Development and validation of an RNA sequencing–based classifier for medullary thyroid carcinoma on thyroid FNA

Medullary thyroid carcinoma (MTC) is an uncommon malignancy of C-cell origin that accounts for <3% of all thyroid cancers.^1,2 However, the prognosis for MTC, which is derived from parafollicular C cells, has generally been worse than the prognosis for differentiated thyrocyte-derived carcinomas, and it thus accounts for a disproportionate number of deaths from thyroid carcinoma. The clinical course of MTC has been found to be largely stage-dependent. Ten-year survival for thyroid-confined MTC is ~90%– 100%, whereas for MTC with nodal and distant spread, 10-year survival is ~32% and ~ 17%, respectively.³ It is well understood that earlier detection of MTC has led to improved outcomes and overall patient survival.^4–8 Better imaging, biochemical testing, and advanced tissue-based diagnostic pathological protocols, all of which rely on sufficient access for patients to an equipped health care system, are all essential to this. As pathologists, though, we must have the tools to consistently make the best diagnosis when slides arrive at our desks. Fine-needle aspiration cytology (FNAC) has limited sensitivity for the diagnosis of MTC, and ancillary testing may be essential for improving our diagnostic accuracy.^9,10 Molecular tests increasingly are being used in the diagnosis of thyroid lesions, and their validation for MTC consequently serves as a significant point of interest. This is the focus of a recent article by Randolph et al.¹¹ published in Thyroid, which serves as the basis for this commentary.

MTC represents a diagnostic cytological conundrum for two main reasons: (1) it is relatively rare in comparison with the far more frequent follicular thyroid lesions, and (2) it shows a diversity of nuclear and architectural cytomorphology that may in some cases overlap with these more frequent follicular thyroid lesions.^12,13 MTC with more extensive epithelioid to plasmacytoid morphology may mimic oncocytes (Hürthle cells). In fact, students of pathology history know that the original cells Dr Karl Hürthle described in the thyroid were parafollicular C cells, and the use of the term Hürthle cell to describe thyroid oncocytes is a misnomer. MTC cases with an extensive oncocytic appearance risk misclassification as a hyperplastic nodule with oncocytic metaplasia or an oncocytic neoplasm.¹² The cells of MTC can also be arranged in tight clusters mimicking the microfollicular architecture of a follicular thyroid neoplasm.¹⁴ MTC cells may demonstrate intranuclear pseudoinclusions, and the amyloid on cytology smears can simulate the thick “bubble gum” colloid of papillary thyroid carcinoma.^13,15 Finally, the differential diagnosis for cases with an extensive spindle cell component might include anaplastic thyroid carcinoma or even mesenchymal tumors.¹⁶

In light of these challenges, it is unsurprising that thyroid FNAC is poorly sensitive for MTC. In the largest meta-analysis of the subject to date, the diagnostic rate of thyroid FNAC for MTC, when “suspicious for malignancy–MTC” and “positive for malignancy–MTC” were both considered to be positive for MTC, was a disheartening 56% (360 of 641). Of the misclassified fine-needle aspiration (FNA) specimens, roughly one third were misdiagnosed as other malignancies, one third were called indeterminate, and one third were considered nondiagnostic or benign.¹⁰ Although there is little that we may be able to do for the latter category beyond improving our recognition of these tumors, for FNA biopsies considered to be indeterminate, neoplastic, or malignant, ancillary testing could be a helpful adjuvant for arriving at the definitive diagnosis. The importance of this for indeterminate cases goes without saying; even for malignant ones, however, arriving at the correct diagnosis is crucial to appropriate management. The standard treatment for MTC involves the assessment of preoperative serum calcitonin and carcinoembryonic antigen levels followed by total thyroidectomy with central neck dissection with or without lateral cervical lymph node dissection depending on ultrasound findings and calcitonin levels.¹⁷

There are additional benefits to the molecular testing of thyroid FNAC beyond improved diagnosis. In the case of MTC, the tumor mutational profile may guide the need for genetic screening for multiple endocrine neoplasia type 2 syndrome. Although these tests would not distinguish between sporadic and germline alterations, the identification of a RET mutation, present in roughly 50% of MTCs, would at least suggest the need for genetic screening in the proper clinical setting.¹⁸ Similarly, the identification of non-RET mutations exclusive to sporadic MTC, including HRAS and KRAS, would obviate the need for screening, but the discovery of these mutations, which are more frequent in follicular and oncocytic neoplasias of the thyroid, may not lend additional clarity in the preoperative diagnosis of MTC.¹⁹ Importantly, in cases of advanced disease, molecular results are certainly required to guide adjuvant therapy and, in some cases, neoadjuvant therapy, as small molecular RET inhibitors are now approved by the Food and Drug Administration for the treatment of thyroid cancers with RET alterations.^20–22

The article by Randolph et al.¹¹ describes the development and validation of an RNA sequencing–based MTC classifier by Afirma for use on thyroid FNAC. In addition to detecting gene mutations and copy number alterations, the RNA sequencing–based genomic sequencing classifier provides for high-fidelity gene expression profiling, which allows for the detection of parathyroid tissue and MTC.²³ The training cohort for the MTC classifier consisted of 483 thyroid FNA specimens with surgical follow-up, including 21 MTC cases and 462 non-MTC cases; the validation cohort consisted of a distinct set of 211 thyroid FNA cases, including 21 MTC cases and 190 non-MTC cases. Cases were sourced from 37 different institutions in order to achieve a sufficient number of MTC FNAC specimens; as a result, this represents the largest independent validation cohort for any available molecular diagnostic tests for MTC.

The candidate genes were established by selecting those differentially expressed in MTC and non-MTC from 483 FNAC specimens and 97 surgical pathology specimens. Eight candidate classifier models of genes, referred to as “feature sets,” were then devised with unique statistical models to define differential expression. These feature sets were subsequently used on the training cohort, which consisted of the 483 FNAC specimens, to generate binary “positive for MTC” or “negative for MTC” results. In the end, all eight competing feature sets performed with perfect sensitivity and specificity for distinguishing MTC; several criteria were then used to choose the most robust model.

So how did the final model perform when it was blindly tested on the validation cohort of 211 FNAC cases? As it had with the training cohort, the model showed 100% sensitivity (21 of 21 MTCs correctly labeled as positive; 95% CI, 83.9%–100%) and 100% specificity (190 of 190 non-MTCs correctly labeled as negative; 95% CI, 98.1%–100%). Importantly, the non-MTC cases in the validation cohort included a number of other thyroid pathologies ranging from nonneoplastic conditions such as nodular follicular disease and chronic lymphocytic thyroiditis to classic differential diagnoses for MTC such as oncocytic neoplasm, papillary thyroid carcinoma, and even hyalinizing trabecular tumor. The Afirma RNA sequencing–based MTC classifier described in this text has also been performed on thyroid FNAC specimens from various independent institutions, with more than 2100 thyroid nodules reported in the literature. There have yet to be any false negatives or false positives.^24–36

In summary, the article by Randolph et al.¹¹ provides insight into the development and validation of the Afirma RNA sequencing–based classifier for distinguishing MTC from non-MTC samples on thyroid FNAC. It presents a fascinating look into the complex process of machine learning and statistical models used to create such a test. Most importantly, however, with 21 MTC cases and 190 non-MTC cases correctly distinguished in the validation cohort, it provides further proof of the test’s essentially perfect performance to date. Given the significance of a missed preoperative MTC diagnosis, cytopathologists should take comfort in knowing that they are equipped with ancillary testing, Afirma being simply one example discussed here, that even the great mimicker cannot evade.