Construction and validation of a signature with ferroptosis-related genes

We compared gene expression levels between tumor and adjacent normal tissues to identify the DEGs using the ‘limma’ R package [¹⁶]. The criterion of DEGs was a false discovery rate (FDR) < 0.05. Cox analysis assisted in examining the ability of ferroptosis-related genes to predict overall survival (OS). Benjamini-Hochberg adjusted p-values were used to decrease FDR. The ‘glmnet’ R package served for a LASSO Cox regression – a powerful technique for variable selection and regularization – to analyze if DEGs could predict the OS and the status of the PTC patients [¹⁷]. The optimal value of the penalty parameter (λ), which corresponds to the minimum of the partial likelihood deviance, was identified by ten-fold cross-validation. We calculated the risk score using the following formula: risk score = esum (the normalized expression level regarding each gene× its regression coefficient) [¹¹]. We then used the median risk score as the standard for classifying patients into group with high risk and group with low risk. We depicted the gene distribution in two groups by performing PCA and t-SNE using the ‘stats’ and the ‘Rtsne’ R package [¹⁸]. The ‘surv_cutpoint’ function of the ‘survminer’ R package served for the survival analysis to identify the optimal cutoff values for gene expression [¹⁹]. We then used the ‘survivalROC’ R package to perform time-dependent ROC analysis to estimate if the gene signature has predictive ability [²⁰]. We then used the univariate and multivariate Cox regression analyses for determining if the risk score could independently predict patients’ prognosis in terms of the OS.