ABSTRACT: The gene expression patterns of favorable histology Wilms tumors (FHWT) that relapsed were compared with those that did not relapse using oligonucleotide arrays; Description: 250 FHWT of all stages enriched for relapses treated on National Wilms Tumor Study 5 passed quality parameters and were suitable for analysis using oligonucleotide arrays. Relapse risk stratification utilized Support Vector Machine; two and ten fold cross-validation was applied. The number of genes associated with relapse was less than that predicted by chance alone for 106 patients (32 relapses) with stages I and II FHWT and no further analyses were performed. This number was greater than expected by chance for 76 local stage III patients. Cross validation including an additional 68 local stage III patients (total 144 patients, 53 relapses) demonstrated that classifiers for relapse composed of 50 genes were associated with a median sensitivity of 47%, specificity 70%, and total error rate of 38%. Analysis of genes differentially expressed in relapse patients revealed apoptosis, Wnt signaling, IGF pathway, and epigenetic modification to be mechanisms important in relapse. Potential therapeutic targets include FRAP/MTOR and CD40. Experiment Overall Design: 144 stage 3 FHWT, with fifty-three relapses (cases) and ninety-one non-relapses (controls) with a minimum of three years follow-up, included all relapses and a 30% random selection of non-relapses for which frozen tumor tissue was available who passed all quality control parameters. The NWTS-5 protocol was approved by the review boards of institutions that registered patients. Histological diagnosis and local stage were confirmed by central review. Experiment Overall Design: Quality control steps taken: Experiment Overall Design: 1. Samples were snap frozen immediately following surgery and were mailed on dry ice to the Tumor Bank and retained at -80C. Experiment Overall Design: 2. Frozen sections were evaluated histologically and tumors with less than 80% viable tumor cellularity were excluded. Experiment Overall Design: 3. Array images were assessed by eye to confirm scanner alignment and the absence of significant bubbles or scratches. Experiment Overall Design: 4. Samples for which the 3'/5' ratios for GAPDH were greater than 3.2 were excluded. Experiment Overall Design: 5. The BioB spike controls were confirmed as present; BioC, BioD and cre were confirmed as increasing intensity. Experiment Overall Design: 6. When scaled to a target intensity of 2500, scaling factors were between 12 and 53; background levels were 34 â 115: Q values were 1.3 â 3.7 and mean intensities were within acceptable limits. Experiment Overall Design: 7. The range of percent present calls was from 38% to 52%. Experiment Overall Design: 8. Verification of gene expression was performed utilizing quantitative RT-PCR for five genes. Â Â Experiment Overall Design: Statistical Analysis: Positional-dependent-nearest-neighbor model (PDNN) software was used to translate the scanned images into expression analysis files and to normalize the data across all arrays (http://odin.mdacc.tmc.edu/~zhangli/PerfectMatch/). Genes with maximum expression less than a log scale of 6 across all tumors and Affymetrix control genes were excluded, resulting in 20,931 probe sets for analysis. Support Vector Machine (SVM) as developed by Chang and Lin (http://www.csie.ntu.edu.tw/~cjlin/libsvm) and implemented in an R software package, e1071 was chosen for relapse risk stratification using the p-value of the t-test comparison between case and control to select the genes. Using all 144 tumors, 109 genes were identified with p-value <0.001 and are provided in the data table labelled "t-test comparison between case and control" (ie, Supplemental table 2, in related publication)." Two and ten fold cross validations were utilized to investigate the ability of classifiers established in randomly selected training set to predict relapse in an independent test set comprised of the remaining tumors. For two-fold cross validation, the dataset was randomly divided 500 times into training and corresponding test sets of equal size, each including half the patients who relapsed. A classifier for relapse was identified for each training set and used to assign tumors in the corresponding test set to low and high risk categories. The training and test sets were then swapped. The number of top K genes in each classifier evaluated ranged from 1-150. Therefore, a total of 150,000 different classifiers were developed, one for each value of K from 1-150, for each of the 1,000 (500*2) training sets. For ten-fold cross validation the dataset was randomly divided 500 times into ten groups of approximately equal size. Each group included approximately the same number of relapses. For each such group, a classifier was built with the remaining 9/10 of the samples and then used to categorize tumors in the group as low or high risk; the process was repeated until all tumor samples were categorized as low or high risk. For all the cross validation procedures, to avoid gene-selection bias, classifiers were completely rebuilt in each cross validation iteration.
