Ion quality remains competitive with the other learners (close to 90 for accuracy, recall and precision). For ALL-AML Leukemia, the top three performers are SVM, Bayes and MDR. All have the same FPR of 4 . SVM and Bayes have the highest recall, accuracy, precision and the lowest FNR. However, MDR has the best AUC of 99.02 , and the second best values for the other measures, with precision lower by only 0.04 , accuracy lower by about 1 , and differences in FNR and recall of about 2 . This is similar to the results obtained in prostate cancer. However, here SVM and Bayes give highly competitive results in all measures except AUC. On the other hand, MDR has reasonably high precision, accuracy and recall (97.87 , 97.22 PubMed ID:https://www.ncbi.nlm.nih.gov/pubmed/26266977 , and 97.87 , respectively). MDR performs very well on the classification ranking of ALL. MDR’s performance on lung cancer is similar to its performance on lymphoma cancer. For lung cancer, SVM, Bayes and MDR consistently performed better than the rest. The AUC and ACC obtained by MDR were 0.39 and 2.76 lower than those of the top performer, but it still has over 90 on almost all measures, with a low 2 FPR. Unfortunately, the 9.68 for FNR obtained by MDR is rather high, but still better than the 12.9 and 35.48 ofPage 3 of(page number not for citation purposes)BMC Genomics 2008, 9(Suppl 2):Shttp://www.biomedcentral.com/1471-2164/9/S2/S0.True Positive Rate0.0.4 C4.5 MDR SVM NaiveBayes ZeroR 3NN 0 0.2 0.4 0.6 0.8 1 False Positive Rate0.Figure 1 ROC curves on prostate cancer expressions ROC curves on prostate cancer expressionsC4.5 and 3NN, respectively. In the lymphoma cancer analysis, Bayes and SVM are top performers, followed by MDR. The AUC for MDR is 93.61 , which is 4.29 lower than that of Bayes and SVM. Compared to MDR, 3NN has 4.17 lower FPR (type I error) but 25.8 higher FNR (type II error), which is usually a more critical measure for tumor prediction. The ACC and AUC obtained by C4.5 and 3NN are not competitive to those obtained by MDR. All the machine learners had poor results for the breast cancer data. Furthermore, the results obtained from different measures are not consistent. For example, Bayes has 100 precision but very low AUC and ACC of only about 50?5 , and an extremely high Type II error of about 95 . This suggests that the data are probably very skewed. The results obtained from SVM are the best with AUC, ACC and recall all close to 70 . However, its precision is only 65 , and it has very high type I II errors of about 30 . MDR does not perform well. Nevertheless, the majority of its results rank second best among all the learners. We now discuss the results obtained in the SP600125 web ALLsubtype classification as shown in Table 3. For the classification of subtypes E2A-PBX1 and T-ALL, MDR has perfect averages in all measures. For the subtype BCR-ABL, although MDR has the highest AUC, 88.48 , and high accuracy, 92.66 , (3.98 lower than the accuracy obtained by the best learner, SVM), it has low precision, low recall, and high FNR, about 80 . The results obtained by the other learners also show similar inconsistency across different measures. Thus, for BCR-ABL classification, all learners are likely to perform well but with high variances. For the rest of the ALL subtypes: Hyperdiploid > 50 chromosomes (HYP > 50), MALL and TEL-AML1, the results are similar in that the top three performers are SVM, 3NN, and C4.5 on almost all measures except AUC. However, MDR gives reasonably high accuracy, ranging from 80.