Fig. 4.

(a) Training samples of the ALL-AML data. The polygons indicate the subsets obtained by SPACC, where the AML samples are divided into two subclasses. (b) Testing samples. SPACC achieves 91.4% classification accuracy, which is higher than the nearest neighbor method. The incorrectly classified testing samples are highlighted. SPACC offers more information than other supervised classifiers. Other supervised methods only classify a testing sample to either the ALL class or the AML class. On the other hand, SPACC is able to classify testing samples into three categories (ALL, AML-1, and AML-2), which may correspond to biologically meaningful subtypes.