Robust assignment of cancer subtypes from expression data using a uni-variate gene expression average as classifier

BMC Cancer Robust assignment of cancer subtypes from expression data using a uni-variate gene expression average as classifier Martin Lauss 0 Attila Frigyesi 2 Tobias Ryden 1 Mattias Hglund 0 0 Department of Oncology, Clinical Sciences, Lund University and Lund University Hospital , SE-221 85 LUND , Sweden 1 Centre for Mathematical Sciences, Lund University , Box 118, SE-221 00 Lund , Sweden 2 Department of Anesthesiology and Intensive Care, Lund University Hospital , SE-221 85 Lund , Sweden Background: Genome wide gene expression data is a rich source for the identification of gene signatures suitable for clinical purposes and a number of statistical algorithms have been described for both identification and evaluation of such signatures. Some employed algorithms are fairly complex and hence sensitive to over-fitting whereas others are more simple and straight forward. Here we present a new type of simple algorithm based on ROC analysis and the use of metagenes that we believe will be a good complement to existing algorithms. Results: The basis for the proposed approach is the use of metagenes, instead of collections of individual genes, and a feature selection using AUC values obtained by ROC analysis. Each gene in a data set is assigned an AUC value relative to the tumor class under investigation and the genes are ranked according to these values. Metagenes are then formed by calculating the mean expression level for an increasing number of ranked genes, and the metagene expression value that optimally discriminates tumor classes in the training set is used for classification of new samples. The performance of the metagene is then evaluated using LOOCV and balanced accuracies. Conclusions: We show that the simple uni-variate gene expression average algorithm performs as well as several alternative algorithms such as discriminant analysis and the more complex approaches such as SVM and neural networks. The R package rocc is freely available at http://cran.r-project.org/web/packages/rocc/index.html. - Background One of the most promising clinical applications of genome wide expression studies is the construction of robust and reliable disease classifiers. Correct identification and sub-classification of diseases such as cancer is a prerequisite for proper and efficient treatment. To date a large number of different algorithms for disease classification have been described. They range in complexity from neural network approaches [1] to the simpler nearest-neighbor classification algorithms [2]. Even though some of the more complex approaches such as neural networks and self organized maps (SOM) [3] have proved to be very efficient, these methods often rely on the tuning of several parameters and hence are liable for over-fitting. Furthermore, simple classifiers seem to perform remarkably well when compared to more sophisticated ones [4]. In the present investigation our aim has been to design a simple predictor system useful for cancer subtype classification. Features to be included in the predictor signatures are selected based on their classification capacity as determined by a receiver operating characteristic (ROC) analysis and area under the curve (AUC) estimates [5,6]. After selection of the appropriate number of genes in the predictor signature, the mean expression level of all genes included is calculated, transforming the ensemble of genes into one vector and used as a uni-variate gene expression average, or a metagene, as classifier. Two features of gene expression are exploited by the merging of genes, genes are often co-regulated and hence correlated, and by using the expression level of the metagene, effects by random noise from single genes are minimized. Most of the commonly used algorithms such as SVM [7] and PAM [8] apply specifications such as support vectors and weights to the individual features included in the predictor gene signatures which potentially complicate their application to independent data [9]. Hence in this investigation we use an alternative way to evaluate the results by using the obtained training set gene signature genes only and then establish new parameters in the validation set to evaluate the performance of the classifier. We show that the proposed metagene classifier produces excellent accuracies, similar to what is obtained with a SVM approach, in several types of cancer data sets using a variety of tumor classification criteria. Implementation Data sets To establish the classifier we used bladder cancer datasets produced by Sanchez-Carbayo et al. [10] (Supplementary Table 10 in [10]) SanchezC, Stransky et al. [11] (ArrayExpress: E-TABM-147) Stransky"; and Blaveri et al. [12] (Supplementary Table 4 in [12]) Blaveri. The remaining datasets were obtained from Gene Expression Omnibus (GEO) [13], except for the vandeVijver breast cancer dataset [14]. The following datasets were downloaded from GEO; for breast GSE2034 (WangY), GSE2990 (Sotiriou), for neuroblastoma GSE3960 (WangQ), GSE12460 (JanoueixL), GSE19274 (Attiyeh), for lung GSE8569 (Angulo), GSE11969 (Takeuchi). For a detailed description of the datasets see Additional file 1. Normal urothelium samples, recurring tumors from the same patient, cell lines, and technical replicates were not included in the final bladder cancer data sets. The SanchezC dataset was quantile-normalized using the normalizeBetweenArrays function of the R package limma [15]. Robust Multi-array Average (RMA) was performed separately for two samples sets of the Stransky dataset (on U95A and U95Av2 respectively) using the affy package [16]. Obtained RMA expression values were de-logged, the samples sets combined, and quantile normalized using limma. The SanchezC and Stransky datasets were both transformed to log2 scale. To obtain gene-centered values the gene expression values were subtracted by the mean expression of the gene in each dataset separately. The Blaveri dataset was imputed for missing values using k-nearest neighbors (k = 10) for genes that had no more than 20% missing data, and genes with >20% missing data were omitted [17]. The HGNC GeneSymbols were updated in all datasets with the official HGNC GeneSymbols from the HGNC webpage [18]. The expression values of GeneSymbols with multiple reporters were merged by taking the median expression value. All reporters in the datasets without a GeneSymbol were discarded. The final SanchezC dataset contained 90 patients and 12761 genes, the Stransky dataset 56 patients and 8955 genes, and the Blaveri dataset 74 patients and 4430 genes. The SanchezC and Stransky datasets share a total of 8518 GeneSymbols and were used to explore the AUC characteristics. For classification, Ta and T1 cases were considered non-muscle invasive (NMI), and T2 cases as muscle-invasive (MI). Grade is discriminated between Grade 2 and 3 in SanchezC, and between Grade1+2 and Grade 3 in Sanchez. Randomized versions of the datasets were generated using the mean and st (...truncated)