..; (3) radial basis kernel: K(x, y) = expx-y; (4) Sigmoid kernel: K(x, y) = tanh [b(x•y)+c], where b, c and σ are parameters. Among these four types of kernel
function, radial basis kernel showed best performance according to the results from similar studies [34, 35]. The correct choice of kernel parameters is crucial for obtaining good results, so an extensive search must be conducted FK228 supplier on the parameter space before results can be trusted. Here we adopted radial basis kernel function and 5-fold cross-validation in the training set to search the best parameters for SVM-based classification in the test set. Figure 1 Classification via SVM (linear separable case). Evaluation of model performance Classification accuracy and the standard deviations of our proposed method (with prior knowledge) were compared with the original one (no prior knowledge) in the training set and test set. The framework of the above mentioned procedures is shown in Figure 2. Figure 2 Framework of our proposed method. Statistical analysis All the statistical analyses were conducted using R statistical software version 2.80 (R foundation for Statistical Computer, Vienna, Austria). Results Genes selected by PAM The number of genes selected by PAM method varied from 4 to 12 with an E7080 mw average 7.81, and the standard deviation 2.21. The combination of genes selected by PAM is shown ID-8 in Table 1. Among them,
CEACAM6, calretinin, VAC-β and TACSTD1 appeared in the results all the time. Table 1 Gene lists selected by Prediction Analysis for Microarrays Gene name GenBank access No. Location at HG_U95Av2 ERBB3 M34309 1585_at CD24 L33930 266_s_at TACSTD2 J04152 291_s_at UPK1B AB015234 32382_at HIST1H2BD M60751 38576_at TITF-1 U43203 33754_at CLDN3 AB000714 33904_at CEACAM6 M18728 36105_at PTGIS D83402 36533_at SFTPB J02761 37004_at caltrtinin X56667 37157_at VAC-β
X16662 37954_at claudin-7 AJ011497 38482_at AGR2 AF038451 38827_at TACSTD1 M93036 575_s_at Gene selection via prior biological knowledge After reviewed the full text of literature, twenty-three lung adenocarcinoma-related genes were selected. Then, Table 2 lists the eight significant genes that passed the multiple testing procedure in the training set provided by Gordon et al. The details of these genes are shown in Table 2. Table 2 Genes as prior biological knowledge Gene name GenBank access No. Location at HG_U95Av2 CXCL1 J03561 408_at IL-18 U90434 1165_at AKAP12 X97335 37680_at KLF6 U51869 37026_at AXL M76125 38433_at MMP-12 L23808 1482_g_at PKP3 Z98265 41359_at CYP2A13 U22028 1553_r_at Evaluation of model performance Our proposed method performed better after incorporating prior knowledge (Figure 3). Accuracy of the Apoptosis inhibitor modified method improved from 98.86% to 100% in training set and from 98.51% to 99.06% in test set. The standard deviation of the modified method decreased from 0.