Identification of Tissue-Specific DNA Methylation Signatures for Thyroid Nodule Diagnostics

This article is featured in Highlights of This Issue, p. 451

Every year, as many as 50,000 patients in the United States receive unnecessary thyroidectomies (1). These unnecessary thyroidectomies occur due to the difficulty in preoperatively distinguishing benign thyroid nodules from thyroid cancers. Although some thyroidectomies are required for patients with thyroid nodules for reasons other than cancer, it has been estimated that as many as 70% of thyroidectomies are unnecessary. Surgical complications from thyroidectomy include damage to the recurrent laryngeal nerves which results in a permanently hoarse voice or, in severe cases, tracheostomy (2, 3). Thyroidectomy may cause damage or removal of the parathyroids which may result in a lifelong need for calcium supplementation. Lifelong thyroid hormone replacement is necessary when the whole thyroid is removed and does occur in a significant percentage of patients when only part of the thyroid is removed.

Papillary thyroid cancer (PTC) represents about 75% to 80% of all thyroid cancers (4). The diagnosis of thyroid cancer is established via fine-needle aspiration (FNA) biopsy of thyroid nodules. In contrast to many other cancers, diagnosis of well-differentiated thyroid cancer is challenging due to frequent overlap of cytologic features between malignant and benign nodules from FNA. Many institutions have introduced the use of molecular tests for thyroid nodules with indeterminate cytology, which are observed in up to 40% of thyroid nodules (5). The most widely used molecular test for thyroid nodules, the Afirma Gene Expression Classifier, is associated with a high rate of overdiagnosis due to low positive predictive value (PPV; 47%) and can be performed only on fresh aspirates due to the instability of the RNA used in the test (6). Another diagnostic approach, ThyroSeqv2, is based on genetic alterations frequently found in benign nodules including H/K/NRAS mutations and RET/PTC rearrangements and is associated with a low PPV (42%–77%; refs. 7–13). These facts suggest that there is a need for a highly accurate thyroid nodule diagnostic test which uses time-insensitive material.

DNA methylation is a stable epigenetic modification, which plays an important role in cancer initiation and progression. Studies comparing thyroid cancer versus adjacent thyroid tissues revealed that DNA methylation patterns in PTC are strongly associated with mRNA and miRNA signatures, BRAFV600E, and H/K/NRAS mutations, and frequently reflect a histologic tumor type (14–17). However, use of this information for development of thyroid cancer testing can be problematic because (i) it is not adjacent thyroid tissue but benign nodules that undergo cancer diagnostic testing and (2) there are potential differences between benign nodules and adjacent thyroid tissues. For instance, in contrast to adjacent thyroid tissues, some benign nodules have genetic alterations frequently found in cancer including H/K/NRAS mutations and RET/PTC rearrangements (7–11). Thus, comparison of malignant to benign nodules rather than adjacent thyroid tissue should be used for the development of thyroid nodule diagnostics.

Here, we demonstrate that benign thyroid nodules have a unique DNA methylation pattern different from normal thyroid tissues and malignant nodules with PTC. Based on this epigenetic signature specific for benign thyroid nodules and a DNA methylation signature specific for PTC, we have developed a highly accurate Diagnostic DNA Methylation Signature (DDMS) approach for thyroid nodule diagnostics. Our study suggests that DNA methylation sites analyzed by DDMS potentially play a functional role in thyroid cancer development because these sites frequently colocalize with active enhancers in thyroid tissues and are associated with genes playing an important role in cancer development.

Developmental and validation retrospective cohorts were represented by surgical leftover thyroid nodule samples which were obtained from the City of Hope Pathology Research Services Core – Biorepository according to the institutional review board (IRB) protocols #97134 (J.H. Yim) and #04069 (P. Chu). IRB protocol #97134 (deidentified retrospective leftover discard specimens) falls under nonhuman subject research criteria according to City of Hope IRB guidelines and does not require patient consent. For IRB protocol #04609 (discard leftover specimens), a waiver of informed consent and HIPAA (Health Insurance Portability and Accountability Act of 1996) authorization was issued by the City of Hope IRB. Both studies were conducted in accordance with recognized ethical guidelines (e.g., Declaration of Helsinki, CIOMS, Belmont Report, and U.S. Common Rule) and were approved by the City of Hope IRB. Tissue hematoxylin and eosin (H&E) slides were blindly evaluated by two pathologists (S.-W.T. Tong and S. Chang). Specimens were included in the study if the pathologists agreed with the diagnosis (benign vs. malignant), and their diagnosis did not conflict with the original diagnosis of the nodule. Disputed cases were included and considered as malignant if they had BRAFV600E mutation or had metastasis.

DNA from fresh-frozen samples was isolated by using the phenol/chloroform extraction approach. For DNA isolation from formalin-fixed, paraffin-embedded (FFPE) samples, QIAamp DNA FFPE Tissue Kit was used (Qiagen Inc.). The reduced representation bisulfite sequencing (RRBS) procedure was previously described by Hahn and colleagues (18). DNA sequencing was performed on an HiSeq2500 Illumina instrument and aligned to the human genome at the City of Hope Integrative Genomics Core (GEO submission: GSE107738). The analysis pipelines were implemented using R statistical language. Refseq genes were downloaded from the UCSC HG19 genome annotation database. For the DDMS diagnostics, because information for all 373 DDMS sites was sometimes not available for each specimen, the benign and malignant scores were calculated based on only available information.

For BRAFV600E mutation screening, BRAF exon 15 was amplified by using primers (5′ GGGCCAAAAATTTAATCAG and 5′ TTCCTTTACTTACTACACCTCAGATA). Sequences of PCR products were evaluated by Sanger sequencing performed by the City of Hope Integrative Genomics Core.

Chromatin immunoprecipitation (ChIP) with fresh-frozen tissue was previously described (19). Enriched thyroid DNA fraction was repaired and linker ligated by using by using the KAPA Hyper Prep Kit (KAPA biosystems). Further, DNA amplified by using Phusion High-Fidelity DNA Polymerase (NEB), sequenced on an HiSeq2500 Illumina instrument, and aligned to the human genome at the City of Hope Integrative Genomics Core (GEO: submission: GSE107738).

We characterized DNA methylation patterns in 109 surgical thyroid specimens including 28 benign nodules, 39 PTCs, and 42 adjacent normal thyroid tissues (Table 1; Supplementary Table S1).

For DNA methylation profiling, we applied the RRBS approach, which provides single-base resolution for DNA methylation information for over 6 million cytosines genome wide (after generating at least 20 million aligned sequencing reads; ref. 18). Using these data, we searched for DNA methylation sites, which are characterized by tissue-specific DNA methylation patterns. We performed the following comparisons: malignant nodules versus adjacent thyroid tissues, malignant nodules versus benign nodules, benign nodules versus adjacent tissues, and benign nodules versus malignant nodules and adjacent thyroid tissues. We identified 18,593 cytosine locations which demonstrated a significant difference (adjusted FDR < 10E–2) between analyzed tissue groups (Fig. 1). Based on the clustering analysis of these locations, PTC is associated with unique DNA methylation patterns. At the same time, benign nodules are frequently characterized by an intermediate epigenetic pattern between thyroid tissues and malignant nodules (Fig. 1; see intermediate clusters #1 and #2). However, 4,575 cytosines demonstrated a unique DNA methylation pattern specific to benign nodules. These DNA locations were associated with intensive DNA hypermethylation in benign nodules compared with malignant nodules and adjacent thyroid tissues. These data suggest that the DNA methylome of benign nodules differs from methylomes of adjacent thyroid tissue and malignant nodules, and this fact should be taken into account for development of diagnostics for thyroid nodules. In this light, it is very important to note that some adjacent thyroid tissues demonstrated a very similar pattern to the adjacent nodules which suggests a field cancerization effect in thyroid cancer (Fig. 1).

We observed that a group of specimens with lymphocytic thyroiditis is clustered together and characterized by unique DNA methylation patterns (Fig. 1). In order to evaluate if a high lymphocyte presence is responsible for the DNA methylation patterns in this “thyroiditis” tissue cluster, we profiled the DNA methylation pattern in lymph nodes as a tissue highly enriched with lymphocytes. Clustering analysis revealed that the lymph node had a very similar DNA methylation pattern to the “thyroiditis” cluster (Fig. 1). These data suggest that a high presence of cells other than follicular cells in the analyzed thyroid tissue may interfere with DNA methylation pattern diagnostics.

By using DNA methylation data for 109 specimens, we evaluated the presence of thyroid cancer–specific and benign nodule–specific patterns. We detected 373 genomic locations with tissue-specific epigenetic patterns in benign and malignant nodules (Supplementary Table S2). Based on these 373 differentially methylated genomic locations, we developed a DDMS approach. According to our study, 62% (234 of 373) of DDMS locations were linked to gene bodies or gene promoters (Fig. 2A). We identified 236 locations which are associated with the cancer-specific pattern and colocalized with 93 genes (Supplementary Table S2). At the same time, 137 DDMS locations were associated with the benign-specific pattern and colocalized with 64 genes. According to Ingenuity Pathway Analysis (Qiagen Inc.), genes associated with tissue-specific signatures are linked to cancer and involved in cell proliferation, invasion, and migration (Fig. 2B; Supplementary Fig. S1; Supplementary Table S3). Genes associated with tissue-specific epigenetic signatures include gastrin-releasing peptide (GRP), high mobility group AT-hook 1 (HMGA1), Thy-1 cell surface antigen (THY1), tyrosine kinase 2 (TYK2), and GATA-binding protein 6 (GATA6). Further, we investigated whether tissue-specific DNA methylation signatures are associated with gene regulatory elements. We performed genome-wide profiling of a chromatin mark of active enhancers, H3 lysine 27 acetylation (H3K27Ac), by using ChIP-sequencing with two normal adjacent thyroid tissues. A representative example of H3K27Ac pattern in adjacent thyroid tissues is shown (Supplementary Fig. S2A). Our data revealed that approximately 35% of DDMS locations are associated with H3K27Ac-enriched regions in at least one of the adjacent tissues samples (Fig. 2C; Supplementary Table S2). Examples of colocalization of active enhancers with DDMS locations are shown for the genes ANGPTL4 (angiopoietin like 4), MUC1 (mucin 1, cell surface associated), NFIC (nuclear factor I C), and EPHA2 (EPH receptor A2; Fig. 2D and E; Supplementary Fig. S1B and S1C). At the same time, approximately 54% of DDMS locations are located within H3K27Ac peaks or located at least within a 1 kb area surrounding H3K27Ac peaks, so-called “active enhancer shores.” These data suggest that DNA methylation at DDMS locations potentially plays a regulatory role in thyroid cancer development.

We proposed that thyroid cancer diagnosis can be accomplished by a combination of both benign and thyroid cancer patterns/signatures, where each individual thyroid nodule is characterized by both the number of epigenetic changes specific to benign nodules and the number of epigenetic alterations specific to malignant nodules.

In order to validate our DDMS approach, we performed a statistical analysis based on the leave-one-out cross-validation technique (20). In order to predict benign and cancer scores for 67 nodules, 67 benign and cancer unique predictive signatures were developed by using 109 specimens.

The leave-one-out cross-validation revealed that 11 of 67 analyzed specimens are associated with fewer than 5 out of the 137 epigenetic changes specific to benign nodules and fewer than 5 out of the 236 epigenetic changes specific to PTC (Supplementary Table S1). We classified these specimens as undiagnosable because (i) our approach is based on the identification of the tissue-specific epigenetic signatures and (ii) undiagnosable tissues are missing tissue-specific epigenetic alterations/signatures. These undiagnosable specimens were excluded from further analysis.

The leave-one-out cross-validation indicated that some specimens are associated with the presence of both epigenetic scores, benign and malignant (Supplementary Table S1). In order to diagnose these types of specimens, we introduced a cancer risk score that is defined as the ratio of the cancer score and the benign score [cancer risk score = (malignant score + 1)/(benign score + 1)]. We classified nodules with a cancer risk score above or equal to 1.0 as thyroid cancer, whereas the specimens with a cancer score below 1.0 were classified as benign nodules (Fig. 3A). Based on these criteria, 23 out of 24 benign nodules were classified as benign, and 30 out of 32 malignancies were classified as cancer (Fig. 3B). Both wrongly diagnosed malignant nodules (#63T and #112T) were Hurthle cell lesions but had more than 75% of adjacent thyroid tissue in the specimens according to H&E staining analysis (Supplementary Table S1). Because (i) sensitivity is defined as the proportion of true positive in all positively classified samples and (ii) we classify thyroid nodules as malignant when a specimen is characterized by at least one epigenetic signature and cancer risk score equal to 1.0 or more, the estimated sensitivity of DDMS is 94% (Fig. 3C). On the other hand, because specificity is determined as a proportion of true negative in all negatively classified samples and (ii) we classify benign nodules as nodules with at least one epigenetic signature and cancer risk score less than 1.0, the estimated specificity of our test is 95% (Fig. 3C). Thus, based on the leave-one-out cross-validation technique in the developing nodule cohort, DDMS has an estimated specificity of 96%, sensitivity of 94%, PPV of 97%, NPV of 92%, and accuracy of 95% (Fig. 3C).

In order to evaluate mechanisms responsible for an absence of epigenetic signatures in undiagnosable specimens, we investigated an association between lymphocytic thyroiditis and undiagnosable specimens (Fig. 1). We found that all nodules associated with the “thyroiditis” tissue cluster were undiagnosable according to DDMS, and 4 out of 11 undiagnosable nodules were associated with thyroiditis (Fig. 1).

Further, we evaluated the potential impact of Hurtle cell presence on the DDMS epigenetic signatures (Fig. 1; Supplementary Table S1). According to the pathologic evaluation, the development cohort contained nine nodules with some presence of Hurthle cells and ten Hurthle cell lesions. From ten Hurthle cell lesions, only two nodules were undiagnosable. Both nodules were associated with the “thyroiditis” tissue cluster and had thyroiditis (Fig. 1). At the same time, only one of nine nodules with some presence of Hurthle cells was undiagnosable.

Because some FNAs are falsely acquired from adjacent thyroid tissue, we evaluated DDMS performance in adjacent thyroid tissue. We performed the leave-one-out cross-validation for 42 adjacent thyroid samples by using our 109-specimen cohort (Supplementary Table S4). According to our study, only 38% (16 out 42) of adjacent tissues had epigenetic signatures. Fourteen out of these 16 adjacent tissues were classified as benign, when 2 adjacent tissues were classified as malignant. One wrongly classified adjacent tissue (113N) was associated with a very similar DNA methylation pattern to the adjacent malignant nodule (Fig. 1). This similarity can be potentially explained by a field cancerization effect. For another wrongly diagnosed normal adjacent tissue sample (55N), it is not clear whether the wrong diagnosis is associated with a field cancerization effect. The status of the nodule #55 located next to this normal adjacent tissue 55N is unknown due to a disagreement of the participating pathologists regarding the diagnosis of the nodule.

Further, we asked whether a high level of adjacent normal tissue in a biopsy might affect a DDMS diagnostic result. First, we determined the proportion of cancer cells in each malignant tissue by using H&E staining. We found 5 specimens 63T, 102T, 112T, 117T, and 119T with approximately 25% or less of malignant cells in the specimen. Two (102T and 117T) out of five of these specimens were undiagnosable when the other two (#63T and #112T) were misdiagnosed as benign. Thus, undiagnosable specimens according to DDMS are frequently associated with a high presence of nonnodule cells including lymphocytes and adjacent tissues.

We performed validation of DDMS performance in a testing cohort by using RRBS. Our testing cohort contained 78 postsurgical thyroid nodules: 41 malignant and 37 benign thyroid nodules (Table 1; Supplementary Table S1). We observed DNA methylation signatures in 34 malignant and 31 benign nodules. It is important to note that 8 out of 13 undiagnosable specimens had areas of lymphocytic thyroiditis (Supplementary Table S1). At the same time, only 3 out of 13 undiagnosable nodules were associated with Hurthle cells, whereas two out of these three nodules had thyroiditis. All 34 cancers with epigenetic signatures were classified as malignant according to DDMS, and 30 out of 31 benign nodules were classified as benign (Fig. 4A; Supplementary Table S1). These data suggest that DDMS has an estimated specificity of 97% [95% confidence interval (CI), 80–100], sensitivity of 100% (95% CI, 86–100), PPV of 97% (95% CI, 82–100), and NPV of 100% (95%, 85–100; Fig. 4B). Therefore, DDMS can serve as a potential diagnostic tool for thyroid cancer in specimens with epigenetic signatures.

In order to reduce costs, we developed a multiplex bisulfite PCR (MB-PCR) approach for the DDMS validation in thyroid nodules (Fig. 5). This method is based on multiplex PCR with bisulfite-converted DNA followed by ligation with a linker (a short double-stranded oligonucleotide) and next-generation sequencing (Fig. 5A). MB-PCR demonstrated high reproducibility (R = 0.97; P < 0.0001) for the lowest range of DNA amount obtained by FNA, 100 ng (Fig. 5B; ref. 21). The DNA methylation status of cytosines from the DDMS panel strongly correlates between RRBS and the newly developed MB-PCR for thyroid nodules with an average of correlation coefficient (R = 0.93; P < 0.0001; Fig. 5C). Thus, this inexpensive MB-PCR approach (under $100 per sample) based on DDMS can be used for thyroid nodule diagnostics.

The evaluation of thyroid cancer has been very problematic due to a frequent overlap of cytologic features in benign and malignant nodules. Though current molecular testing methods have improved the accuracy of evaluation, their performance still has the potential for improvement. For instance, a classification of “suspicious” from the most widely used molecular test, the Afirma Gene Expression Classifier, is seen in up to 60% of benign thyroid nodules (6). ThyroSeqv2, which is based on known gene mutations and translocations, deems up to 58% of benign thyroid nodules as “suspicious” (12, 13). By contrast, our test is associated with PPV of 97% (95% CI, 82–100) according to the validation study.

According to our analysis, DDMS locations are strongly linked to genes involved in cancer development; and a third of DDMS locations are colocalized with active enhancers, gene regulatory elements (Fig. 2; Supplementary Table S2). These data suggest that DDMS locations play a role in functional cancer development.

In our study, DNA methylation is used for thyroid nodule diagnostics. Because DNA methylation is a very stable epigenetic mark, DNA methylation–based testing does not require special handling and can be performed on FFPE-preserved samples like cell blocks and even on FNA slides. Sometimes special sample handling is impossible when an unexpected incident, such as bleeding, occurs during a biopsy procedure. In this light, DDMS stands out in comparison with the RNA-based test, Afirma Gene Expression Classifier, that requires special tissue handling due to RNA instability.

Our approach is very different from standard cancer diagnostics. The standard diagnostics searches for cancer-associated alterations and classifies specimen as cancer negative in an absence of these characteristics. By contrast, our test defines both cancer-negative and cancer-positive states according to the nodule-specific signatures. Therefore, an absence of those epigenetic changes makes the thyroid tissues undiagnosable.

We found that 17% of thyroid nodules are missing DNA methylation changes specific for malignant or benign nodules. We observed that the absence of epigenetic signatures in thyroid nodules is frequently linked to lymphocytic thyroiditis. In fact, a group of thyroid nodules with thyroiditis had a very similar DNA methylation pattern to lymph node (Fig. 1). At the same time, 62% of adjacent thyroid tissues and 40% (2 out of 5) of malignant specimens with high levels (75% or more) of adjacent thyroid tissues were undiagnosable according to DDMS. These data suggest that the presence (contamination) of cells other than follicular epithelial cells from a nodule in a specimen may lead to the absence of thyroid nodule epigenetic signatures. Such contamination is frequently found in FNA biopsies of thyroid nodules. Contaminating cells may include white blood cells, skeletal muscle, blood vessels, or adjacent normal tissue. In order to overcome diagnostic difficulties associated with nonfollicular cell contamination, current molecular thyroid nodule diagnostics such as ThyroSeqV2 perform an “assessment of sample adequacy for molecular testing.” The assessment is completed by analyzing an expression activity of a gene specifically expressed in thyroid epithelium, KRT7 (1).

Regarding undiagnosable samples which are not linked to thyroiditis, at this point, it is very difficult to determine whether these specimens have their own epigenetic signature because benign and malignant groups with the undiagnosable specimens are fairly small. To solve this, a potential combination of DDMS and BRAFV600E genetic testing can reduce the number of undiagnosed patients. The BRAFV600E genetic screening is currently incorporated in Afirma Gene Expression Classifier and ThyroSeqv2 (13, 22). According to developing and testing cohort data, the incorporation of BRAFV600E in DDMS resulted in 86% of diagnosable specimens. (Supplementary Table S1). The positive side of the undiagnosable classification is that it does not incorrectly diagnose but rather gives room for other testing approaches. Alternatively, microdissection such as laser capture microdissection may allow for an enrichment of the sample for noncontaminating cells.

In our study, the DNA methylation patterns specific to benign nodules were used not only for a better characterization of cancer-specific alterations but also for the development of a benign-specific signature. In fact, if we had not used the benign signature, the malignant undiagnosable specimens would have been erroneously diagnosed as benign. Therefore, the introduction of the benign signature dramatically improves DDMS accuracy.

In contrast to current molecular thyroid nodule diagnostics, our test is based on differences between benign and malignant nodules. Our study suggests that benign nodules are distinct from adjacent thyroid tissues and malignant tissues. Therefore, not only adjacent tissues but the benign nodules themselves should be used for the development of thyroid nodule diagnostics. Furthermore, according to our study, some adjacent thyroid tissues have DNA methylation patterns (13N, 113N) that are very similar to the thyroid nodules to which these “normal tissues” are adjacent (Fig. 1). This fact can be explained by a field cancerization effect that is observed in many types of cancer (23–26). Field cancerization also explains classification of 113N thyroid tissues adjacent to a malignant nodule as malignant. Thus, adjacent thyroid tissues may carry cancer-associated alterations due to a field cancerization effect and have different DNA methylation compared with benign nodules. Our data shift the emphasis of the research paradigm to compare the benign thyroid nodule instead of any thyroid tissue versus the cancer for the development of thyroid cancer diagnostics.

Taken together, we have developed a new, inexpensive, high precision diagnostic test based on DNA methylation patterns of benign and malignant thyroid nodules. We plan on performing a multicenter clinical trial to determine the performance of this test in FNAs in the clinic.

J.H. Yim reports receiving speakers bureau honoraria from Veracyte. No potential conflicts of interest were disclosed by the other authors.

The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH.

Conception and design: J.H. Yim, Y. Fong, M.A. Hahn

Development of methodology: J.H. Yim, Y. Fong, M.A. Hahn

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.H. Yim, A.H. Choi, H. Qin, S. Chang, S.-W.T. Tong, P. Chu, B.-W. Kim, R. Lew, Y. Ibrahim, A. Baker, X. Wu, Y. Fong, M.A. Hahn

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.H. Yim, A.X. Li, D. Schmolze, V.A. Poroyko, X. Wu, Y. Fong, M.A. Hahn

Writing, review, and/or revision of the manuscript: J.H. Yim, A.H. Choi, A.X. Li, H. Qin, S. Chang, S.-W.T. Tong, D. Schmolze, R. Lew, Y. Ibrahim, X. Wu, G.P. Pfeifer, M.A. Hahn

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.H. Yim, S. Salvatierra, A. Baker, J. Wang, Y. Fong, M.A. Hahn

Study supervision: J.H. Yim, M.A. Hahn

Research reported in this publication was supported by the NCI of the NIH under award numbers R21CA223367 (to M.A. Hahn and J.H. Yim) and P30CA033572 to City of Hope. Research reported in this publication included work performed in the Pathology Research Services Core – Biorepository and the City of Hope Integrative Genomics Core.

The authors thank Karen Miller, Michael Lewallen, and Maria De Lao for their support and outstanding work with patient specimens.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.