D from the TCGA public domain. For the UVA-51 cohort, we

D from the TCGA public domain. For the UVA-51 cohort, we obtained and used the archived patient samples and deidentified clinical data which were consented for general research purpose and approved by the Institutional Review Board (PRC# 455-07) at the University of Virginia; its full description has been published elsewhere [8].Statistical AnalysisMultivariate models for predicting patient therapeutic responses to three chemotherapy drugs, paclitaxel, cyclophosphamide, and topotecan, were derived by integrating in vitro drug sensitivity data for the NCI-60 cell lines and clinical outcome information from EOC patients after standard chemotherapy. The schematic procedures for our model training and validation are summarized in Figure 1. First, initial gene expression biomarkers highly associated with in vitro drug sensitivity were identified from the NCI-60 microarray data by correlating each drug’s GI50 values for the NCI-60 with their genomic expression data for cyclophosphamide and topotecan treatment and by identifying differentially expressed biomarkers between sensitive and resistant cell lines of the NCI-60 to paclitaxel. These chemosensitivity biomarkers were then triaged based on the COXEN coefficient, which represents the degree of concordance of expression regulation between the NCI-60 cell lines and a general EOC patient population prior to standard chemotherapy [16]. In brief, derivation of the COXEN coefficient is based on a so-called “AC220 dose correlation of correlations,” which first calculates the expression correlations within each set for the identical set of genes and then evaluates gene-by-gene correlation between the correlation matrices of the two sets. This kind of second-order correlation has proven useful by us and others for investigating various gene networks to identify concordant data sets [17?9]. More detailed description of the COXEN algorithm can be found elsewhere [7,10]. The above biomarkers were further screened with ovarian cancer patient data: the Bonome-185 set for paclitaxel and cyclophosphamide and the TCGA-UW set for topotecan. A subset of each drug’s biomarkers significantly associated with patient survival was identified by a Cox regression survival analysis. Therefore, these final biomarkers were the genes significantly associated with both in vitro drug sensitivity and patient survival and preserved consistent expression patterns between the cell lines and EOC patients. These biomarkers, which were discovered by simultaneously utilizing in vitro drug sensitivity and patient outcome information, were then used for our prediction modeling of each drug response. Using both principal component and crossvalidated regression analyses sequentially on the final biomarker set, we avoided model overfitting with the training NCI-60 set. For practical interpretation and use of our gene expression model prediction values without loss of information, the predicted scores were converted into rank-based percentile scores between zero and unity within each set. Trained models were evaluated with theMethods Patient DataIn vitro drug activity and microarray data for the 60 NCI cancer cell lines (NCI-60) were previously described [10]. In brief, publicly available drug sensitivity data for 50 growth inhibition (GI50) for the NCI-60 were obtained from the NCI Developmental AC220MedChemExpress AC220 Therapeutics Program (http://dtp.nci.nih.gov). NCI-60 expression profiling data with HG-U133A GeneChipH arrays (Affymetrix, Santa Clara, CA) were al.D from the TCGA public domain. For the UVA-51 cohort, we obtained and used the archived patient samples and deidentified clinical data which were consented for general research purpose and approved by the Institutional Review Board (PRC# 455-07) at the University of Virginia; its full description has been published elsewhere [8].Statistical AnalysisMultivariate models for predicting patient therapeutic responses to three chemotherapy drugs, paclitaxel, cyclophosphamide, and topotecan, were derived by integrating in vitro drug sensitivity data for the NCI-60 cell lines and clinical outcome information from EOC patients after standard chemotherapy. The schematic procedures for our model training and validation are summarized in Figure 1. First, initial gene expression biomarkers highly associated with in vitro drug sensitivity were identified from the NCI-60 microarray data by correlating each drug’s GI50 values for the NCI-60 with their genomic expression data for cyclophosphamide and topotecan treatment and by identifying differentially expressed biomarkers between sensitive and resistant cell lines of the NCI-60 to paclitaxel. These chemosensitivity biomarkers were then triaged based on the COXEN coefficient, which represents the degree of concordance of expression regulation between the NCI-60 cell lines and a general EOC patient population prior to standard chemotherapy [16]. In brief, derivation of the COXEN coefficient is based on a so-called “correlation of correlations,” which first calculates the expression correlations within each set for the identical set of genes and then evaluates gene-by-gene correlation between the correlation matrices of the two sets. This kind of second-order correlation has proven useful by us and others for investigating various gene networks to identify concordant data sets [17?9]. More detailed description of the COXEN algorithm can be found elsewhere [7,10]. The above biomarkers were further screened with ovarian cancer patient data: the Bonome-185 set for paclitaxel and cyclophosphamide and the TCGA-UW set for topotecan. A subset of each drug’s biomarkers significantly associated with patient survival was identified by a Cox regression survival analysis. Therefore, these final biomarkers were the genes significantly associated with both in vitro drug sensitivity and patient survival and preserved consistent expression patterns between the cell lines and EOC patients. These biomarkers, which were discovered by simultaneously utilizing in vitro drug sensitivity and patient outcome information, were then used for our prediction modeling of each drug response. Using both principal component and crossvalidated regression analyses sequentially on the final biomarker set, we avoided model overfitting with the training NCI-60 set. For practical interpretation and use of our gene expression model prediction values without loss of information, the predicted scores were converted into rank-based percentile scores between zero and unity within each set. Trained models were evaluated with theMethods Patient DataIn vitro drug activity and microarray data for the 60 NCI cancer cell lines (NCI-60) were previously described [10]. In brief, publicly available drug sensitivity data for 50 growth inhibition (GI50) for the NCI-60 were obtained from the NCI Developmental Therapeutics Program (http://dtp.nci.nih.gov). NCI-60 expression profiling data with HG-U133A GeneChipH arrays (Affymetrix, Santa Clara, CA) were al.