Fig. 5

Permutation validation and enrichment analysis. In permutation validation, ten thousand protein combinations were generated randomly. The model with each combination was trained on the training set using SVM. The ability of each combination to separate the testing set into good- and poor-prognosis groups was evaluated using the log-rank test. Histograms of the χ² value from the log-rank test on ADC testing set (a) and SCC testing set (b) were illustrated. The x axis indicates the χ² value. A larger χ² value indicates a lower P value and a more statistically significant ability to separate the testing set. The y axis shows the frequency and higher values indicate a larger fraction of the population. The performance of the ADC/SCC signature is marked with a red arrow. In enrichment analysis, ten thousand signatures were identified on 10,000 randomly partitioned training sets using random forest algorithm. For each protein, the fraction of the signatures containing the protein (i.e. percentage of subsets) in ADC (c) and SCC (d) patients from BJ cohort was calculated. A zoom-in on the 15 most enriched proteins is also shown. Each column corresponds to a protein, the signature proteins are denoted in red