Construction of a Prognostic ATG Signature

The absolute median difference (MAD) was used to select the ATGs (Guinney et al., 2015). MAD is a robust statistic of statistical divergence that is more adaptable to outliers in a dataset than the standard deviation. To construct a prognostic ATG signature, we focused on 655 ATGs from eight gene sets identified via MSigDB (version 6.2) (Subramanian et al., 2005; Liberzon et al., 2011; Liberzon et al., 2015) with the keyword “autophagy.” Only 617 genes measured on all platforms involved in this study with high variation (MAD >0.5) were selected. After 1,000 times random Cox univariate regressions, genes with repeated significance, which indicated a strong relationship between ATGs and patients’ disease-free survival (DFS), were selected as candidates for the signature. A Cox proportional hazard regression model on CRC samples together with the least absolute shrinkage and selection operator (LASSO) was applied to minimize the risk of overfitting as well as to generate a risk model.

Patients were divided into low and high autophagy risk groups in accordance with the optimal ATG signature cutoff, which was defined by the time-dependent receiver operating characteristic (ROC) curve analysis at 5 years of DFS in the training cohort. The ATG signature score with the largest Youden’s index in the ROC curve was deemed as the cutoff value.