"... In PAGE 13: ... We ranked them according to their mean expression levels in our experiments. The top 15 highly expressed housekeeping genes are listed in Table2 . It is evident that several of these highly expressed genes (9 out of 15) are ribosomal protein coding genes that carry out important cellular functions.... ..."

"... In PAGE 9: ... Gene expression profiles. The remaining eleven genes fell into two categories: novel genes for which a function has been predicted by sequence similarity to known genes, and novel genes with no known function and no or limited similarity to known genes ( Table2 ). We examined the UniGene clusters associated with these novel genes.... In PAGE 9: ... As EST sets made from hypothalamus RNA are represented in the mouse EST databases, we expected to see representation of hypothalamic cDNAs within the Unigene clusters corresponding to these genes. Indeed, all eleven genes had corresponding ESTs derived from whole brain, with some specifically including hypothalamus (4) or diencephalon (2) ( Table2 ). The information in the UniGene clusters did not point to these genes being particularly hypothalamus-specific, in contrast to the Unigene cluster for arginine vasopressin, for example, in which twelve of the nineteen ESTs were derived from the hypothalamus.... ..."

"... In PAGE 2: ... We compiled from human stud- ies [7-9,12,18], transcriptional differences between the preterm myometrium (PTNIL), term myometrium not in labor (TNIL) and term myometrium in labor (TIL). Methods standing, we used combinations of searches through the literature referenced in public databases ( Table1 ) and the Onto-Tools software developed by the Draghici apos;s group at Wayne State University (Detroit, MI, USA) [19]. The Onto-Express module helps to recognize functional pro- files (using gene ontology terms) for the categories: bio- chemical function; biological process; cellular role; cellular component; molecular function, and chromo- some location [20].... ..."

"... In PAGE 22: ... Similarly, we performed a thorough optimization of the KNN classifier over all possible numbers of neighbors K ranging from 1 to the total number of instances in the training dataset based on cross-validation error. Datasets and data preparatory steps The datasets used in this work are described in Table2 . In addition to nine multicategory datasets which were most of the multicategory cancer diagnosis datasets in humans found in the public domain at the time the present study was initiated, two binary datasets (i.... In PAGE 23: ...tering for a minimal lev ecutive image analysis performed by DeArray Software and fil el of expression [Khan2001]. The genes or oligonucleotides with absent calls in all samples were excluded from the analysis to reduce the amount of noise in the datasets ([Lu2002] and [Wouters2003]), and if this was the case, the number of genes is listed in bold in Table2 . While setting up datasets for experiments, we took advantage of all available documentation in order to increase the number of categories or diagnoses for the outcome variable.... In PAGE 31: ...as employed, we obtained 76.60% in 14_Tumors and 92% in Prostate_Tumor. The difference can be explained by the difficulties of the classification problems - 14_Tumors is much harder (it has 26 classes with prior of the most frequent class 9.7%, see Table2 ) than Prostate_Tumor (it is a binary problem with prior 51%, see Table 2). According to Table 3, in 8 out of 11 datasets, MC-SVMs perform cancer diagnoses with accuracies gt; 90%.... ..."

"... In PAGE 5: ...5. Datasets and data preparatory steps The datasets used in this work are described in Table1 . In addition to nine multicategory datasets which were most of the multicategory cancer diagnosis datasets in humans found in the public domain at the time the present study was initiated, two binary datasets (i.... ..."

"... In PAGE 7: ... The average rank of the left-out PFRs gives an indication of how similar the PFRs are to each other and how distinct from background. To explore the robustness of the PFR-Sampler program, we used inputs of 10, 12, 14, 17 and 19 blastodermally expressed genes( Table1 ). Ineachoftheseinputs, thecoresetof10genes was the same, drawn from a previous TF-clustering based CRM prediction study (Berman et al.... In PAGE 9: ... Since 80% of the left-out PFRs were at rank 2/1000 or better and 88% at rank 4/1000 or better, we elected to take as a threshold the midway point of the score ofthe3rdrankedPFRoutof1000. Queryingthecompletegen- omeandselectingthosewithscoresabovethesetpointyielded 207 PFRs, of which 24 are the input PFRs (see Supplemental Table1 , http://arep.... ..."