Shrunken centroid analysis with a 29-gene subset. (A) Class prediction analysis using 29 genes derived from the 22 215 genes. The class prediction analysis was able to accurately classify 19 of the 24 patients with AL and 23 of the 28 patients with MM with an observed accuracy of 90% and a cross-validated accuracy of 81%. The y-axis shows the threshold value with samples closer to 1 having the highest probability of being an AL or MM sample, respectively. The x-axis denotes each of the 24 AL and 28 MM samples. Circled symbols indicate misclassified patients. (B) Identification of the 12 genes used to classify patients in the 2 groups. The shrunken differences for the 12 genes used for class prediction are shown. The size of the bars indicates relative distance from the centroid, with the larger bars having more significance in predicting the class. The set of 12 genes that could classify patients with AL or patients with MM with 92% observed accuracy is listed by their Affymetrix probe identification numbers. The probe sets included (in order) TNFRSF7, SDF-1 (CXCL12), JUN, PSMA2, DEFA1, NDUFA4, PGK1, and TXN.