Expression based biomarkers and models to classify early and late stage samples of Papillary Thyroid Carcinoma

Abstract

In this study, we describe the key transcripts and machine learning models developed for classifying the early and late stage samples of Papillary Thyroid Cancer (PTC), using transcripts expression data from The Cancer Genome Atlas (TCGA). First, we rank all the transcripts on the basis of area under receiver operating characteristic curve, (AUROC) value to discriminate the early and late stage, based on an expression threshold. With the expression of a single transcript DCN, we can classify the stage samples with a 68.5% accuracy and AUROC of 0.66. Then we implemented various combination of multiple gene panels, selected using various gold standard feature selection techniques. The model based on the expression of 36 multiple transcripts (protein coding and non-coding) selected using SVC-L1 achieves the maximum accuracy of 74.51% with AUROC of 0.75 on independent validation dataset with balanced sensitivity and specificity. Further, these signatures also performed well on external microarray data obtained from GEO, predicting nearly 70% (12 samples out of 17 samples) early stage samples correctly. Further, multiclass model, classifying the normal, early and late stage samples achieves the accuracy of 75.43% with AUROC of 0.80 on independent validation dataset. With correlation analysis, we found that transcripts with maximum change in correlation of their expression in both the stages are significantly enriched in neuroactive ligand receptor interaction pathway. We also propose a panel of five protein coding transcripts, which on the basis of their expression, can segregate cancer and normal samples with 97.32% accuracy and AUROC of 0.99 on independent validation dataset. All the models and dataset used in this study are available from the web server CancerTSP (http://webs.iiitd.edu.in/raghava/cancertsp/).