0021-972X/05/$15.00/0 The Journal of Clinical Endocrinology & Metabolism 90(5):2512–2521 Printed in U.S.A. Copyright © 2005 by The Endocrine Society doi: 10.1210/jc.2004-2028 Genetic Classification of Benign and Malignant Thyroid Follicular Neoplasia Base

Thyroid carcinoma is a common endocrine cancer with a fa- vorable prognosis if subjected to timely treatment. However, the clinical identification of follicular thyroid carcinoma (FTC) among patients with benign thyroid nodules is still a with a total of 114 samples. We were able to identify three genes, cyclin D2 ( CCND2 ), protein convertase 2 ( PCSK2 ), and prostatedifferentiationfactor( PLAB ),thatallowtheaccurate molecular classification of FTC and FA. Two independent validation sets revealed that the combination of these three genes could differentiate FTC from FA with a sensitivity of 100%, specificity of 94.7%, and accuracy of 96.7%. In addition, our model allowed the identification of follicular variants of papillarythyroidcarcinomawithanaccuracyof85.7%.Three-geneprofilingofthyroidnodulescanaccuratelypredictthediagnosisofFTCandFAwithhighsensitivityandspecificity,thusidentifyingpromisingtargetsforfurtherinvestigationtoultimatelyimprovepreoperativediagnosis.( Clin nol Metab 90:

Thyroid carcinoma is a common endocrine cancer with a favorable prognosis if subjected to timely treatment. However, the clinical identification of follicular thyroid carcinoma (FTC) among patients with benign thyroid nodules is still a challenge. Preoperative fine needle aspiration-based cytology cannot always differentiate follicular carcinomas from benign follicular neoplasias. Because current methods fail to improve preoperative diagnosis of thyroid nodules, new molecular-based diagnoses should be explored. We conducted a microarray-based study to reveal the genetic profiles unique to FTC and follicular adenomas (FAs), to identify the most parsimonious number of genes that could accurately differentiate between benign and malignant follicular thyroid neoplasia. We confirmed our data by quantitative RT-PCR and immunohistochemistry in two independent validation sets with a total of 114 samples. We were able to identify three genes, cyclin D2 (CCND2), protein convertase 2 (PCSK2), and prostate differentiation factor (PLAB), that allow the accurate molecular classification of FTC and FA. Two independent validation sets revealed that the combination of these three genes could differentiate FTC from FA with a sensitivity of 100%, specificity of 94.7%, and accuracy of 96.7%. In addition, our model allowed the identification of follicular variants of papillary thyroid carcinoma with an accuracy of 85.7%. Threegene profiling of thyroid nodules can accurately predict the diagnosis of FTC and FA with high sensitivity and specificity, thus identifying promising targets for further investigation to ultimately improve preoperative diagnosis. (J Clin Endocrinol Metab 90: 2512-2521, 2005) T HYROID CARCINOMA DERIVED from the follicular epithelial cell is the most common endocrine cancer. Papillary thyroid carcinoma (PTC) and follicular thyroid carcinoma (FTC) account for the great majority of all thyroid malignancies (1). An estimated 7% of the adult population (275,000 in 1999 in the United States alone) develops clinically significant thyroid nodules during their lifetime (2). The advent of thyroid ultrasound now allows for an increasing number of nodules to be diagnosed, and it is now recognized that nodules are present in an estimated 50% of the general population and are detected at a subclinical level. Because only 10% of these nodules will be a true malignancy, preoperative testing to differentiate benign from malignant nodules has been developed (3,4). Currently, fine needle aspi-ration (FNA) biopsy is the best diagnostic tool available for preoperative diagnosis. The FNA-based cytological diagnosis can be straightforward. However, approximately 20% (ranging from 9.2-42%) of all FNA will result in an inconclusive or suspicious outcome, especially if a follicular proliferation is found; the differentiation between a benign follicular neoplasia, especially follicular adenomas (FAs), and FTC based on the morphological features on FNA cytology is virtually impossible (5)(6)(7)(8). Therefore, because of the obvious difficulty in such preoperative diagnoses, surgical removal of the involved thyroid gland is routinely performed for diagnostic purposes in the setting of thyroid nodules and follicular cytology. However, in only 10 -20% of these cases would a follicular thyroid malignancy be found on final histology, resulting in unnecessary surgery for the vast majority of patients (4 -6, 8, 9). More importantly, false-negative cytologies can lead to delayed treatment with potentially serious consequences for the patient (10). Regarding the obvious limitation of FNA cytology in the preoperative diagnosis, there is a clinical need for new, reliable preoperative markers to distinguish benign from malignant thyroid nodules. Nonetheless, whereas numerous assays have been developed in an attempt to reduce these inconclusive preoperative diagnoses, none has yet proven more successful than FNA cytology in the clinical setting (4,(11)(12)(13). A possible underlying cause for this clinical problem is the continued limited understanding of the biological relationship of the different benign thyroid neoplasias to each other and to thyroid carcinoma, despite much research in this field (11, 14 -17). Therefore, to directly address the clinically relevant issue, we sought to elucidate further the molecular differences between benign follicular neoplasia and FTC. We took a global expression array approach to dissect out the minimal number of genes that can play a fundamental role in the early steps of FTC carcinogenesis, thus not only giving new biological insight but also allowing us to differentiate FTC, even at the minimally invasive stage, from benign follicular neoplasia based on a limited set of genes. We believe that this may form a basis for further investigation, in the hope that objective molecular markers will serve as an adjunct in the preoperative diagnosis of follicular thyroid cancer.

Tissue specimens
In total, 55 samples (24 FTC and 31 benign thyroid samples) were independently acquired for gene expression analysis in our training and validation set mentioned below. All tissue specimens were snap frozen in liquid nitrogen after surgical removal and stored at Ϫ80 C. Final histological classification for these samples was obtained from paraffinembedded tissue. In addition, sections from each snap-frozen tumor sample were independently subjected to hematoxylin and eosin stain and evaluated by a pathologist. A panel (training set) of 12 FTCs and 12 FAs were accrued for microarray (GeneChip) analysis (Table 1). No atypical variant or Hurthle cell adenoma was included in our set of 12 FAs. RNA extraction of these 24 samples was performed for GeneChip analysis and quantitative RT-PCR. Furthermore, seven follicular variants of PTCs (FV-PTCs) and additional tissue samples from five normal thyroids have been obtained from unrelated patients and RNA was extracted for quantitative RT-PCR. To validate our findings from the training set, two independent validation sets were also obtained as follows. The first validation set comprised in total 31 samples among which were 12 FTCs, 12 nonfunctioning thyroid nodules (five FAs and seven adenomatous nodules), five autonomous adenomas (hot nodules), and two normal thyroid tissues. The first validation series was subjected to quantitative RT-PCR. The second independent validation set comprised paraffin-embedded archival material from 57 patients with FTC [including 14 minimally invasive FTC and seven minimally invasive Hurthle cell carcinomas (HCC)] and 26 patients with benign thyroid nodules (17 FA and nine follicular hyperplasia) was subjected to immunohistochemistry (IHC). These samples were obtained through the Department of Pathology, The Ohio State University (Columbus, OH) and independently analyzed for histological diagnosis by the collaborating pathologist. All samples were obtained as anonymized materials without linked identifiers, with the approval of The Ohio State University's Institutional Review Board for Human Subjects' Protection.

RNA extraction
Total RNA was isolated from 0.2 g of snap-frozen tissue using the TRIzol Reagent (Invitrogen, Carlsbad, CA) and purified with the RNeasy Kit (QIAGEN, Valencia, CA). Aliquots of 1 g of total RNA were pretreated with DNase I (Invitrogen), after which 500 ng were reverse transcribed into cDNA using the SuperScript II System (Invitrogen) and a random hexamer anchored primer (Roche, Indianapolis, IN) according to the manufacturers' recommendations.

Oligonucleotide expression microarray analysis
Sample preparation, hybridization, and analysis were performed as described previously, except that version U133A GeneChips were used, which contain 22283 probe sets (17). In addition RNA quality was assured by using the Bioanalyzer 2100 (Agilent, Palo Alto, CA) in accordance to the standards described by Auer et al. (18). Furthermore, a detailed description of the microarray experiment, according to the MIAME criteria, is available online at http://www.ebi.ac.uk/ miamexpress/ (accession number E-MEXP-97). The cell intensity files (.CEL) were interrogated using the Affymetrix Microarray Suite 5.0 software. The percentage of probe-sets called present, the ratio of 3Јsignal to 5Ј-signal of two housekeeping genes, the intensity of four hybridization controls, the scale factor between arrays and signal-tobackground ratio were used for quality control assessment and to validate the in vitro transcription procedure. Furthermore, each array was cross-referenced to other arrays to identify array or single outliers by the method described by Li and Wong (19). All arrays passed these quality control steps. The DNA-Chip Analyzer Software (dChip) developed by Li and Wong (http://www.dchip.org) was used to normalize all arrays to a common array having a median overall brightness by using an invariant set of probes (19). A perfect match/mismatch difference model of the dChip software developed by Li and Wong was used to compute the model-based expression index (MBEI) (19). Raw data and computed expression values are available at http://www.ebi.ac.uk/miamexpress/. A summary table of the 80 differentially expressed genes is published as supplemental data on The Endocrine Society's Journals Online web site at http://jcem.endojournals.org (supplemental Fig. 1).

Quantitative RT-PCR
Quantitative RT-PCR was performed using the primers noted below and the iQ SYBR Green RT-PCR system (Bio-Rad, Hercules, CA) on an iCycler Instrument (Bio-Rad) using the comparative threshold cycle (Ct) method (20). Equal efficiency of the reference and target amplification was determined by a validation experiment for all reference and target genes. Samples were analyzed in triplicate for the target gene and normalized to the average Ct value of the two reference genes, ␤-actin and glyceraldehyde-3-phosphate dehydrogenase, the latter two of which were analyzed as duplicate. ⌬⌬Ct was determined by normalizing to the average ⌬Ct of five normal thyroid samples, indicating the relative difference in the expression level of the target gene between neoplasia and normal sample. The fold difference between FTC and FA is calculated by two to the power of the absolute difference in ⌬⌬Ct between the two groups. All values are given as means and 95% confidence intervals of each group. Primer sequences were as follows: glyceraldehyde-3-phosphate dehydrogenase, 5Ј-GGGCTGCTTTTA-ACTCTGGTAA and 5Ј-ATGGGTGGAATCATATTGGAAC; ␤-actin, 5Ј-CGTCATACTCCTGCTTGCTG and 5Ј-CCAGATCATTGCTCCTC-CTGA; cyclin D2 (CCND2), 5Ј-CACTTGTGATGCCCTGACTG and 5Ј-ACGGTACTGCTGCAGGCTAT; prostate differentiation factor (PLAB), 5Ј-CAACCAGAGCTGGGAAGATT and 5Ј-AGAGATACGCAGGTG-CAGGT; and protein convertase 2 (PCSK2), 5Ј-GCCATGGTGAAAAT-GGCTAA and 5Ј-GAGTGTCAGCACCAACTTGC. Primer sequence for ARHI and CITED1 have been described previously (21). Primers for quantitative RT-PCR were designed to span an exon-exon boundary or an intronic sequence, to avoid amplification of any genomic DNA. All quantitative RT-PCR products were initially visualized on a 2% agarose gel to ensure the presence of only a single amplicon product. The average sd between replicates was 0.15 and the average interassay sd for control genes was 0.32.

IHC
IHC was performed as described previously (22). Antibodies against CCND2 (Santa Cruz Biotechnology, Santa Cruz, CA) were used at a dilution 1:150 and against PCSK2A (US Biological, Swampscott, MA) were used at a dilution of 1:100. A total of 83 sections were analyzed, consisting of 57 FTCs and 26 benign thyroid nodules (17 FA and nine follicular hyperplasia). Additional sections from five normal thyroid glands and adjacent normal thyroid tissue were used for comparison. All slides were scored in a blinded fashion, and a second individual randomly validated the results. We regarded cells as immunoreactive when an obvious nuclear (CCND2) or cytoplasmic (PCSK2) expression was seen. We scored immunoreactivity as follows: retained (ϩϩ) when more than 50% of nuclei/cytoplasm were strongly immunoreactive, reduced (ϩ) when 10 -50% of the nuclei/cytoplasm were immunoreactive, and absent (Ϫ) when less than 10% of the nuclei/cytoplasm were immunoreactive or all cells' nuclei showed no immunoreactivity at all [supplemental Figs. 2 and 3 (published as supplemental data on The Endocrine Society's Journals Online web site at http://jcem.endojournals. org)]. The absence of a commercially available antibody that could reliable allow staining of thyroid tissue led to refine the IHC analysis to CCND2 and PCSK2.

Statistical methods
Two-tailed Student's t test for independent samples, assuming equal variance, was used to determine difference between mean gene expression determined by RT-PCR of the three selected genes with 22 degrees of freedom ( Table 2).
The hierarchical cluster analysis we used to present our data are based on 96 probe sets that we filtered from the 22283 probe sets present on the HG-U133A chip by setting the thresholds to 2-fold expressional changes at the lower 90% confidence bound in either direction, a P value less than 0.05 for the difference in expression and no less than 50% present call for each gene in all 24 arrays. For our cluster analysis we choose the commonly used average linkage method. The distance measure in the clustering analysis is 1 minus the correlation coefficient (23).
When the expression of a single gene is used for diagnosis, it becomes necessary to find a desirable threshold value that is used to distinguish the two groups. We obtained for each possible threshold value the sensitivity and specificity of diagnoses, which are percentages of FTC ("test positive") and FA ("test negative", i.e. not FTC) samples correctly identified, respectively. The best threshold value is the one that maximizes an appropriate combination of the two. To use multiple genes in combination for the purpose of diagnosis, we applied linear discriminant analysis, which is based on the assumption of multivariate normal distributions of the joint expressions, and finds the best linear combination of the expression values that discriminates the two groups. In a first round, we applied the technique of cross-validation to the training set to assess the performances of the diagnostic tests, in which each sample is in turn left out of the data, a test developed based on the remaining samples and then applied to the sample being left out. The diagnoses can be compared with the true classes of the samples to indicate the performance of the method leading to the diagnostic test. In a second round, we applied the same technique of linear discriminant analysis, but this time using our validation set, to independently confirm our findings from the first round.

Results
To dissect out the most parsimonious gene expressional differences that accurately classify FTC from benign follicular neoplasias, in particular FAs, we used a global expression array approach on 12 FTCs and 12 FAs ("training set"). So that we could also differentiate the earliest signs of malignancy from benign neoplasia, we included two minimally invasive FTCs and two minimally invasive HCC within our set of FTCs (Table 1). Using the dChip compare sample function, we used, as a first step, a straightforward but conservative approach to identify those genes that could reliably differentiate between FTC and FA. Using these criteria defined in the Materials and Methods section, we identified 96 probe sets, which represent 80 genes. To statistically validate these finding, we performed a random permutation analysis, in which we randomly permuted the labels of FTCs and FAs a large number of times, repeated the gene selection procedure using the same criteria, and recorded the number of genes identified (24). It demonstrated that these 80 genes were uncovered due to biological relevance and not by random coincidence (i.e. chance). Hierarchical cluster analysis showed that based on this set of 80 genes, FTCs and FAs could be accurately classified according to their histological group ( Fig. 1 and supplemental Fig. 1). Notably, three of four minimally invasive carcinomas and all HCC clustered within the FTC group. Only sample 03E192, a minimally invasive FTC, clustered with the FA group. From this set of 80 genes, we set out to find the smallest number of genes that could reliably classify FTC from FA in an independent validation set. After ranking the probe sets based on their fold change and significance (P value and t statistics), we identified those genes that also showed the highest difference in expression levels between minimally invasive FTCs and FA and we excluded expressed sequence tags and hypothetical proteins. Based on these criteria, we identified a list of 11 genes, and we focused, in the first instance, on the two highest ranking genes CCND2 (fold change Ϫ11.72; P value 0.0025), and PLAB (fold change 7.86; P value 0.0039; this gene has been annotated under different names, such as GDF-15, MIC-1, or com1) (25). Besides CCND2, we also found CD44, a gene targeted by the Wnt signaling pathway, markedly underex- pressed in FTC (fold change Ϫ4.5; P value 0.0016). In addition, Frizzled-1, the membranous receptor for Wnt ligands, is also dysregulated (fold change Ϫ4.39; P value 0.0081). Neither CCND2 nor PLAB have been previously associated with thyroid carcinogenesis. As a second step, we analyzed our gene expression data for probe sets with very high absent calls in only one group, either FTC or FA but not both, expecting that this approach will identify strongly under-expressed or silenced genes, which would in theory reliably differentiate these two histologies. Such high absent calls can lead to high P values, and consequently, the gene will not be detected by standard selection process. This approach revealed the gene encoding PCSK2 [present call 7% (MBEI 12.05) in FTC vs. 75% (MBEI 1743.51) in FA; fold change 144.7, P value 0.011] on further analysis. Expressional differences of each of the three genes between FTC vs. FA in the training set was confirmed using quantitative RT-PCR (summarized in Table 2).

Genetic classification of FTC and FA
Based on our microarray data from the training set of 12 FTCs and 12 FAs, we then employed different statistical methods to predict the performance of our selected three genes in the accurate and reliable classification of FTC and FA. We employed receiver-operated characteristics (ROC) curve analysis to evaluate the performance of our genetic classification using the expression of each of the three genes (CCND2, PCSK2, and PLAB) individually. The ROC curve shows the sensitivity (proportion of FTC samples correctly classified) and one minus the specificity (where specificity is defined as proportion of FA samples correctly classified, i.e. not carcinoma) from using all possible threshold values of expression in the classification (graph not shown). Because a very low false-negative rate is desired, and we note that to perfectly identify all FTC samples (12 of 12), the minimum proportions of misclassified FA samples based on our data are 33% (four of 12), 16.7% (two of 12), and 75% (nine of 12) when the expression values of CCND2, PCSK2, and PLAB are used separately. Of significance, when expression values of CCND2 and PCSK2 were used jointly in the classification by applying the method of linear discriminant analysis, the two groups of samples, FTC and FA, can be distinguished perfectly (24 of 24) (Fig. 2A). PCSK2 and PLAB have the same joint effect (Fig. 2B). To validate this microarray-based classification, we blindly analyzed the expression levels of CCND2, PCSK2, and PLAB in an independent validation set of 12 FTCs, 12 nonfunctioning thyroid nodules (five FAs and seven follicular hyperplasia), two normal thyroids and five autonomous adenomas (hot nodules). Linear discriminant analysis of this independent validation series confirmed that dual combinations of CCND2 and PCSK2 or PCSK2 and PLAB were able to distinguish between FTCs and FAs with an accuracy of 87.1% (exact 95% confidence interval 70.2-96.4%) (27 of 31 samples) and 93.5% (exact 95% confidence interval 78.6 -99.2%) (29 of 31 samples), respectively ( Fig. 2 and Table 3). Furthermore, because both hot as well as cold nodules could be accurately identified, we showed that the differences between the two groups are independent from functional status of the thyroid nodule but due to malignant transformation. For an honest estimate of the clinical performance using all three genes together, i.e. CCND2, PCSK2, and PLAB jointly, we applied the classifier from linear discriminant analysis, which correctly identified all 12 FTC samples from the validation set, and we estimated a false-positive rate of 5.3% (exact 95% confidence interval 0.13-26.03%) (1 of 19 samples) allowing an accuracy of 96.7% (exact 95% confidence interval 83.3-99.9%) (30 of 31 samples) ( Fig. 3 and Table 3).
Furthermore, we validated our data by means of IHC for the most promising combination of two genes, CCND2 and  Table 3. The joint performance of all three genes is demonstrated in Fig. 3.
PCSK2, in a second independent validation set of 57 FTCs and 26 benign thyroid nodules (supplemental Figs. 2 and 3). Using PCSK2 and CCND2 jointly (ROC curve in Fig. 4), we observed a sensitivity of 89.5% (exact 95% confidence interval 78.5-96.0%), specificity of 80.8% (exact 95% confidence interval 60.6 -93.4%) and accuracy of this test of 86.7% (exact 95% confidence interval 77.5-93.2%) when we chose the cutoff value for identifying FTC to be category 3 or larger (Table  4). Of note, complete absence of expression of PCSK2 and/or CCND2 was only seen in FTCs but never in benign thyroid nodules (Table 4). Furthermore, only 1 of 14 minimally invasive FTCs was misclassified due to retained immunostain for both antibodies, PCSK2 and CCND2. These observations affirm the accuracy of our genes to identify even minimally invasive neoplasias.

Genetic classification of FV-PTC
About 10% of suspicious FNA biopsies will be classified as FV-PTC in final histology. Therefore, we employed our three-gene based classifier system on a set of seven FV-PTC ( Table 5). Six of seven FV-PTC samples analyzed were correctly identified as a malignant thyroid neoplasia (85.7%). In addition, we used CITED1 and ARHI, two other markers previously described by us, to further characterize these samples. It is of note that one sample (FV-PTC_269) does not show expression of CITED1, a predictive marker for FV-PTC and PTCs. Interestingly, only in this sample we see a clear under-expression of CCND2 as seen in all other FTCs analyzed. Furthermore, sample FV-PTC_345 shows expression of CITED1, but was not identified by our three-gene profile as a malignancy. It is noteworthy that we found strong expression of the imprinted tumor suppressor gene ARHI in this sample. As we showed previously, silencing of this gene is associated with FTC carcinogenesis (21). These data might indicate that histological diagnosis of FV-PTC addresses a heterogeneous group of follicular neoplasia-an aspect that needs further elucidation. We note, by including the seven FV-PTC in our validation set, we can accurately identify 94.7% of all malignant samples (18 of 19) and 94.7% of all benign samples (18 of 19) as well.

Discussion
Currently, the diagnosis of thyroid nodules relies primarily on cytology (4,8). For the majority of patients with PTC,  Table 3.  (Table 4). When categories 3, 4, and 5 are considered to represent test positive cases (FTC), the classification of follicular neoplasias based on the protein expression of CCND2 and PCSK2 shows a sensitivity of 89.5% (exact 95% confidence interval, 78.5-96.0%), a specificity of 80.8% (exact 95% confidence interval, 60.6 -93.4%) and accuracy of 86.7% (exact 95% confidence interval, 77.5-93.2%; indicated by arrow in the curve), thus supporting the data derived from the more quantitative gene expression analysis (Fig. 3).
non-FTC, or inflammatory lesions, FNA-based cytology can make a diagnosis with high accuracy (4). However, there is a significant proportion of follicular neoplasias in which this FNA-based preoperative cytologic diagnosis fails (4 -6, 8 -10). Several reports show that individual skill and experience largely affect the sensitivity of this diagnostic test, ranging from as low as 57% to as excellent as 98% (10). However, an estimated 20% (ranging from 9.2-42%) of all performed FNA-based cytologies will describe a suspicious follicular neoplasia, but only 10 -20% of the patients that undergo surgery based on this diagnosis will actually have a malignant thyroid nodule (4,5,8). Based on investigative studies, immunohistochemical analysis has been proposed as a reliable marker for differentiating between FTC and FA (26). However, most of these markers showed their limitations in clinical practice and failed to become established (4,27). One underlying reason might be that neoplasias do not show their distinct malignant phenotype and therefore cannot be diagnosed by these methods. Different global gene expression studies have been conducted over the last years to identify novel targets. A recent study employing serial analysis of gene expression proposed a four-gene profile to improve preoperative diagnosis of FTC, but the accuracy of 80% for the gene expression based model is not superior to other algorithms (28). In addition other microarray-based studies, that allowed the highly accurate differentiation between FTC and FA by employing a 105-genes profile, still failed to identify minimally invasive FTCs, which comprise a large proportion of all FTCs (5,14). Our approach overcame this problem by including diverse phenotypes of follicular thyroid malignancies, especially minimally invasive variants, in the microarray-based training set. The inclusion of oncocytic variants of FTC (HCC) might appear distracting at first, because they are considered by some as a distinct clinicopathological entity and display unique molecular alterations (12,29). Other groups have identified molecular alterations such as RET/PTC translocations or BRAF mutations in a subset of oncocytic thyroid cancer (29 -31). Both these somatic alterations are common in PTC (15,29). However, it is acknowledged that morphological features defining PTC and FV-PTC can be found in Huerthle cell carcinoma as well (29). Therefore, other reports endorse the idea of Huerthle cell PTC or FV of Huerthle cell PTC (29). Unsupervised cluster analysis and multidimensional scaling failed to differentiate FTC and HCC into two distinct classes, indicating that in our sample set, the similarities in gene expression out-weigh in FTC and HCC the differences. These findings and other reports support our hypothesis that FTC and some HCC may result from shared molecular alterations (21). Nonetheless, this area requires further clarification and it remains important to identify HCC separately.
Our approach has allowed us to begin to identify genetic nuances in the initiation of follicular carcinogenesis. The dysregulation of CCND2, a cell cycle regulator, is intriguing because over-expression is associated with cancer progression and malignant transformation (32,33). However, there are emerging data that CCND2 may act in different ways beyond cell cycle control. Other reports showed that CCND2 is under-expressed in various cancers due to hypermethylation of its promoter (34,35). Our findings might provide further insight into the biological mechanism of CCND2 inactivation. Previous reports indicated that the dysregulation of the Wnt signaling pathway might play an important role in thyroid carcinogenesis (36). The membranous Frizzled receptors serve as binding targets for the Wnt proteins and subsequent activation of its intracellular Dishevelled proteins lead to transcription of targets genes such as CCND2 and CD44 (36,37). Our data demonstrate dysregulation of this pathway from the receptor to the target genes in FTC. Corroborating our findings, a previous report identified 11 genes of the Wnt pathway, including CCND2 and CD44, under-expressed in prostate cancer (37). This seeming paradox that both over-and underexpression of the same gene can result in carcinogenesis is being explained by accumulating data showing that different signaling pathways and its downstream targets may act as oncogenes in some neoplasms and tumor suppressors in others (38,39). Thus, fur- Values for the three-gene classifiers CCND2, PLAB, and PCSK2 are given in ⌬⌬Ct (see also Table 2). Six of seven FV-PTC (85.7%) have been correctly identified as malignant. Sample FV-PTC_345 was not identified as malignant. LDA value is the value of the linear combination of the three genes used to discriminate malignant and nonmalignant samples. Expressions of the two genes CITED1 and ARHI are marked with ϩ, whereas no detectable expression is labeled Ϫ. ther investigation would be required to determine how a profile of concurrent signaling pathways feed into directly opposed phenotypes.
The second gene we identified, PLAB, encodes a member of the TGF-␤ superfamily that is known to prevent apoptosis by activating the Akt pathway (25). The importance of Akt activation in follicular thyroid carcinogenesis has been previously shown by us (40). Therefore, PLAB might provide an upstream target of this pathway. Furthermore, an estimated 10% of all FNA do not result in sufficient material for a cytological diagnosis (4). Due to the lack of serum biomarkers that could identify FTCs, no preoperative noninvasive diagnosis is currently available for these patients. In this context, PLAB, a secreted protein, should be considered for further investigation to determine its feasibility as a diagnostic tool to identify thyroid malignancies from a simple blood test (41).
The third gene identified in our analysis is PCSK2. The members of this family process latent precursor proteins into their biologically active products. The mechanism by which the disruption of proprotein processing can promote tumorigenesis in thyroid tissue remains unknown. However, it has been shown that the inhibition of proprotein convertases enhances cell migration and metastases development of human colon carcinoma cells (42). Such a mechanism is plausible as well in thyroid carcinogenesis.
Even when we used only a combination of two of our three identified genes (CCND2 and PCSK2 or PLAB and PCSK2) we were still able to correctly classify 100% of the FTCs, including four minimally invasive ones, and all FAs. Indeed, using an independent validation series of 31 samples, we demonstrated that the combination of all three genes CCND2, PCSK2, and PLAB performed well in differentiating FTC from FA, resulting in an accuracy of 96.7% (exact 95% confidence interval of 83.3-99.9%). Furthermore, we were able to use a second validation series and a different technique, IHC, to examine a combination of only CCND2 and PCSK2, which resulted in an accuracy of 86.7%. Thus, our results appear to be superior to those reported using RT-PCR methods to detect gene expression of telomerase, galectin-3, or a number of other markers to discriminate benign from malignant follicular thyroid tumors (4,13,43,44). The employment of galectin-3 IHC has been reported to reliably identify malignant thyroid lesions (26,45). However, we and others have shown previously that this method does not succeed in improving the differentiation between FTCs and FAs in all cases (27,43). Furthermore, analysis by means of IHC often has its limitations, not only due to variability of antibodies or interinstitutional variation but also because of nonuniform classification and interpretation. In contrast, our gene expression analysis in a total of 24 FTCs and 31 benign thyroid nodules, using the combination of three genes, resulted in 100% of FTCs being identified and 30 of 31 of benign thyroid nodules definitively identified as well. A very recent FNAbased study employing hTERT as a molecular differentiator succeeded with recognizable sensitivity and specificity (46). However, the data indicate that this test performs much better in the identification of PTC and FV-PTC compared with FTC. Indeed, a full 20% of FTCs were missed. In addition, the performance of this test in identifying minimally invasive FTCs is unclear, and the authors conclude that additional molecular-based markers need to be explored (46). The robust results from our initial testing/training set confirmed by two independent validation sets have lent confidence that our three-gene test might help to establish a new and reliable molecular adjunct for diagnosis of follicular thyroid nodules in the near future.
There exist other studies that reported accurate differentiation of thyroid carcinomas, but notably, all these models were either based on high-density gene profiles (100 or more genes), which would not work in a presurgical diagnostic setting due to limited tissue and RNA available in such a setting, or do not provide the accuracy needed (13,14,28,47). Our classification model based on the limited number of genes, only three, provides the basis to pursue further evaluation. Whereas the technique to perform gene expression analysis in limited cell material has been well established (48), it needs to be shown how inadequate and/or contaminated FNA will affect the accuracy of our proposed test.
FV-PTC will be found in about 10% (range 0 -22%) of inconclusive FNA cytologies (5,6,49,50) and it is of note that when we employed our three-gene profile, we were able to identify FV-PTCs with an accuracy of 85.7%. Still, we need to acknowledge that FV-PTC might pose a special challenge when employing the three-gene predictor model into an FNA based setting. Our data indicate that the histological diagnosis of FV-PTC might describe a heterogeneous group of thyroid neoplasias. In this regard, it is of note that in a recent study by Lloyd et al. a concordant diagnosis of FV-PTC among 10 pathologists was made only in 39% of all cases (51). This high degree of observer variation can lead to a considerable bias of data if analysis is based on the unreviewed diagnosis of FV-PTC.
However, considering the recent studies that reported the differentiation between FV-PTC and FA using hTERT or CITED1, it may be plausible to use a four-gene test comprising CCND2, PCSK2, and PLAB plus hTERT (46,52). Therefore, there is accumulating molecular evidence that suggest that, in the near future, the majority of, if not all, thyroid malignancies can be targeted for definitive surgery, abolishing the requirement of a completion surgery (46,53,54). More importantly, most of the FAs that currently would have gone to unnecessary surgery would have been spared an extensive operation.
In summary, we have demonstrated that genetic classification of follicular thyroid neoplasia with a minimal number of three genes is highly accurate and may provide a tool to overcome the difficulties in today's preoperative diagnosis of follicular malignancies. It is hoped that the quantitative nature of such a test will be a useful gene-based objective adjunct to the preoperative diagnosis of a disease that currently relies solely on cytology.