support the diagnosis of indeterminate nodules. In fact, the gene expression levels of selected genes seemed promising not only in differentiating benign from malignant lesions, which can rarely be distinguished by FNAC, but also in differentiating between low- and high-risk malignant lesions, providing further information that could be useful in the surgical and therapeutic approach.

Declaration of interest

The authors declare that there is no conflict of interest that could be perceived as prejudicing the impartiality of the research reported.

Funding

This work was supported by AIRC (IG_10316_2010), PRA (PRA_2016_19) and PRIN (2015HPMLFY).