Development and validation of mRNA expression-based classifiers to predict low-risk thyroid tumors.

Background: Molecular variants and fusions in thyroid nodules can provide prognostic information at a population level. However, thyroid cancers harboring the same molecular alterations may exhibit diverse clinical behavior. Leveraging exome-enriched gene expression analysis may overcome the limitations seen in models based on a small number of point mutations or fusions. Here, we developed and validated mRNA-based classifiers with high negative predictive values to preoperatively rule out thyroid tumor invasion and lymph node metastases. Materials and methods: In this retrospective cohort study, histopathology reports from the Afirma Genomic Sequencing Classifier (GSC) algorithm training and consecutive thyroid cancer patients with Bethesda III-VI thyroid nodules in clinical practice (total 697 and ~50%, respectively) were scored for invasion and metastases. mRNA expression-based classifiers were developed utilizing literature-derived signatures as well as differentially expressed genes between samples with or without clinically significant invasion/metastases as the basic building blocks. Machine learning algorithms were employed to develop the final candidate classifiers. The final locked classifiers were validated on a retrospective cohort of 259 patients with Afirma testing who had thyroid surgery and had invasion and metastasis scores assigned based on histopathology while blinded to the classifier results. Results: A total of 697 (88% female) patient Afirma samples and scored histology reports were used for classifier development. In development, patients had a median age of 51 years. Ten percent of samples were assigned a high risk for invasion label, and 11.3% were assigned a high risk for lymph node metastasis (LNM) label. A low-risk invasion classifier result was assigned to 41.3% of the cohort with a negative predictive value (NPV) of 97.6%, and a low-risk LNM classifier result was assigned to 49.8% of the cohort with an NPV of 98.6%. In the validation cohort, made up of 75% women with a median age of 53 years, 51% of the samples were ruled out for high risk for invasion label with a 99% [95-100] NPV, and 53% were ruled out for high risk for LNM label with 100% [97-100] NPV. Discussion: Gene expression-based classifiers that confidently, preoperatively rule out thyroid tumor invasion and lymph node metastasis may help personalize the surgical approach for individuals, reducing overtreatment, surgical complications, and postoperative hypothyroidism.