A gene expression signature that distinguishes malignant from benign thyroid nodules.

Abstract

e21011 Background: Thyroid nodules can be detected in as high as 67% of the population. Distinguishing thyroid cancers from benign lesions is crucial for determining an appropriate treatment plan. For years a gene expression signature for discriminating malignant from benign thyroid nodules has been sought by clinicians. In this study, multivariate bioinformatics tools were used to generate a qPCR based gene expression signature for determining malignancy in thyroid nodules. Methods: Multiple mathematical models, such as Random Forest, Support Vector Machine (SVM), and Nearest Shrunken Centroid (NSC), were used to analyze published microarray data sets and select 366 putative classifier (biomarker) mRNA targets. The selected 366 genes were further evaluated for their expression pattern by real-time PCR using a panel of 49 pathology assessed thyroid nodule samples (fresh frozen, 23 malignant and 26 benign). Results: Using the qPCR data set, Random Forest was compared with SVM and NSC classifier methods and was found to be more successful in finding genes with better discriminative powers. A Random Forest method identified a panel of 7 genes together with 5 reference genes as a gene expression signature for thyroid malignancy, which led to the development of a companion classifying algorithm to provide a probability score to assess malignancy of thyroid nodules. In our limited sample set, this signature was shown to distinguish malignant and benign thyroid nodules with 92% accuracy and 100% specificity. Conclusions: Our results suggest that a combination of multiple bioinformatics analysis tools is the proper approach for biomarker candidate selection from high-throughput gene expression data. As demonstrated here, panel of 12 genes and a companion classification algorithm has the potential to successfully discriminate malignant thyroid nodule with high accuracy and specificity. This panel of twelve genes is for molecular biology applications only.