Nearest shrunken centroids classification. A Illustration of training the pamr prediction profile on osteosarcoma biopsies. At thresholds of 4.9-5.1, the misclassification error rate was minimal. B True versus predicted values from the nearest shrunken centroid fit. C Probabilities of each biopsy to belong to any of the three histological subtypes. Samples are separated (dotted lines) based on their true subtypes. Cross-validated probabilities for each histological subtype are shown on the y-axis, so that for every sample three open dots are present (blue, red, and green dots for osteo-, chondro-, and fibroblastic osteosarcoma, respectively). A sample is classified into a specific subtype if the probability to belong to that specific subtype is higher than the probabilities to belong to the other subtypes. D The FDR plotted against different thresholds of the prediction profile. At a threshold of 5.0, 24 genes are included in the prediction profile. These 24 genes have a FDR < 5%.