DNA Copy Number Variations Characterize Benign and Malignant Thyroid Tumors

Abstract

Context:

Fine-needle aspiration (FNA) is the best diagnostic tool for preoperative evaluation of thyroid nodules but is often inconclusive as a guide for surgical management.

Objective:

Our hypothesis was that thyroid tumor subtypes may show characteristic DNA copy number variation (CNV) patterns, which may further improve the preoperative classification.

Design:

Our study cohorts included benign follicular adenomas (FAs), classic papillary thyroid carcinomas (PTCs), and follicular variant PTCs (FVPTCs), the three subtypes most commonly associated with inconclusive preoperative cytopathology.

Setting:

Tissue and FNA samples were obtained at an academic tertiary referral center.

Patients:

Cases were identified that underwent partial or complete thyroidectomy for malignant or indeterminate thyroid lesions between 2000 and 2008 and had adequate snap-frozen tissue.

Interventions:

Pairs of tumor tissue and matching normal thyroid tissue-derived DNA were compared using 550K single-nucleotide polymorphism arrays.

Main Outcome Measure:

Statistically significant differences in CNV patterns between tumor subtypes were identified.

Results:

Segmental amplifications in chromosomes (Ch) 7 and 12 were more common in FAs than in PTCs or FVPTCs. Additionally, a subset of FAs and FVPTCs showed deletions in Ch22. We identified the 5 CNV-associated genes best at discriminating between FAs and PTCs/FVPTCs, which correctly classified 90% of cases. These 5 Ch12 genes were validated by quantitative genomic PCR and gene expression array analyses on the same patient cohort. The 5-gene signature was then successfully validated against an independent test cohort of benign and malignant tumor samples. Finally, we performed a feasibility study on matched FA-derived intraoperative FNA samples and were able to correctly identify FAs harboring the Ch12 amplification signature, whereas FAs without amplification showed a normal Ch12 signature.

Conclusions:

Thyroid tumor subtypes possess characteristic genomic profiles that may further our understanding of structural genetic changes in thyroid tumor subtypes and may lead to the development of new diagnostic biomarkers in FNA samples.

Fine-needle aspiration (FNA) is currently the best diagnostic tool for the preoperative evaluation of a thyroid nodule, but it is often inconclusive as a guide for subsequent surgical management, because 15%–20% of FNAs yield indeterminate results (1, 2). Recent studies have demonstrated that detecting mutations in BRAF, RAS, RET/PTC, and PAX8/PPARγ in clinical FNA samples contributes to the diagnostic accuracy of FNA cytology (3–6), although current assays are still insufficiently sensitive and specific.

In past decades, genetic gains and losses in thyroid cancers have been studied extensively (7–18), and comparative genomic hybridization (CGH) has shown that DNA copy number changes are frequent in follicular adenomas (FAs), and most of the changes are gains, often involving entire chromosomes, including chromosomes (Ch) 7, 5, 9 12, 14, and 17 (12). In comparison, DNA copy number changes and large chromosomal aberrations appear to be less common in papillary thyroid carcinomas (PTCs) and follicular variant PTCs (FVPTCs) (16).

Identifying characteristic copy number variation (CNV) patterns in these tumors may lead to an improved understanding of these clinically vexing lesions, which would be particularly useful in the case of FVPTCs, where morphological classification is subject to significant inter-observer and even intra-observer variation (19, 20). In this study, we used high-resolution single-nucleotide polymorphism (SNP) arrays to refine and extend current genomic characterizations of thyroid tumors. To maximize the statistical power of our initial analysis, we selected the 3 tumor subtypes most commonly leading to an ambiguous preoperative diagnosis: PTCs, FVPTCs, and FAs. Follicular carcinomas (FCs) are a much less common and were therefore not included in our initial genome-wide screen. Our focus on improving the preoperative diagnosis also led to selection of early-stage, well-differentiated tumors or adenomas, because later-stage tumors rarely pose a diagnostic dilemma.

Materials and Methods

A more detailed description of our methods and statistical procedures is available in Supplemental Methods (published on The Endocrine Society's Journals Online web site at http://jcem.endojournals.org).

Tissue samples and DNA isolation

The array study cohort included 14 FAs, 13 FVPTCs, and 12 PTCs as well as paired adjacent normal thyroid tissue. All samples were collected under an Institutional Review Board protocol approving molecular studies on excess tissue. Additional sets of tumor-normal paired tissue samples from 12 adenomatoid nodules (ANs), 18 FAs, 9 PTCs, 10 FVPTCs, 7 FCs, and 5 Hürthle cell carcinomas (HCs) were collected for an independent validation study. The clinical data on the study cohorts are summarized in Table 1 and in Supplemental Table 1. Genomic DNA was isolated using the DNeasy blood and tissue kit (QIAGEN, Valencia, California) according to the manufacturer's protocol.

Table 1.

Clinical Information Summary of Tissue Sample Cases Used in This Study

Tumor Type	Total (M/F)	Median Age, y	Median Size, cm	Tumor Stage (n)
Discovery patient cohort for SNP array analysis
FA	3/11	42	3.2
FVPTC	2/11	47	4	I (8), II (2), III (2), IV (1)
PTC	3/9	42.5	2.5	I (7), II (1), III (1), IV (3)
Validation patient cohort
FA	6/12	51	2.7
FVPTC	2/8	37	3.2	I (6), II (2), III (1), IV (1)
PTC	3/6	48	2	I (6), II (1), III (1), IV (1)
FC	5/2	55	4	I (4), III (3)
HC	2/3	56	3.5	I (1), II (1), III (2), IV (1)
AN	2/10	50.5	2.9
Total	23/61	46	3.2

Abbreviations: F, female; M, male.

SNP array analyses

A total of 750 ng genomic DNA was used to genotype samples using the Illumina 550K SNP array (Illumina, San Diego, California), following the manufacturer's guidelines (21).

Unsupervised cluster analysis

For heat maps and cluster analyses, segmented and smoothed copy number estimates for each sample were summarized at 25 000 base pair intervals over the entire genome to limit the dimensionality of the data while ensuring complete genomic coverage. For additional analyses, data were further filtered to select the regions of the genome showing the most variation across samples, as determined by calculating SDs for each 25-kilobase interval across the 39 thyroid tumors. Hierarchical cluster analysis was performed in the R statistical software suite (22) using agglomerative clustering with the Euclidean distance and complete linkage.

Identifying genes for which copy number is associated with tumor type

The smoothed, segmented data were mapped to genes, and an ANOVA as implemented in the Limma software package from www.bioconductor.org was performed.

We then evaluated the predictive capability of the most prominent CNVs identified to differentiate tumor subtypes, estimating error rates using leave-one-out cross-validation. To predict tumor type based on the predominant CNVs identified, we first used the ANOVA to select the genes for which copy number had the strongest association to tumor type and then calculated the average log-fold copy number change of those genes as a score. For an unbiased evaluation of performance, we embedded the entire gene selection and scoring procedure within a leave-one-out cross-validation and then performed a receiver-operator characteristic (ROC) analysis, reporting the area under the ROC curve as the primary performance measure (23).

Designing a validation study

To design an appropriate, independent validation study, we assumed that quantitative PCR (QPCR) measurements of copy number would be quantitatively similar to results obtained on the array, and sized the study to offer 90% power for detecting a difference between benign and malignant tumors using a 2-sided, 2-sample t test. From the SNP array, the average log ratio for the 5 genes in FAs was 0.143 (±0.133). In the PTC/FVPTC group, the average log ratio was 0.0002 (±0.052). Assuming a 2-sided, 2-sample t test, and conservatively estimating the overall SD to be 0.13, we determined that a total sample size of 37, or 18 per group, would meet this goal.

Real-time QPCR

Three sets of amplification primers were designed for each gene using Primer Express version 3.0 (Applied Biosystems, Foster City, California) with default parameter settings. The primer sequences used in this study are listed in Supplemental Table 4. QPCR was performed on all thyroid tumors and their matched normal samples, including the original tumors used for SNP array analysis and the additional validation set.

QPCR of FNA samples

FNA samples that matched FA tissue samples in this study were enriched for epithelial cells using a Dynal Epithelial Enrich kit (Life Technologies, Grand Island, New York) in accordance with the manufacturer's instructions. We calculated the normalized cycle threshold (Ct) value (ie, −ΔCt_[Target-Alu]) to represent the copy number relative to internal Alu sequence signal in thyroid FNA samples. As reference, 3 white blood cell (WBC) samples from patients with benign thyroid disease (multinodular hyperplasia) were used as normal control of Ch12 copy numbers.

Results

FAs, FVPTCs, and PTCs have characteristic genomic patterns

Using Illumina 550K SNP arrays, we investigated the genome-wide DNA copy number changes in 39 thyroid tumors (14 FAs, 13 FVPTCs, and 12 PTCs) with paired normal thyroid tissue samples from the same patients as controls (See Table 1 and Supplemental Table 1 for clinical patient information).

We performed an unsupervised hierarchical cluster analysis of our segmented and smoothed copy number estimates for each sample, summarized at 25 000 base pair intervals, and selected the 10% of segments with the greatest sample-to-sample variation in copy number. These regions were not evenly distributed throughout the genome but were concentrated over several chromosomes, most notably 7, 12, and 22, although all chromosomes were represented to some extent (see Supplemental Figure 2). The results are shown as a heat map in Figure 1, with three clusters standing out. Cluster 1 consists of 7 of 14 (50%) of the FAs, and 1 of 12 PTCs screened. These tumors exhibited a genomic amplification pattern predominantly involving Ch7 and Ch12, which is consistent with previous studies (8, 12, 15), although the rate observed here is higher than previous estimates. Most of the PTCs and FVPTCs clustered together in the center of the heat map, identified as cluster 2, where few CNVs were observed. This too is consistent with previous studies, which have found PTCs to be relatively stable genomically (10, 16). Finally, in cluster 3, a distinct subset of FVPTCs and FAs were characterized by large deletions in Ch22q, which are indistinguishable from monosomy 22 because of the lack of probes on the acrocentric 22p arm. Two of the samples with the Ch7 and Ch12 amplifications also harbored this deletion. Upon analysis of clinical and pathological parameters, we found the Ch22 deletion pattern to be associated with younger patients (32 vs. 46 years, P < .01, by 2-sided t test). No other significant associations with clinical indices or specific histopathological features, such as tumor stage or degree of encapsulation, were observed. All cases showing a BRAF mutation, including 2 cases of FVPTC, were in cluster 2.

Figure 1.

Unsupervised hierarchical clustering of 39 thyroid tumors. Only the 10% of segments with the greatest sample-to-sample variation in copy number, as measured by Illumina 550K SNP array, are shown. The tumor samples have been formally clustered on the x-axis in this analysis, whereas copy number is presented in genomic order on the y-axis. Individual tumors are shown as columns, with tumor subtypes shown in the colored annotation band along the top: FA (n = 14) in blue, PTC (n = 12) in deep pink, and FVPTC (n = 13) in orange. Each row of the heat map summarizes copy number in one 25-kilobase region of the genome, and in all, 11 426 such regions are represented here, selected for highly variable copy number and sorted in chromosome order. In the body of the heat map, copy number is color coded from bright green (homozygous deletion) to bright red (high-amplitude amplifications), as shown in the figure legend.

Chromosomal amplifications and deletions in FAs, FVPTCs, and PTCs

We performed a statistical analysis to identify significant CNVs as genomic amplifications and deletions (Supplemental Figure 2). The rule for calling significant CNVs depended on the number of SNPs involved as well as the magnitude of the copy number change and was designed to ensure that type I error did not exceed 10%. We identified a total of 464 CNVs as significant genomic aberrations (Supplemental Table 2). Chromosomal amplifications were more frequent in FAs than in FVPTCs or in PTCs (P < .01, χ² test, see Figure 2), occurring in ≥3 FAs at 7p, 7q, 12p, 12q, 17q, and 20q13.12. In PTCs, an amplification of 1q41 region occurred in 3 of 12 samples; and a deletion of 5q32 occurred in 2 samples. In FVPTCs, 7p11.21 was amplified in 4 of 13 samples, and deletions at 12p13.31 and the whole arm of 22q were also common.

Figure 2.

Overview of statistically significant copy number changes. The horizontal axis is the same for all 3 panels, showing genomic location, with chromosomal boundaries depicted as vertical lines. In the middle panel, where the vertical axis shows the 39 tumor samples grouped by subtype, all of the CNVs we identified as statistically significant by permutation test are represented, with deletions in green and amplifications in red. The remaining panels offer a view of the same data, summarized by tumor subtype, depicting the proportion of samples within each subtype having amplifications (top panel) or deletions (bottom panel) on each chromosome.

Identifying genes distinguishing benign FAs from malignant FVPTCs and PTCs

Our next step was to identify genes in which copy number differed by tumor type. We mapped the original segmented data to genes and performed an ANOVA, controlling type I error by the Benjamini-Hochberg false discovery rate at less than 10%. We found a total of 1209 genes for which DNA copy number showed significant differences (adjusted P < .05) between FAs and FVPTCs/PTCs. Most these genes were located on Ch7, Ch12, and Ch17. We found the dominant CNV pattern to be low-level but widespread copy number gain of Ch12 in FAs, as illustrated in Figure 3A-C, showing the mean fold changes across all samples on Ch7, Ch12, and Ch22, separated by tumor subtype.

Figure 3.

Mean copy number fold changes on Ch7, Ch12, and Ch22 in thyroid tumor subtypes. Calculations were performed after summarizing copy number by gene for each sample. A–C, Mean relative copy number on Ch7, Ch12, and Ch22, respectively. FAs are shown in blue, FVPTCs in orange, and PTCs in pink. In each case, the x-axis gives the physical position of each gene on the chromosome; with log fold copy number shown on the y-axis. Ch7 and Ch12 show widespread amplifications in many FAs, while Ch22 deletions are seen in subsets of the FVPTC and FA samples. A value of 0 corresponds to a ratio of tumor copy number to normal tissue copy number of 1. D, Log fold copy number for each sample on Ch12, calculated by averaging 10 genes selected by ANOVA to distinguish FAs from PTCs and FVPTCs. The horizontal line at log fold = 0.07 optimally demarcates benign and malignant tumors. E, Results of a cross-validated evaluation of this Ch12 gene panel by ROC, achieving an AUC of 0.88.

To obtain a gene set whose CNVs could distinguish benign FAs from malignant PTCs and FVPTCs, we selected the top 10 ranked genes on Ch12, ordered according to their statistical significances, and calculated their mean copy number changes within each sample. This resulted in a significant difference in mean copy number change (P < .001). Discrimination between classes was optimal at a cutoff of 0.07 for mean log fold copy number change, where we found that the 10-gene set could accurately classify 11 of 14 FAs and 24 of 25 PTCs and FVPTCs (Figure 3D). To evaluate the performance of this gene set in classifying our tumor series, we applied an ROC analysis for this 10-gene set, which resulted in an area under the curve (AUC) of 0.88 (Figure 3E). Leave-one-out cross-validation confirmed our result, accurately classifying 10 of 14 FAs and 23 of 25 PTCs/FVPTCs, with an AUC of 0.84, using the same cutoff of 0.07. Results were not sensitive to the number of genes used, remaining stable from 5 genes (AUC = 0.85) to at least 50 genes (AUC = 0.82).

Validation of Ch12 copy number changes

Our goals for the validation experiments were 2-fold: 1) a technical validation of the Ch12 signature using an independent, PCR-based assay and 2) investigating whether the CNV signature found in FAs was in fact specific to FAs or also present in FCs/HCs and FVPTCs on the one hand or in ANs on the other, given the morphological similarities between these follicular neoplasms. We selected 5 genes: NDUFA12, NR2C1, FGD6, VEZT (the top 4 ranked genes according to their statistical significance by ANOVA), and GDF3 (located at 12p13.31, a region showing amplifications in FAs and deletions in FVPTCs), averaging copy number levels across the 5 to obtain a single estimated value for each sample. The GenBank annotation for these genes can be found in Supplemental Table 3. Based on the distributions of the 5-gene score in benign and malignant tumors on the SNP array (Figure 4A), we performed a power analysis, estimating that we would require 18 additional FAs and 18 PTC/FVPTCs to have 90% power for detecting a difference in Ch12 amplification in an independent validation sample. The real-time QPCR analysis of copy number changes for these 5 genes independently confirmed our SNP array finding that FAs most frequently harbor Ch12 amplifications, both in the original 39 tumors (Figure 4C) as well as in an independent test set of 18 FAs and 19 malignant tumors, including 9 PTCs and 10 FVPTCs. We also tested 12 ANs and 12 samples from additional malignant tumor subtypes (7 FCs and 5 HCs). Although a small number of ANs showed elevated Ch12 CNV scores, FCs and HCs did not. Our gene expression array analysis of these 39 thyroid tumors (see Supplemental Methods) also showed that the average expression level of these 5 genes presented the same trend, confirming our finding on a complementary assay platform (Figure 4B).

Figure 4.

Validation of Ch12 copy number changes. Five genes selected for validation, NDUFA12, NR2C1, FGD6, VEZT, and GDF3, were averaged to obtain a single, composite value for each sample. Brackets identify statistically significant between-group differences using Welch's t test. *P < .05; **P < .01. A, Average relative copy number of the 5 selected genes for all samples of each tumor subtype, as measured on the SNP arrays. B, Expression of the 5 genes as measured by cDNA array. The log intensities from expression arrays normalized by matching normal thyroid tissue were averaged across genes to obtain a single estimated value for each sample. C, Copy number estimates as measured by real-time QPCR of genomic DNA. Estimated copy number changes from 15 primer pairs (3 primer pairs for each of the 5 genes) were averaged to obtain a single estimate of Ch12 relative copy number for each sample. In total, 100 thyroid tumor-normal paired samples were assayed, including the discovery set of 39 cases and additional samples from a test set of 7 FCs, 5 HCs, 10 FVPTCs, 9 PTCs, 18 FAs, and 12 ANs. For reference, the observed copy number changes for a chromosome 21 region in 3 Down's syndrome patients is shown as an example of a trisomy, whereas an X-chromosome region is measured in 9 normal males compared with 3 normal females as a surrogate for a monosomy.

Detection of the Ch12 amplification signature of FAs in matched FNA samples

To determine the clinical applicability of detecting CNVs in thyroid FNA samples, given the expected contamination with blood and WBCs, we then performed a small FNA feasibility study. Matching FNAs were available from 18 of the FA cases under study. All FNA samples were obtained intraoperatively after surgical isolation of the target lesion and stored in 95% ethanol. FNA samples were enriched for epithelial cells using magnetic beads, resulting in a total of 10 matching FNA samples with detectable amounts of DNA, as determined by achieving identifiable real-time PCR threshold cycle numbers. The results of the successful QPCR assays of this subset are shown in Figure 5. The samples are plotted separately based on their amplification status as determined by the tissue-based assays. The results clearly indicate that the Ch12 amplification signature is detectable and distinguishable from wild type in thyroid FNA-derived DNA, as long as sufficient epithelial cells are present in the sample.

Figure 5.

Real-time PCR assay of Ch12 amplification signature in thyroid tissue and matched FNA samples. Box plots on the left show fold copy number changes (relative to matching normal thyroid tissue) of Ch12 genes in 10 FAs for which both tissue and FNA samples were available. The left panel shows 8 cases (amplified, AMP) with fold copy number values consistent with amplification in tissue-derived DNA, whereas 2 cases (wild type, WT) showed no amplification. The right panel shows the result of the same real-time PCR assay in matched FNA samples after enrichment for epithelial cells. The normalized Ct value (−ΔCt_[Target-Alu]) represents copy number changes for FNA samples normalized for Alu elements, because no matching normal cell sample was available. For reference, results of the same assay on three WBC samples from patients with benign thyroid disease (multinodular hyperplasia) are shown.

Discussion

We have characterized the somatic genomic alterations in one benign (FA) and two malignant (PTC and FVPTC) thyroid tumor subtypes, focusing on these 3 because they are the most commonly associated with a suspicious but inconclusive preoperative cytopathology, reserving our much more limited FC samples for a validation of our screening results. In total, 39 thyroid tumor/normal pairs, including 14 FAs, 13 FVPTCs, and 12 PTCs, were analyzed using the Illumina 550K SNP array platform. To our knowledge, this is the first study to report genome-wide DNA copy number profiles comparing FA, PTC, and FVPTC thyroid tumors based on a high-resolution SNP array analysis.

The most frequent genomic aberrations occurred in FAs and included amplifications of Ch7 and Ch12, supporting results reported in prior CGH and array-CGH studies (8, 12, 15). Significantly, we can report that the frequency of such events in FAs is much higher than previously estimated using lower-resolution techniques. Conversely, with the notable exception of Ch22 deletions observed in several FVPTCs, both PTCs and FVPTCs showed relatively few copy number changes, confirming the current consensus that these are genomically relatively stable neoplasms (10, 14, 16), at least in their initial, well-differentiated stages.

The unsupervised hierarchical cluster analysis of detected CNVs in our series clearly shows some distinct patterns, which are simply identified in Figure 1 as cluster 1, 2, and 3. The consistent CNV patterns in cluster 1 found in many FAs on chromosomes 7 and 12 suggest that FAs showing these changes may represent a subset that may harbor a different developmental potential than structurally more stable FAs. Furthermore, because we did not find Ch12 amplifications in malignant tumor subtypes, this could indicate that FAs harboring this CNV signature are unlikely to progress and are not precursor lesions, in possible contrast to FAs showing Ch22 deletions, as discussed below. Because follicular neoplasms reflect a spectrum of disease with considerable morphological overlap rather than discreet entities, and the malignant potential of early-stage FVPTCs is often unclear and not always easily distinguishable from other follicular neoplasms (24, 25), it is possible that CNV patterns may help identify subsets of follicular neoplasms with different biological potential.

Similarly, although the number of cases showing Ch22 deletions is small, the consistency of the deletion patterns seen in several FAs and FVPTCs on Ch22 suggests that this genetic lesion may represent a distinct subset of these tumors. In this context, it is also worth noting that large Ch22 deletions and monosomy 22 have been associated with subsets of malignant follicular neoplasms (26, 27) and may therefore be indicative of precursor lesions. Except for a statistically significant association of the Ch22 deletion cluster with younger age, however, we were unable to correlate any clinical or pathological parameter with a particular CNV cluster. Of note, the 2 FVPTCs harboring BRAF mutations were in the PTC-associated cluster 2, supporting the notion that FVPTCs may broadly belong to either follicular or papillary tumors, each with distinct molecular and clinical signatures (24). Given the current inability to observe any evolution of these tumors, the biological implications of these findings will be resolved only by long-term studies of large cohorts of genetically well-defined cases.

Our most striking results, however, arose from a gene-by-gene comparison of copy number in our 14 benign and 25 malignant lesions of our discovery cohort. As seen in the cluster analysis in Figure 1, as many as 50% of the FAs showed distinctive amplification of Ch7 and Ch12. In particular, the panel of the top 10 genes showing significant copy number changes by ANOVA could distinguish FAs and PTCs/FVPTCs in all but 4 of 39 cases. The estimated copy numbers, although elevated, were moderate, suggesting that not all adenoma cells harbor a detectable copy number change, reflecting intra-tumor heterogeneity. The stromal component of well-differentiated thyroid tumors is typically minor and is therefore unlikely to strongly affect CNV patterns.

To confirm this result by independent methodologies, 5 genes, NDUFA12, NR2C1, FGD6, VEZT, and GDF3, were selected for validation using real-time genomic QPCR. We also analyzed gene expression array data for the same samples to see whether the amplification on Ch12 could be detected by such an approach as well. Both copy number changes, as assessed by QPCR, and gene expression, as assessed by transcriptome array, supported the presence of gene amplifications on Ch12 in FAs. In addition, a number of genes identified in an integrated analysis of gene expression and DNA copy number showed concordant results between DNA copy number change and gene expression levels. Not surprisingly, Ch12 was over-represented in this set, but similar results were observed in other regions as well.

Ch12 copy number changes were also confirmed in an independent test cohort that included both benign and malignant tumors, which again showed amplification in FAs, whereas other tumor subtypes, regardless of dignity or presence or absence of oncocytic cells, generally did not. This would suggest that FAs with amplifications on Ch12 are less likely to progress to thyroid cancer, because that genetic change would not be expected to disappear as FAs progressed. If confirmed, the ability to positively identify FAs with a low chance of malignant progression would be a useful adjunct to our current set of diagnostic tests that are focused on identifying oncogenic mutations and translocations in malignant thyroid tumors.

In light of these results, we then asked both study pathologists whether any distinct morphological patterns matching the Ch12 CNVs could be identified. Independent initial blinded and subsequent open reviews failed to identify a morphological subset in our FA cohort.

It is also noteworthy that among our samples in the morphological continuum ranging from AN to FA to FVPTC, small numbers of both ANs and FVPTCs harbored the Ch12 amplification characteristic of FAs, which may support a re-evaluation of these lesions based on molecular traits in addition to morphological characteristics.

It remains to be seen whether the 5 genes that we used to represent Ch12 have any functional roles in thyroid tissues or thyroid neoplasia, because they were selected based on the structural chromosomal changes we detected by our CNV analysis.

Finally, we performed an initial feasibility study to determine whether we could detect the Ch12 amplification signature in cytological specimens. The principal challenge in applying our quantitative genomic PCR assay to FNA samples is the unavoidable presence of varying amounts of blood contamination. We therefore fractionated our archival FNA samples using a commercially available magnetic bead separation approach, and the epithelial cell enrichment did lead to the correct classification of all 10 amplifiable DNA preparations, as shown in Figure 5. Of note, the magnetic bead separation was successful on archival FNA samples preserved in 95% ethanol for several years, and it is possible that yields may improve if the separation is performed on freshly obtained FNA material. How such an approach may fit into current clinical practice remains to be determined by follow-up studies aimed at assessing the sensitivity and specificity of PCR-based CNV marker assays in FNAs from the various thyroid tumor subtypes.

In summary, our study provides a high-resolution analysis of somatic copy number aberrations in FA, PTC, and FVPTC thyroid tumors. Our study showed distinct genomic patterns of copy number changes associated with benign and malignant thyroid tumors, of which the gene copy number gains in Ch12 were the most distinctive and which were limited to benign tumors. We have verified these amplifications using real-time PCR of genomic DNA and transcriptome arrays of the same 39 tumor-normal paired thyroid samples, and we validated the specificity of this result on an additional independent test set of benign and malignant thyroid tumors and demonstrated the feasibility of assessing CNV signatures in thyroid FNA samples.

Because FAs are a common source of inconclusive preoperative cytopathology results, a molecular signature such as Ch12 amplifications that positively identifies a subset of follicular neoplasms with no malignant potential could be a useful adjunct to the currently available tests for oncogenic genetic changes in thyroid cancers. Similarly, the presence of Ch22 deletions in FAs could be an indication of a premalignant state ultimately leading to invasive disease. Our results illustrate the value of the molecular characterization of benign thyroid tumors and well-differentiated thyroid cancer, which continue to confound the preoperative diagnosis of thyroid nodules, and may help justify the clinical development of molecular assays based on an epithelial cell-enriched fraction of the standard FNA sample.