Lung squamous cell carcinoma LUSC

An elastic-net logistic regression was trained on the 300 top-ranked candidates to tell tumor from normal tissue. Its held-out AUCHeld-out AUCHow well an elastic-net model separates tumor from normal tissue, measured with nested cross-validation on samples the model did not train on. It describes tissue, not the accuracy of a blood test. is > 0.999, from nested five-fold cross-validation.

Before training, the TCGA-versus-GTEx offset was removed from the expression matrix with the tumor/normal contrast protected, as in the tissue comparison. That correction is fitted once on all samples, so the held-out AUC is not fully independent of it.

This number describes tumor and normal tissue, where separation is expected to be near perfect. It is not the accuracy of a blood test: that has to be measured in plasma from patients and controls.

For the 150 top-ranked candidates, a Cox model relates tumor expression, as a continuous value, to overall survival in 491 patients (211 events). No high/low cutpoint is searched for, since optimized cutpoints inflate false positives.

A hazard ratioHazard ratio per SDFrom a Cox model with the gene’s tumor expression as a continuous variable: the change in hazard for each one-standard-deviation increase. Above 1, higher expression goes with a shorter time to the event; below 1, with a longer one. above 1 means higher expression goes with a shorter overall survival. These are associations in tissue, not evidence that a blood level predicts outcome.