Analytical performance of ThyroSeq v3 Genomic Classifier for cancer diagnosis in thyroid nodules

Abstract

Background

Molecular tests have clinical utility for thyroid nodules with indeterminate fine-needle aspiration (FNA) cytology, although their performance requires further improvement. In this study, we evaluated the analytical performance of the newly created ThyroSeq v3 test.

Methods

ThyroSeq version 3 is a DNA and RNA-based next-generation sequencing assay that analyzes 112 genes for a variety of genetic alterations including point mutations, indels, gene fusions, copy number alterations, and abnormal gene expression and uses a Genomic Classifier (GC) to separate malignant from benign lesions. It was validated in 238 tissue and 175 FNA samples with known surgical follow-up. Analytical performance studies were conducted.

Results

Using the training tissue set, ThyroSeq GC detected >100 genetic alterations, including BRAF, RAS, TERT, DICER1 mutations, NTRK1/3, BRAF and RET fusions, 22q loss, and gene expression alterations. GC cutoffs were established to distinguish cancer from benign nodules with 93.9% sensitivity, 89.4% specificity, and 92.1% accuracy. This correctly classified most papillary, follicular, and Hurthle cell lesions, medullary thyroid carcinomas and parathyroid lesions. In the FNA validation set, the GC sensitivity was 98.0%, specificity 81.8%, and accuracy 90.9%. Analytical accuracy studies demonstrated minimal required nucleic acid input of 2.5 ng, a 12% minimal acceptable tumor content, and reproducible test results under variable stress conditions.

Conclusions

ThyroSeq v3 GC analyzes five different classes of molecular alterations and provides high accuracy for detecting all common types of thyroid cancer and parathyroid lesions. Analytical sensitivity, specificity, and robustness of the test were successfully validated, indicating its suitability for clinical use.

Introduction

Thyroid cancer is the most common type of endocrine malignancy and its incidence has been continuously growing over the last several decades in the U.S. and many other countries. Thyroid cancer frequently presents as a thyroid nodule. However, thyroid nodules are far more common than cancer, with only a small fraction of all medically-evaluated nodular thyroid lesions found to be malignant or even neoplastic [1,2]. Nevertheless, the growing use of high-resolution imaging propels the detection of more and more thyroid nodules and subsequently thyroid cancer in what some pronounce to be an “epidemic” of this disease [3,4].

In thyroid nodules with clinical or ultrasonographic (US) suspicion for malignancy, fine-needle aspiration (FNA) with subsequent cytological examination of collected cells is typically performed [2]. FNA cytology provides a definitive diagnosis of benign or malignant nodule in 70-75% of cases, whereas the remainder of aspirates falls into one of three categories of indeterminate cytology defined by the Bethesda system [5,6]. These include Bethesda categories III, IV, and V, with the expected risk of cancer at 5-15%, 20-30%, and 50-75%, respectively. The uncertainty in cancer risk in these nodules precludes the optimal medical or surgical management of these patients and many of them undergo diagnostic surgery, which could be avoided in many patients with benign nodules.

Over the last decade, molecular testing has been increasingly used to improve the diagnosis and optimize management of patients with thyroid nodules that carry indeterminate cytological diagnosis. It progressed from single gene and small gene panels to broad genomic panels, multi-gene classifiers, and use of other molecular markers [7,8]. However, most of these approaches still have not reached the highest possible accuracy in the detection of all main types of thyroid cancer and may have limited performance in the populations with high pre-test probability of cancer. Detection of Hurthle cell (oncocytic) carcinomas remains problematic, particularly for low-grade cancers. In addition, for optimal performance, such diagnostic tests should accurately identify medullary thyroid cancer and non-thyroidal nodular lesions that occur in this location, particularly parathyroid lesions.

Further enhancement of the diagnostic performance can be achieved by expanding the existing tests and incorporating additional, more recently discovered molecular markers of thyroid cancer. Indeed, over the last several years a number of new driver mutations and gene fusions in different types of thyroid cancer have been discovered [9-13]. Furthermore, it was shown that a small but distinct proportion of thyroid cancers carry other types molecular alterations, such as copy number variations [10,12]. Importantly, rapid development and improvement of sequencing assays allow to detect multiple and various types of genetic alterations using a limited amount of cells collected by thyroid FNA [14].

We have previously developed and clinically used a next-generation sequencing-based ThyroSeq test, version 2, of which includes 56 thyroid-related genes analyzed for point mutations, gene fusions, and abnormal gene expression [15,16]. In this study, we report the development of a new, 112-gene version of the test. ThyroSeq version 3 was built with the goals to (i) expand the current ThyroSeq v2 test panel by including recently discovered genetic markers related to thyroid nodules and cancer, (ii) enable the analysis of additional classes of genetic alterations that were not previously tested, i.e. copy number alterations (CNAs), and (iii) improve test accuracy for detecting various types of Hurthle cell (oncocytic) tumors. Furthermore, we report the results of the studies of analytical performance of the new test which are required for its clinical use.

Materials and Methods

Tissue and FNA samples

The study samples included 238 surgically removed fresh-frozen and formalin-fixed paraffin-embedded (FFPE) tissue samples, 175 FNA samples with indeterminate cytology from nodules that were subsequently removed with known surgical pathology diagnosis, 16 cell lines, and 4 reference controls (Table 1, Supplemental Table 1). The study was approved by the University of Pittsburgh Institutional Review Board. Pathology glass slides from tissue sample of Hurthle cell lesions were blindly reviewed to verify the diagnosis and subclassify these nodules.

Table 1.

Tissue and FNA sample used for training and validation of ThyroSeq Genomic Classifier.

Tissue Type/ Pathology Diagnosis	Tissue training set (n = 238)	FNA validation set, (n=175)	FNA validation set by cytology diagnosis
			B-III	B-IV	B-V
Normal thyroid tissue	19	0	0	0	0
Hashimoto Thyroiditis	5	0	0	0	0
Hyperplastic nodule	25	55	35	19	1
Follicular adenoma	10	17	13	4	0
Hürthle cell hyperplasia	11	1	1	0	0
Hürthle cell adenoma	15	4	2	2	0
NIFTP	2	6	3	3	0
Follicular thyroid carcinoma	11	1	0	1	0
Hürthle cell carcinoma	29	2	0	2	0
Papillary thyroid carcinoma (PTC)	45	82	29	37	16
PTC, Hürthle cell variant	12	7	1	6	0
PDTC, ATC	6	0	0	0	0
Medullary thyroid carcinoma	15	0	0	0	0
Parathyroid lesion	13	0	0	0	0
Non-thyroidal tissue	20	0	0	0	0
Total	238	175	84	74	17

NIFTP, non-invasive follicular thyroid neoplasm with papillary-like nuclear features; PDTC, poorly differentiated thyroid carcinoma; ATC, anaplastic thyroid carcinoma

Next-generation sequencing and data analysis

Version 3 of the ThyroSeq test contains 112 genes, including all genes that were tested by ThyroSeq v2 (Supplemental Table 2). It utilizes targeted amplification-based next-generation sequencing technology to detect 12,135 single nucleotide variations (SNVs) and insertions/deletions (indels) (COSMIC hotspots), more than 120 gene fusion (GF) types, abnormal gene expression alterations (GEAs) of 90 genes, and copy number alterations (CNAs) in 10 genomic regions in FNA samples and in up to 27 genomic regions in tissue samples. In details, DNA and mRNA are isolated either from FNA sample collected directly to the nucleic acid preservative solution or from FFPE tissue or cell blocks [14,16] using MagNa Pure instrument (Roche). Nucleic acids concentrations are determined by the GloMax Discover System fluorometer (Promega) and Nanodrop 8000 spectrophotometer (Thermo Fisher Scientific). NGS libraries are generated from 10 ng DNA and 10 ng RNA using the Ion AmpliSeq Library kit 2.0 and Ion Xpress Barcode Adapters as previously described [14,17]. The libraries are normalized for template preparation on the Ion Chef and sequenced on the Ion Proton (Thermo Fisher Scientific) following the manufacturer's protocol. The Torrent Suite Software v5.2.2, (Thermo Fisher Scientific) and an in-house developed software Variant Explorer v2 are used for data analysis and interpretation.

The accuracy of next generation sequencing was determined independently for each class of genomic alteration by analyzing normal and pathogenic reference cell lines, reference sequencing controls and patient samples by the ThyroSeq v3 and an orthogonal method as a “gold standard” as per current NGS guidelines [18]. The sequencing performance characteristics are reported in the Supplemental Table 3.

Genomic Classifier (GC) Score

Each detected genetic alteration was annotated to receive a value from 0 to 2 based on the strength of its association with malignancy. The values were derived from (i) extensive literature and searchable database review (TCGA, cBioPortal, COSMIC, etc.), (ii) in-house database of >1000 thyroid surgical and FNA samples with known surgical outcome, (iii) RNA-Seq analysis of thyroid cancer tissue and FNA samples [9,19], and (iv) CytoScan analysis of 17 thyroid cancer tissue samples. Total GC score for each sample was calculated as a sum of individual values of detected genomic alterations (GC score = (x^SNV/I)n+x^GF+x^GEA+x^CNA; x = weighted value 0 - 2; n=number of SNVs/Indels; SNV/I, GF, GEA, CNA are indicators of genomic alteration type).

ThyroSeq v3 GC pre-analytical and analytical performance studies

The assay was developed in the University of Pittsburgh Medical Center Molecular and Genomic Pathology laboratory, which is certified under the Clinical Laboratory Improvement Amendments (CLIA) Act of 1988 to perform high complexity testing and accredited by the College of American Pathologists (CAP). The analytical performance studies were performed as recommended by the CAP Laboratory Accreditation Program (http://www.cap.org/) and New York State regulations (https://www.wadsworth.org/regulatory/). Specifically, DNA and RNA stability, analytical sensitivity, specificity, reproducibility and repeatability studies were performed (Supplemental Table 4).

Statistical Analysis

We calculated empirical sensitivity, specificity, accuracy and positive and negative predictive values with 95% Wilson confidence intervals [20]. The true positive fraction (sensitivity) and the false positive fraction (1-specificity) was plotted against the false positive rate in a receiver-operator-characteristic (ROC) curve. Area under the curve (AUC) was computed by the trapezoidal method with confidence intervals by the method of DeLong [21]. A two-step approach was applied; first the most accurate classifier was tested in the tissue training set. Next, the established score was evaluated in an independent validation set of FNA specimens with known surgical outcome. For the purpose of these analyses, cases of non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) were considered together with malignant nodules, as both tumor types require surgical excision. Twenty-five repetitions of 10-fold cross-validation were applied to test the classifier parameters stability both the training and the validation set.

Results

ThyroSeq v3 is a 112-gene, DNA and RNA-based, targeted next-generation sequencing assay that tests for five classes of genetic alterations: (i) point mutations, (ii) indels, (iii) gene fusions (GF), (iv) copy number alterations (CNAs), and (v) gene expression alterations (GEAs). The complexity of the analysis of the output data necessitated the establishing of a Genomic Classifier (GC) to classify the test result as negative (likely benign) or positive (likely malignant) (Fig. 1).

Figure 1.

Workflow of ThyroSeq v3 GC. Total nucleic acids from FNA or tissue samples are used to prepare and quantitate the DNA and RNA sequencing libraries and determine the overall sample adequacy. Samples with acceptable adequacy are sequenced. Data analysis starts with evaluation of cellular composition of the sample (follicular thyroid cells, parathyroid cells, C-cells, or non-thyroidal cells). Then, sequencing data from 112 genes undergo bioinformatics analysis to detect SNVs, indels, gene fusions, GEAs, and CNAs. Next, Genomic Classifier is applied to annotate the detected alterations and generation a score to classify the test result as Negative or Positive.

Establishing Genomic Classifier (GC) Using Training Set of Tissues

To establish a classifier, ThyroSeq v3 NGS analysis was performed on 238 surgically removed tissue samples. This set consisted of 205 thyroid tissue samples representing all main types of benign and malignant tumors and non-tumoral conditions and enriched in tumors negative for common genetic alterations, 13 parathyroid tissue samples, and 20 samples of other tissue types (Table 1). Overall, cancer-associated gene mutations (SNVs) and indels were identified in 55 samples, gene fusions in 35, CNAs in 47, and GEAs in 105 samples (Table 2). The most commonly detected alterations were SNVs/indels in the BRAF, RAS, EIF1AX, TERT, RET, DICER1, TP53, PTEN, and PIK3CA genes, gene fusions affecting the RET, PPARG, NTRK1, NTRK3, BRAF, ALK, THADA, IGF2BP3 genes, CNAs at 22q and NF2 genomic regions, and GEAs affecting MET, CALCA, and PTH. The GC score was calculated for each sample and found to range between 0 to 8. A ROC curve revealed an area under the curve (AUC) of 93.1% (95%CI:89.5-96.8) (Fig. 2A). Cross-validation analysis revealed similar test performance, with the AUC of 91.7%. The most accurate separation between benign and malignant samples was found with the GC score value of 1.5. Based on this analysis, test result yielding a score of less than 1.5 was designated as “Negative” (likely benign), whereas score ≥1.5 as “Positive” (likely malignant). Using this cutoff, the test sensitivity in the training set was 93.9% (95%CI: 88.4% – 96.9%), specificity 89.4% (95%CI: 81.1% – 94.3%), and accuracy 92.1% (95%CI: 87.8% – 95.0%). The test correctly classified 52 (91.2%) papillary thyroid carcinomas (PTC), 27 (93.1%) Hurtle cell (oncocytic) carcinomas (HCC), and 10 (90.9%) conventional follicular thyroid carcinomas (FTC). Two NIFTP cases were classified as positive. In addition, 13 (100%) parathyroid nodules and 15 (100%) MTC were correctly detected. All 20 non-thyroidal tissues were identified as being of non-thyroid origin.

Table 2. Prevalence of cancer-associated genetic alterations in the tissue training set.

Disease groups	Tissue type/Pathology Diagnosis	n	Genetic alterations
			SNV/indels	Gene fusions	CNA	GEA
Non-clonal, benign	Normal thyroid tissue	19	0	0	0	0
	Hashimoto Thyroiditis	5	0	0	0	0
	Hyperplastic nodule	25	0	0	0	0
	Hurthle cell hyperplasia	11	0	0	0	0
Clonal, benign	Follicular adenoma	10	0	0	0	1 (10%)
	Hurthle cell adenoma	15	1 (7%)	0	5 (33%)	5 (33%)
Pre-malignant	NIFTP	2	2 (100%)	0	0	2 (100%)
Malignant	Follicular thyroid carcinoma	11	4 (36%)	1 (9%)	7 (64%)	3 (27%)
	Papillary thyroid carcinoma	45	18 (40%)	22 (49%)	8 (18%)	32 (71%)
	Hurthle cell carcinoma	29	9 (31%)	4 (14%)	18 (62%)	5 (17%)
	PTC, Hurthle cell variant	12	4 (33%)	2 (17%)	4 (33%)	6 (50%)
	PDTC, ATC	6	5 (83%)	1 (17%)	1 (17%)	3 (50%)
Non-follicular cell	Medullary carcinoma	15	12 (80%)	0	4 (27%)	15 (100%)
	Parathyroid lesion	13	0	0	0	13 (100%)
	Non-thyroidal tissue	20	na	5 (25%)	na	20 (100%)

SNV, single nucleotide variant; Indels, insertions or deletions; CNA, copy number alterations; GEA, gene expression alterations; PTC, papillary thyroid carcinoma; ATC, anaplastic thyroid carcinoma; PDTC, poorly differentiated thyroid carcinoma; na, not applicable

Figure 2.

ROC curves for ThyroSeq v3 Genomic Classifier. A - ROC curve trained on surgically removed thyroid tissues. B - ROC curve based on the independent validation set of FNA samples with known surgical outcome.

ThyroSeq v3 GC Performance in a Validation Set of FNA Samples

Next, we tested GC in an independent set of 175 FNAs with indeterminate cytology (84 Bethesda III, 74 Bethesda IV, 17 Bethesda V) and known surgical outcome (Table 1). This set was purposely enriched in nodules found to be malignant after surgery, which constituted 52.6% of the set. In this sample set, gene mutations and indels were identified in 115 samples, gene fusions in 1, CNAs in 35, and GEAs in 79 samples. The total GC score ranged from 0 to 6. Applying the proposed cutoff of 1.5, ThyroSeq v3 GC revealed a sensitivity of 98.0% (95%CI:92.9% - 99.4%), specificity of 81.8% (95%CI: 71.8% - 88/9%), and accuracy of 90.9% (95%CI: 85.7% - 94.3%). The test ROC curve showed an AUC of 90.6% (95%CI: 85.6-95.6%) (Fig. 2B). The cross-validation analysis of the test results in this set of FNA samples revealed a similar AUC of 89.9%.

Performance of ThyroSeq v3 GC in Hurthle Cell Lesions

One of the goals of creating ThyroSeq v3 was to achieve a robust test performance in Hurthle cell (oncocytic) thyroid nodules. To achieve this goal, the training tissue set was enriched in all types of oncocytic lesions. Furthermore, blind re-review of histopathologic slides was performed to provide better separation between Hurthle cell hyperplasia (HCH) and Hurthle cell adenoma (HCA) based on histopathological features. Criteria for adenoma included complete moderately thick to thick capsule and a growth pattern clearly different from adjacent thyroid parenchyma, typically microfollicular or solid (Fig. 3, left and center). Based on the re-review of pathology, the training set included 11 HCH, 15 Hurthle cell adenomas (HCA), 29 Hurthle cell carcinomas (HCC), and 13 Hurthle cell variant of PTC (PTC,HCV). The analysis revealed cancer-related SNVs/indes in 14 of these samples, gene fusions in 6, CNAs in 27, and GEAs in 16 (Table 2). With the established threshold of 1.5, the classifier correctly detected 39 out of 42 (92.9%) Hurthle cell cancers. All 11 HCN had a classifier score 0, yielding a negative test result. Among HCA, 7 (46.7%) had a GC score 0-1, yielding a negative test result, and 8 (53.3%) had a GC score 2-4 yielding a positive test result (Fig. 3, right). Overall, in oncocytic lesions included in the training tissue set, ThyroSeq v3 GC had a sensitivity of 92.9% (95%CI: 80.52% - 98.50%) and specificity of 69.3% (95%CI: 48.21% - 85.67%).

Figure 3.

Histology and ThyroSeq v3 GC performance in Hurthle cell lesions. Left – Representative microscopic image of Hurthle cell hyperplasia (HCH) with no well-defined capsule and large follicles; Center - Representative microscopic image of Hurthle cell adenoma (HCA) with a complete thick fibrotic capsule and a microfollicular growth pattern which is clearly different from adjacent thyroid parenchyma. Right - Schematic representation of GC classification of various types of Hurthle cell nodules. HCC - Hurthle cell carcinoma; PTC,HCV - papillary thyroid carcinomas, Hurthle cell variant.

The validation FNA set included 5 benign Hurthle cell nodules (2 HCN and 3 HCA) and 9 cancers (2 HCC and 7 PTC, HCV). All 9 cancers were detected, with the GC score ranging from 2 to 5. Among benign Hurthle cell nodules, all five yielded a negative test result with a GC score of 0-1.

Sample Stability

Standard sample collection procedure for the ThyroSeq test includes collection into a tube with preservative solution with subsequent handling of the specimen at room temperature until shipment overnight in chilled boxes or storing at -20°C for a longer period of time. To test the stability of sample DNA and RNA under different stressed conditions, three ex-vivo FNA samples were created and tested in triplicate for different time and temperature conditions that simulate variations in FNA samples storage and shipping. The samples maintained acceptable quantity, quality and purity of DNA and RNA material under various tested conditions (Supplemental Table 5).

Next, we compared ThyroSeq performance in different sample types. While FNA samples collected directly into a preservative solution have the highest quality of nucleic acids, this test can be performed on fixed FNA cells from a cytology cell block and on formalin-fixed paraffin-embedded (FFPE) tissue samples. We performed sequencing analysis of 20 normal and tumor FFPE samples and compared the quality of sequencing to 20 FNA samples collected into a preservative solution. Although FFPE samples showed on average lower number of reads, it significantly exceeded the minimal acceptable depth of sequencing at 500 reads required for the assay (Figure 4).

Figure 4.

Comparison of ThyroSeq v3 sequencing depth in FNA and FFPE tissue samples. 500× is a sequencing depth of coverage cutoff for the assay.

Minimal Acceptable Nucleic Acid Input

The recommended input of nucleic acids for library preparations used for semiconductor sequencing technology per manufacturer's recommendation is 10 ng. We investigated the tolerance of the test results to variability in the nucleic acid quantity in order to determine the minimal acceptable input. Ten samples from cancer nodules (5 thyroid FNAs and 5 FFPE resected tumor samples) were diluted at 10 ng, 5 ng, 2.5 ng, 1 ng and 0.5 ng of nucleic acid input for library generation and sequencing. ThyroSeq v3 GC was able to correctly identify genetic alterations and classify as positive 100% of samples including both FNA and FFPE samples at 2.5-10 ng nucleic acids input, whereas only 50% of samples were correctly classified as positive at 1 ng (all FNA samples), and none of samples at 0.5 ng input of nucleic acids (Supplemental Table 6). These results indicated that the minimal input of nucleic acids tolerated by the assay is 2.5 ng.

Minimal acceptable tumor content

The proportion of neoplastic cells in thyroid FNA samples can have significant variability. To find the minimal acceptable tumor content that yield an accurate test result, we performed serial dilutions of DNA and RNA from 10 thyroid cancer samples (5 FNA and 5 tissue samples) in normal thyroid tissue samples. All 10 samples were correctly classified as positive down to the 12% tumor sample/88% normal thyroid sample dilution, whereas the proportion of correctly classified samples gradually decreased with further dilutions (Supplemental Table 7). Normal thyroid tissues were all classified as negative by GC. This data indicate that the accurate ThyroSeq v3 GC performance should be expected as long as the collected thyroid FNA sample contains ∼12% of tumor cells in a background of non-neoplastic thyroid cells.

Dilution with non-thyroidal cells

Thyroid FNA samples may contain various amounts of blood and other cells, such as white blood cells, tissue infiltrating lymphocytes, stromal fibroblasts, etc. To determine the minimal acceptable number of thyroid cells within FNA sample, we used three thyroid cancer samples (1 FNA and two tissue samples). The samples were diluted in a mixture of whole blood samples creating 24 in vitro mixtures at 8 concentrations of thyroid cells (Supplemental Table 8). ThyroSeq v3 GC correctly classified mixtures down to 12% of thyroid cells in all samples, down to 6% in 2 samples, and down to 3% in 2 samples, demonstrating that 12% of thyroid cell content is sufficient for correct classification. In addition, the samples were evaluated to define the test tolerance to blood content or method specificity. The results demonstrated that presence of blood did not interfere with any of the NGS analysis steps. Sample classification remained accurate up to 88% of blood content (Supplemental Table 8). All undiluted blood samples were called negative by GC.

In addition, the assay reproducibility, repeatability, and reagent stability studies were performed and demonstrated high accuracy in the assay performance (Supplemental Table 4).

Discussion

In this study, we report the results of analytical validation of a new targeted NGS-based multi-analyte test for thyroid FNA and tissue samples. ThyroSeq v3 GC interrogates five different classes of molecular alterations and was created to further improve sensitivity of cancer detection as compared to the currently available molecular tests [15,16,22]. Indeed, in this study the test, which was trained for maximum accuracy, achieved a sensitivity of 94% in the training set and 98% in the validation set of samples. If confirmed in a prospective clinical validation study, this would suggest that even in populations with high pre-test disease prevalence, the residual cancer risk in thyroid nodules with negative test result could remain at a sufficiently low level to consider observation in lieu of surgical excision for many of these patients (www.nccn.org).

High sensitivity of cancer detection in this study was achieved by testing for multiple classes of genetic alterations and also by offering high analytical sensitivity for detecting various types of diagnostic alterations in thyroid FNA samples. Indeed, as compared to the previous version of ThyroSeq, the current version was expanded not only to include larger number of genes tested for mutations, indels, fusions, and abnormal expression, but also to examine copy number alterations. The importance of chromosomal copy number changes is that they are found in ∼7% of papillary carcinomas lacking other driver mutations [10] and in other types of tumors such as Hurthle cell cancers [12].

The results of the study showed high sensitivity of detecting all main types of thyroid cancer by ThyroSeq v3 GC. This included not only most common papillary and follicular thyroid cancer types, but also Hurthle cell cancers by exploiting common occurrence of widespread copy number alterations resulting in a near-homozygous genome in these tumors [12]. Furthermore, ThyroSeq v3 GC retained accurate detection of medullary thyroid cancers and parathyroid nodules demonstrated by the previous version of the test [16,23].

An important parameter of the performance of each diagnostic test is its specificity for cancer detection. However, the existence of borderline tumors makes the assessment of specificity of tests for thyroid nodules more complex. One of the recently defined borderline tumor is non-invasive follicular thyroid neoplasm with papillary-like nuclear features (NIFTP) [24], and several other thyroid cancer precursor lesions are likely to exist, such as follicular and Hurthle cell adenomas. In contrast to hyperplastic nodules, such lesions are clonal neoplasms with at least some potential for progress to cancer. ThyroSeq v3 GC correctly classified as negative (likely benign) all histologically defined hyperplastic nodules, although about half of Hurthle cell adenomas were classified as positive. Further studies are required to identify molecular markers that can distinguish the precursor tumors from cancer, although it is likely that clinical benefit from excising such lesions exists as it prevents their potential progression to cancer.

Diagnostic tests intended for clinical use should demonstrate accurate and reproducible performance under variable stress conditions that may be encountered in clinical practice. The results of this study showed that ThyroSeq v3 has reliable test performance using as low as 2.5 ng input of the nucleic acids as long as at least 12% of tumor cells are present within the collected sample. This was demonstrated in the experiments performed in this study, which, together with the established high analytical sensitivity and specificity of the test indicate its suitability for use in clinical practice from the prospective of analytical performance. The clinical validity of this test for thyroid nodules with indeterminate FNA cytology is being evaluated in a prospective, blinded, multicenter study.