Download test set = training set Internal validation

A
B
Internal validation (1)
External validation
Dataset = 105 late biopsies
= training set
105 late biopsies
= training set
= test set
10-fold cross-validation:
Validation set 1:
73 early biopsies from the U
of Alberta and U of Illinois
Validation set 2:
48 late biopsies from the
University of Minnesota
C
Internal validation (2)
Dataset = 173 early and late
biopsies
…
…
Repeat
100 times
10-fold cross-validation:
…
…
Repeat
100 times
Supplementary Figure 1. Validation procedure. A) For internal validation, 10-fold cross-validation is used. In each of the ten folds, the
classifier, including the gene selection phase, is rebuilt from scratch. A single estimate of the molecular risk score is made for each of the 105
samples, based on the predictons from the left-out samples. This entire procedure is repeated 100 times. The final reported risk score for each
biopsy (i.e. those in Fig. 2) is the average risk score over all 100 iterations. B) For external validation, the classifier is built from the full set of
105 late biopsies, then used to predict either early biopsies (validation set 1) or late biopsies from a different centre (validation set 2). C)
Classifier built as in A, but using all 173 patients (early + late) from the U of Alberta and U of Illinois.
Suppl. Fig 1
High
High risk
risk
Low
Low risk
risk
Supplementary Figure 2. Boxplots of molecular risk scores in individual biopsies over all 100 iterations of the classifier. See Figure
2 for details. Each biopsy’s risk score is estimated 100 times in separate iterations of a 10-fold cross-validation procedure. The boxes
represent the interquartile range of the 100 estimates. The whiskers extend to the most extreme data point which is no more than 1.5 times the
interquartile range from each box.
AUC
0.83
0.79
0.80
0.75
Supplementary Figure 3. ROC curves for single gene classifiers. The main classifier result based on the supervised PCA classifier is compared to
those from single-gene classifiers. Ranking of genes in each iteration of the training sets was based on the p-value from a Cox regression.